fix(graph): collapse Docling InlineGroup and Picture children (#197)

Two patterns in Docling's serialization were mirrored 1:1 by the graph
projection and produced node explosions on real documents:

- An InlineGroup (paragraph of mixed style runs) emits one `groups[]`
  entry plus N `texts[]` runs. Naive iteration created one Paragraph
  node per run.
- A Picture's `children` carry internal text labels extracted by the
  layout model (flowchart boxes, chart axis labels, diagram callouts).
  Each child became its own Paragraph node, drowning the figure.

`build_collapse_index` (in the shared `infra.docling_tree` helper) now
returns the `skip_refs` set + `inline_meta` overrides for both cases.
The Neo4j `tree_writer` and the in-memory `docling_graph` consume the
same index, so both projections stay in sync.

InlineGroups are projected as a single :Paragraph carrying the
concatenated text and the union of children's provs (re-indexed).
Pictures keep their :Figure node and prov; their descendants are
dropped. Captions live in the picture's separate `captions` field, not
in `children`, so they are unaffected.

.env.example

@@ -48,3 +48,8 @@
 
 # Embedding vector dimension (default: 384 for Granite Embedding 30M / all-MiniLM-L6-v2)
 # EMBEDDING_DIMENSION=384
+
+# Neo4j — graph-native document structure (used by docker-compose.dev.yml)
+# NEO4J_URI=bolt://neo4j:7687
+# NEO4J_USER=neo4j
+# NEO4J_PASSWORD=changeme

.github/workflows/ci.yml

@@ -29,7 +29,7 @@ jobs:
         with:
           python-version: "3.12"
           cache: pip
-          cache-dependency-path: document-parser/requirements.txt
+          cache-dependency-path: document-parser/requirements-test.txt
 
       - name: Install system dependencies
         run: sudo apt-get update && sudo apt-get install -y --no-install-recommends poppler-utils
@@ -37,8 +37,8 @@ jobs:
       - name: Install Python dependencies
         run: |
           pip install --upgrade pip
-          pip install -r requirements.txt
+          pip install -r requirements-test.txt
-          pip install pytest pytest-asyncio httpx ruff
+          pip install httpx ruff
 
       - name: Lint
         run: ruff check .

.trivyignore.yaml

@@ -0,0 +1,12 @@
+vulnerabilities:
+  - id: CVE-2026-40393
+    paths:
+      - "usr/lib/x86_64-linux-gnu/libgbm.so.*"
+      - "usr/lib/x86_64-linux-gnu/dri/*"
+      - "usr/lib/x86_64-linux-gnu/libGLX_mesa.so.*"
+      - "usr/lib/x86_64-linux-gnu/libgallium-*.so"
+    statement: >-
+      Mesa OOB read (fix: Mesa 25.3.6 / 26.0.1). No Debian 13 backport available.
+      Pulled transitively by libgl1 (required by OpenCV/RapidOCR). Tracked in #189
+      with follow-up to drop libgl1 via opencv-python-headless.
+    expired_at: 2026-06-30

README.md

@@ -33,6 +33,7 @@ Upload a PDF, configure the extraction pipeline, and visualize the results — t
 - **Per-page results** — right panel syncs with the current PDF page
 - **Chunking** — split extracted content into semantic chunks (hierarchical, hybrid, or page-based) with configurable token limits and inline editing
 - **Ingestion pipeline** — Docling → chunking → embedding → OpenSearch vector indexing (one-click from Studio)
+- **Graph storage (Neo4j)** — full DoclingDocument tree (sections, paragraphs, tables, pages, chunks) mirrored as a graph with `PARENT_OF`, `NEXT`, `ON_PAGE`, `HAS_CHUNK`, `DERIVED_FROM` relations, with an in-app graph view powered by Cytoscape.js
 - **Markdown & HTML export** of extracted content
 - **Document management** — upload, list, delete, search, filter by indexing status
 - **Analysis history** — re-visit and open past analyses
@@ -58,7 +59,7 @@ Upload a PDF, configure the extraction pipeline, and visualize the results — t
 | **frontend** | Vue 3, TypeScript, Vite, Pinia | UI, PDF viewer, results display |
 | **document-parser** | FastAPI, Docling, SQLite, pdf2image | REST API, document parsing, storage |
 
-### Backend structure (clean architecture)
+### Backend structure (hexagonal architecture — ports & adapters)
 
 ```
 document-parser/
@@ -244,6 +245,69 @@ When ingestion is enabled, the UI shows:
 | `EMBEDDING_URL` | — | Embedding service endpoint (empty = ingestion disabled) |
 | `EMBEDDING_DIMENSION` | `384` | Vector dimension (must match embedding model) |
 
+## Graph storage with Neo4j (opt-in)
+
+Docling Studio can mirror the full **DoclingDocument tree** into a [Neo4j](https://neo4j.com/) graph: sections, paragraphs, tables, figures, pages, and chunks all become first-class nodes connected by `HAS_ROOT`, `PARENT_OF`, `NEXT`, `ON_PAGE`, `HAS_CHUNK`, and `DERIVED_FROM` edges. This enables queries that are impossible with a flat chunk store — navigating a document's outline, finding all tables under a given section, or tracing a chunk back to its source elements.
+
+Enable Neo4j with the ingestion profile (it ships alongside OpenSearch):
+
+```bash
+docker compose --profile ingestion \
+  -f docker-compose.yml -f docker-compose.ingestion.yml \
+  up --build
+```
+
+The Neo4j Browser is available at <http://localhost:7474> (user `neo4j`, password `changeme` by default).
+
+### Schema at a glance
+
+```mermaid
+graph TD
+    D[Document] -->|HAS_ROOT| SH[SectionHeader]
+    D -->|HAS_CHUNK| C[Chunk]
+    SH -->|PARENT_OF| P[Paragraph]
+    SH -->|PARENT_OF| T[Table]
+    P -->|NEXT| T
+    P -->|ON_PAGE| PG[Page]
+    T -->|ON_PAGE| PG
+    C -->|DERIVED_FROM| P
+    C -->|DERIVED_FROM| T
+```
+
+### Example Cypher queries
+
+Find all "Methods" sections across documents (impossible in vector-only stores):
+
+```cypher
+MATCH (d:Document)-[:HAS_ROOT]->(:Element)-[:PARENT_OF*]->(s:SectionHeader)
+WHERE toLower(s.text) CONTAINS 'method'
+RETURN d.title, s.text, s.level
+```
+
+Get the parent section and sibling elements of a chunk (context for RAG):
+
+```cypher
+MATCH (c:Chunk {id: $chunk_id})-[:DERIVED_FROM]->(e:Element)
+MATCH (e)<-[:PARENT_OF]-(parent:Element)-[:PARENT_OF]->(sibling:Element)
+RETURN parent, collect(sibling) AS siblings
+```
+
+List all tables from documents ingested from an `invoices/` path:
+
+```cypher
+MATCH (d:Document)-[:HAS_ROOT]->(:Element)-[:PARENT_OF*]->(t:Table)
+WHERE d.source_uri CONTAINS 'invoices/'
+RETURN d.title, t.caption, t.cells_json
+```
+
+| Variable | Default | Description |
+|----------|---------|-------------|
+| `NEO4J_URI` | — | Neo4j Bolt endpoint (empty = graph storage disabled) |
+| `NEO4J_USER` | `neo4j` | Neo4j username |
+| `NEO4J_PASSWORD` | `changeme` | Neo4j password |
+
+The in-app **Graph** tab (under *Results*) renders the per-document graph with [Cytoscape.js](https://js.cytoscape.org/) (see [ADR-001](docs/architecture/adrs/ADR-001-graph-visualization-library.md) for the library choice). Documents with more than **200 pages** return `HTTP 413` from `GET /api/documents/{id}/graph`; pagination ships in v0.6.
+
 ## CI / Release
 
 GitHub Actions pipelines (see [`.github/workflows/`](.github/workflows/)):

docker-compose.dev.yml

@@ -9,6 +9,26 @@
 # =============================================================================
 
 services:
+  # --- Neo4j (graph-native document structure) ---
+  neo4j:
+    image: neo4j:5.15-community
+    environment:
+      NEO4J_AUTH: ${NEO4J_USER:-neo4j}/${NEO4J_PASSWORD:-changeme}
+      NEO4J_server_memory_heap_initial__size: 512m
+      NEO4J_server_memory_heap_max__size: 1g
+    ports:
+      - "7474:7474"
+      - "7687:7687"
+    volumes:
+      - neo4j_data:/data
+      - neo4j_logs:/logs
+    healthcheck:
+      test: ["CMD-SHELL", "cypher-shell -u $${NEO4J_USER:-neo4j} -p $${NEO4J_PASSWORD:-changeme} 'RETURN 1' || exit 1"]
+      interval: 10s
+      timeout: 5s
+      retries: 10
+      start_period: 30s
+
   # --- OpenSearch (single-node, security disabled for local dev) ---
   opensearch:
     image: opensearchproject/opensearch:2
@@ -77,12 +97,17 @@ services:
       BATCH_PAGE_SIZE: ${BATCH_PAGE_SIZE:-10}
       OPENSEARCH_URL: http://opensearch:9200
       EMBEDDING_URL: http://embedding:8001
+      NEO4J_URI: bolt://neo4j:7687
+      NEO4J_USER: ${NEO4J_USER:-neo4j}
+      NEO4J_PASSWORD: ${NEO4J_PASSWORD:-changeme}
     command: ["uvicorn", "main:app", "--host", "0.0.0.0", "--port", "8000", "--reload"]
     depends_on:
       opensearch:
         condition: service_healthy
       embedding:
         condition: service_healthy
+      neo4j:
+        condition: service_healthy
     deploy:
       resources:
         limits:
@@ -105,6 +130,8 @@ services:
 
 volumes:
   opensearch_data:
+  neo4j_data:
+  neo4j_logs:
   uploads_data:
   db_data:
   frontend_node_modules:

docker-compose.yml

@@ -1,4 +1,22 @@
 services:
+  # --- Neo4j (graph-native document structure) ---
+  neo4j:
+    profiles: ["ingestion"]
+    image: neo4j:5.15-community
+    environment:
+      NEO4J_AUTH: ${NEO4J_USER:-neo4j}/${NEO4J_PASSWORD:-changeme}
+      NEO4J_server_memory_heap_initial__size: 512m
+      NEO4J_server_memory_heap_max__size: 1g
+    volumes:
+      - neo4j_data:/data
+      - neo4j_logs:/logs
+    healthcheck:
+      test: ["CMD-SHELL", "cypher-shell -u $${NEO4J_USER:-neo4j} -p $${NEO4J_PASSWORD:-changeme} 'RETURN 1' || exit 1"]
+      interval: 10s
+      timeout: 5s
+      retries: 10
+      start_period: 30s
+
   # --- OpenSearch (single-node, security disabled) ---
   opensearch:
     profiles: ["ingestion"]
@@ -53,6 +71,9 @@ services:
       BATCH_PAGE_SIZE: ${BATCH_PAGE_SIZE:-10}
       OPENSEARCH_URL: ${OPENSEARCH_URL:-}
       EMBEDDING_URL: ${EMBEDDING_URL:-}
+      NEO4J_URI: ${NEO4J_URI:-}
+      NEO4J_USER: ${NEO4J_USER:-neo4j}
+      NEO4J_PASSWORD: ${NEO4J_PASSWORD:-changeme}
     deploy:
       resources:
         limits:
@@ -69,5 +90,7 @@ services:
 
 volumes:
   opensearch_data:
+  neo4j_data:
+  neo4j_logs:
   uploads_data:
   db_data:

docs/architecture.md

@@ -8,7 +8,7 @@ Two services communicating via REST. The frontend is a Vue 3 SPA served by Nginx
 
 ### Zooming into the backend
 
-The schema above shows the macro view. Inside the backend, the code follows a **Clean Architecture** with strict layer boundaries:
+The schema above shows the macro view. Inside the backend, the code follows a **Hexagonal Architecture** (ports & adapters) with strict layer boundaries:
 
 ```
 ┌──────────────────────────────────────────────────────┐
@@ -34,9 +34,9 @@ The schema above shows the macro view. Inside the backend, the code follows a **
 
 Dependencies flow **inward**: `api → services → domain`. The domain layer has zero knowledge of HTTP or database.
 
-## Backend — Clean Architecture
+## Backend — Hexagonal Architecture (ports & adapters)
 
-The backend follows a strict layered architecture. Dependencies flow inward: API → Services → Domain. The domain layer has zero knowledge of HTTP or database.
+The backend follows the hexagonal / ports-and-adapters pattern. The domain layer defines **ports** (abstract protocols in `domain/ports.py`); `infra/` provides **adapters** that implement them. Dependencies flow inward: API → Services → Domain. The domain layer has zero knowledge of HTTP, database, or any framework.
 
 ```
 document-parser/

docs/architecture/adr-guide.md

@@ -34,7 +34,7 @@ These decisions were made before the ADR process was introduced. They are docume
 
 | Decision | Rationale | Date |
 |----------|-----------|------|
-| Clean Architecture (hexagonal) for backend | Decouple domain from framework — enable converter swapping (local/remote) | 2025-01 |
+| Hexagonal Architecture (ports & adapters) for backend | Decouple domain from framework — enable converter swapping (local/remote) via ports | 2025-01 |
 | FastAPI over Django | Async-first, lightweight, Pydantic-native — better fit for a single-purpose API | 2025-01 |
 | Vue 3 + Pinia over React | Composition API + built-in reactivity — smaller bundle for this use case | 2025-01 |
 | SQLite over PostgreSQL | Single-file DB, zero ops — appropriate for a document processing tool | 2025-01 |

docs/architecture/adrs/ADR-001-graph-visualization-library.md

@@ -0,0 +1,142 @@
+# ADR-001: Graph visualization library for the Neo4j graph view
+
+**Date**: 2026-04-17
+**Status**: Proposed
+**Deciders**: Pier-Jean Malandrino
+
+## Context
+
+v0.5.0 introduces Neo4j as a graph-native storage layer for parsed documents
+(see [docs/design/neo4j-integration.md](../../design/neo4j-integration.md)
+and [#186](https://github.com/scub-france/Docling-Studio/issues/186)). We need
+an in-app visualization of that graph: the `DoclingDocument` tree as rendered
+in Neo4j, with nodes colored by element type (`SectionHeader`, `Paragraph`,
+`Table`, `Figure`, `ListItem`, `Formula`) and edges (`PARENT_OF`, `NEXT`,
+`ON_PAGE`, `HAS_CHUNK`, `DERIVED_FROM`).
+
+The view lives in the existing Vue 3 debug panel. It is the **primary demo
+artifact** for the Hackernoon hackathon (Neo4j partner), so polish matters as
+much as correctness.
+
+### Constraints
+
+- Vue 3 + Vite frontend, no framework change
+- Must render the full tree of a 200-page document (worst case ≈ a few
+  thousand nodes; see graph endpoint cap in the design doc §8.4)
+- Needs a **clean hierarchical layout** — documents are trees, not arbitrary
+  graphs; a good tree layout is the single biggest UX lever
+- Needs per-node styling (shape + color by label), click, hover, zoom, pan
+- Must be installable without Java/Python-side changes
+- License compatible with the repo (MIT-ish preferred)
+
+### Non-goals for v0.5.0
+
+- 3D rendering
+- Force-directed simulation as the primary layout (we have a tree)
+- Editing nodes in place (read-only view)
+- Rendering millions of nodes
+
+## Decision
+
+Use **Cytoscape.js** via a thin Vue wrapper (`vue-cytoscape` or a bespoke
+`GraphView.vue` that imports `cytoscape` directly and uses the
+`dagre`/`breadthfirst` layouts).
+
+## Consequences
+
+### Positive
+
+- Battle-tested library (13k+ GitHub stars, maintained since 2013, used by
+  Neo4j's own "Bloom"-style visualizations in the community)
+- First-class support for hierarchical layouts via `cytoscape-dagre` (hub-and-
+  spoke / tree) and built-in `breadthfirst` — both map naturally to our
+  `PARENT_OF` structure
+- CSS-like selector syntax for styling (`node[label = "Table"] { ... }`),
+  which is pleasant to evolve as we add node types
+- Permissive licensing (MIT)
+- Headless mode available, so it can be tested outside a DOM (Jest + jsdom
+  works cleanly)
+- Active ecosystem: `cytoscape-cola`, `cytoscape-klay`, `cytoscape-popper` for
+  tooltips, all maintained
+- Bundle size is reasonable for a demo: ~300 KB min+gz for core + dagre, well
+  below our current frontend budget
+
+### Negative
+
+- Styling DSL is powerful but has its own syntax to learn; not plain CSS
+- Large graphs (>10k nodes) benefit from canvas+WebGL libraries
+  (sigma.js, reagraph) — we are explicitly not in that regime for v0.5, but
+  we would need to swap if we later visualize the cross-document graph
+- No Vue 3 component library that is both maintained and popular — we wrap it
+  ourselves in `GraphView.vue` (the wrapper is ~50 LOC, so this is minor)
+
+### Neutral
+
+- Not "Neo4j-branded": we do not use Neovis.js, which is a thin Cytoscape
+  wrapper around the Bolt protocol. Our graph API already returns shaped
+  JSON, so the Neovis convenience is not worth the lock-in
+- We take on one runtime dependency (`cytoscape` + `cytoscape-dagre`)
+
+## Alternatives Considered
+
+### Alternative 1: vis-network (vis.js)
+
+- **Pros**: Very easy to get started, built-in physics, shipped by Neo4j
+  Browser historically
+- **Cons**: Maintenance has been rocky (original vis.js split into several
+  forks; `vis-network` is the maintained branch but releases are sparse);
+  hierarchical layout is OK but less configurable than dagre; styling API is
+  less expressive; TypeScript types lag behind the JS API
+- **Why rejected**: Hierarchical layout quality is the single most important
+  criterion for a document tree, and vis-network is clearly a notch below
+  Cytoscape + dagre here. Maintenance trajectory is also a concern for a
+  release we want to keep shipping on
+
+### Alternative 2: Neovis.js
+
+- **Pros**: Built by Neo4j Labs, connects directly to a Bolt endpoint, nice
+  out-of-the-box "Neo4j look"
+- **Cons**: Wraps Cytoscape anyway, so everything it can do we can do with
+  Cytoscape directly; expects the browser to talk Bolt, which forces us to
+  expose Neo4j creds in the frontend OR to proxy Bolt through the backend
+  (both worse than our current "backend returns JSON" design); limited
+  customization compared to raw Cytoscape
+- **Why rejected**: The auth story is a non-starter for a hackathon demo we
+  want to show publicly, and we lose nothing vs. Cytoscape by going one
+  layer lower
+
+### Alternative 3: D3 (d3-hierarchy + d3-force)
+
+- **Pros**: Maximum flexibility; beautiful, publication-grade output; full
+  SVG control
+- **Cons**: Much more code for the same result — layout, zoom, pan, hover,
+  selection all hand-rolled; steeper learning curve for future contributors
+  to the project; no built-in graph data model
+- **Why rejected**: We're building a product feature, not a data-viz
+  artefact. The time budget (1 day of Day 3) doesn't fit a D3 build-your-own
+
+### Alternative 4: Reagraph / react-force-graph / sigma.js (WebGL)
+
+- **Pros**: Scales to tens of thousands of nodes at 60 FPS; good for future
+  cross-document visualization
+- **Cons**: Optimized for force-directed layouts, weaker hierarchical
+  support; Reagraph is React-only (requires a React island inside Vue);
+  sigma.js's tree layout is immature
+- **Why rejected**: Wrong regime for a single-document tree. Worth
+  reconsidering if/when we visualize the full corpus graph in a later release
+
+### Alternative 5: Mermaid
+
+- **Pros**: Trivial to embed, already used in docs
+- **Cons**: Static rendering, no interactivity, not designed for thousands of
+  nodes, no per-node click/hover
+- **Why rejected**: A viewer, not a visualizer. We need interactivity
+
+## References
+
+- [Neo4j integration design doc](../../design/neo4j-integration.md) §8.3
+- [Issue #186 — Neo4j integration](https://github.com/scub-france/Docling-Studio/issues/186)
+- [Cytoscape.js](https://js.cytoscape.org/)
+- [cytoscape-dagre](https://github.com/cytoscape/cytoscape.js-dagre)
+- [vis-network](https://visjs.github.io/vis-network/docs/network/)
+- [Neovis.js](https://github.com/neo4j-contrib/neovis.js)

docs/audit/audits/01-clean-architecture.md

@@ -1,6 +1,6 @@
-# Audit 01 — Clean Architecture
+# Audit 01 — Hexagonal Architecture (ports & adapters)
 
-**Objectif** : verifier que le backend respecte le flux de dependances strict `api -> services -> domain` et que chaque couche a une responsabilite claire.
+**Objectif** : verifier que le backend respecte le pattern hexagonal (ports dans `domain/ports.py`, adaptateurs dans `infra/`), le flux de dependances strict `api -> services -> domain`, et que chaque couche a une responsabilite claire.
 
 **Cible** : `document-parser/` (hors `.venv/`, `__pycache__/`, `tests/`)
 

docs/audit/master.md

@@ -74,7 +74,7 @@ Les audits sont executes dans l'ordre ci-dessous. Chacun est une fiche autonome
 
 | # | Audit | Fichier | Focus |
 |---|-------|---------|-------|
-| 01 | Clean Architecture | [01-clean-architecture.md](audits/01-clean-architecture.md) | Respect des couches, flux de dependances |
+| 01 | Hexagonal Architecture | [01-clean-architecture.md](audits/01-clean-architecture.md) | Ports & adapters, respect des couches, flux de dependances |
 | 02 | DDD | [02-ddd.md](audits/02-ddd.md) | Bounded contexts, entites, value objects, ubiquitous language |
 | 03 | Clean Code | [03-clean-code.md](audits/03-clean-code.md) | Nommage, taille, lisibilite |
 | 04 | KISS | [04-kiss.md](audits/04-kiss.md) | Simplicite, pas de sur-ingenierie |
@@ -161,7 +161,7 @@ Le fichier `reports/release-X.Y.Z/summary.md` consolide tous les audits :
 
 | # | Audit | Score | CRIT | MAJ | MIN | INFO | Verdict |
 |---|-------|-------|------|-----|-----|------|---------|
-| 01 | Clean Architecture | XX | N | N | N | N | GO |
+| 01 | Hexagonal Architecture | XX | N | N | N | N | GO |
 | 02 | DDD | XX | N | N | N | N | GO |
 | ... | ... | ... | ... | ... | ... | ... | ... |
 

docs/design/195-copy-paste-image-verify-mode.md

@@ -0,0 +1,404 @@
+# Design: Copy paste image in Verify mode
+
+<!--
+Design doc template for Docling Studio.
+
+One design doc per tracked issue. File path convention:
+  docs/design/<issue-number>-<kebab-slug>.md
+
+Status lifecycle: Draft → In review → Accepted → Implemented (or Superseded).
+Bump the Status line as the doc progresses; do not delete sections on the way.
+
+This template is tailored to the project's architecture and conventions:
+  - Backend Hexagonal Architecture / ports & adapters
+    (domain → api/services/persistence/infra)
+    see docs/architecture.md
+  - Backend coding standards (FastAPI + Pydantic camelCase, aiosqlite,
+    Python snake_case internal, max 300 lines/file, 30 lines/function)
+    see docs/architecture/coding-standards.md
+  - Frontend feature-based organization (Vue 3 + Pinia, one store per
+    feature, Composition API, TypeScript strict, data-e2e selectors)
+  - E2E with Karate UI (NOT Playwright) — see e2e/CONVENTIONS.md
+  - Audit dimensions used at release gate — see docs/audit/master.md
+  - ADR process for load-bearing decisions — see docs/architecture/adr-guide.md
+
+The `/conception` command pre-fills the header block and §1 / §2 / §12 from
+the linked issue. Everything else is on the author.
+-->
+
+- **Issue:** #195
+- **Title on issue:** [ENHANCEMENT] Copy paste image in Verify mode
+- **Author:** Pier-Jean Malandrino
+- **Date:** 2026-04-23
+- **Status:** Accepted
+- **Target milestone:** 0.5.0
+- **Impacted layers:** <backend: domain | api | services | persistence | infra> · <frontend: features/<name> | shared | app> · <e2e> · <infra/CI>
+- **Audit dimensions likely touched:** <pick from: Hexagonal Architecture · DDD · Clean Code · KISS · DRY · SOLID · Decoupling · Security · Tests · CI/Build · Documentation · Performance>
+- **ADR spawned?:** <no>  *(write an ADR when choosing a library, moving a boundary, or deciding **not** to do something — see `docs/architecture/adr-guide.md`)*
+
+---
+
+## 1. Problem
+
+<!--
+What hurts today, and for whom. Pull from the issue's Context + Current
+behavior sections — keep the user's voice, do not paraphrase aggressively.
+Two or three short paragraphs is usually enough. If you can't state the
+problem in plain language, you are not ready to design a solution.
+-->
+
+TODO: Why this issue exists. Link the user story, incident, or upstream discussion that motivates it.
+
+Today in Verify mode regarding image handling: TODO — describe the current baseline (upload-only? no paste target? no drag-drop?).
+
+## 2. Goals
+
+<!--
+Concrete, verifiable outcomes. Convert the issue's acceptance criteria into
+checkboxes here; the design is "done" when all are satisfied. Keep the list
+small — five or fewer goals is a good smell.
+-->
+
+Users should be able to copy/paste images directly into Verify mode (e.g. from clipboard) instead of only via file upload.
+
+- [ ] Define paste source (OS clipboard, drag-drop, screenshot)
+- [ ] Define target area in Verify mode UI
+- [ ] Define supported image formats and size limits
+- [ ] Size / type limits are env-var configurable, carry sane defaults, and are documented in `README.md` + `docs/deployment/*` (e.g. `MAX_PASTE_IMAGE_SIZE_MB`, `PASTE_ALLOWED_IMAGE_TYPES`)
+
+## 3. Non-goals
+
+<!--
+What this design explicitly does NOT try to solve — and, for each, where it
+*should* be solved (follow-up issue, next milestone, different audit area).
+This is the section that saves the review: naming the off-ramps up front
+prevents scope creep. If you leave this empty, reviewers will fill it in
+for you, badly.
+-->
+
+- ...
+- ...
+
+## 4. Context & constraints
+
+<!--
+The surrounding reality the design has to live in.
+
+### Existing code surface
+List the modules / files / stores this change touches. Prefer concrete paths
+over prose:
+  - Backend: document-parser/<layer>/<file>.py
+  - Frontend: frontend/src/features/<name>/{store,api,ui}.ts|.vue
+  - Persistence: document-parser/persistence/<repo>.py + schema in database.py
+  - E2E: e2e/<feature>.feature
+
+### Hexagonal Architecture constraints (backend)
+The domain layer has zero imports from api / persistence / infra, and
+defines ports (abstract protocols) that `infra/` adapters implement.
+Persistence imports only from domain. API never imports persistence
+directly — it goes through services. Call out any change that crosses
+these lines or adds / moves a port.
+
+### Deployment modes
+Docling Studio ships two images (`latest-local`, `latest-remote`) driven by
+`CONVERSION_ENGINE` — and a HF Space deployment on top of `latest-remote`.
+State which modes this design supports, which it does not, and how the
+frontend's feature flags (`chunking`, `disclaimer`) are affected.
+
+### Hard constraints
+Compatibility (SQLite schema, API contract, Pydantic DTOs), deadlines
+(milestone due date), deployment target (Docker Compose, HF Space),
+performance budget (matters for Performance audit), license / privacy
+(matters for Security audit).
+-->
+
+### SQLite & storage limits (must be enforced upstream of the DB)
+
+Pasted images follow the existing `documents` pattern: bytes land on disk
+under `UPLOAD_DIR`, the row stores `storage_path: TEXT`. We do **not**
+store base64 or BLOBs. Even so, app-level size guards must stay below
+SQLite's structural limits so a malformed request can never wedge the
+engine.
+
+Relevant SQLite defaults (see https://www.sqlite.org/limits.html):
+
+| SQLite limit | Default | Relevance |
+|---|---|---|
+| `SQLITE_MAX_LENGTH` | 1 GB (1 × 10⁹ bytes) | Max size of any single `TEXT` / `BLOB` cell. Irrelevant as long as we keep bytes off the DB. |
+| `SQLITE_MAX_SQL_LENGTH` | 1 MB | Max length of an SQL statement incl. inlined literals. Always use parameter binding — never inline image bytes. |
+| Page cache / WAL growth | n/a | Large writes bloat WAL until checkpoint; another reason to stay off-DB. |
+
+Our own app-level limits guard against ever reaching those ceilings.
+All such limits **must**: (1) carry a sane default in `infra/settings.py`,
+(2) be overridable via env var, (3) be documented in `README.md` and
+`docs/deployment/*`. This is consistent with how `MAX_FILE_SIZE_MB`
+(default 50) is handled today.
+
+## 5. Proposed design
+
+<!--
+The recommended approach, in enough detail that a competent engineer
+outside the immediate context can implement it. Describe contracts, not
+code — the PR is where code lives.
+
+Structure this section by layer. Skip a layer if it is genuinely untouched;
+do not pad.
+
+### 5.1 Domain
+New or changed dataclasses / value objects / ports in `document-parser/domain/`.
+No HTTP or DB concerns here. If you are adding a port (`Protocol`), give its
+full signature.
+
+### 5.2 Persistence
+Schema changes (table, columns, indexes), migration plan, aiosqlite query
+shape. Note whether existing rows need a backfill.
+
+### 5.3 Infra adapters
+New or changed adapters in `document-parser/infra/` (converter, chunker,
+rate limiter, settings). For new env vars, give name / default / allowed
+values.
+
+### 5.4 Services
+Use-case orchestration in `document-parser/services/`. Services do NOT
+implement — they delegate. Describe the call sequence, error handling,
+and concurrency (how does this interact with `MAX_CONCURRENT_ANALYSES`?).
+
+### 5.5 API
+Endpoint additions / changes in `document-parser/api/`. For each:
+  - Method + path
+  - Request DTO (Pydantic, camelCase via alias_generator)
+  - Response DTO (camelCase; remember `pages_json` stays snake_case)
+  - Error responses (status codes, shape)
+  - Whether it is excluded from the rate limiter (like `/api/health`)
+
+### 5.6 Frontend — feature module
+Which `frontend/src/features/<name>/` folder, which Pinia store actions,
+which API client calls in `api.ts`, which Vue components in `ui/`. Name
+new `data-e2e` attributes here (Karate needs them).
+
+### 5.7 Cross-cutting
+Feature flags (how the backend advertises capability via `/api/health` and
+how the frontend reacts), i18n strings (`shared/i18n.ts`), shared types
+(`shared/types.ts`).
+
+Prefer mermaid / ASCII for sequence and data flow. Interfaces are more
+valuable than pseudocode.
+-->
+
+### 5.1 Domain
+
+### 5.2 Persistence
+
+### 5.3 Infra adapters
+
+Extend `document-parser/infra/settings.py` with paste-specific limits.
+Follow the existing `MAX_FILE_SIZE_MB` pattern: typed field on the
+settings dataclass, `os.environ.get(...)` with a string default, cast
+at load time.
+
+| Setting | Env var | Default | Allowed | Notes |
+|---|---|---|---|---|
+| `max_paste_image_size_mb` | `MAX_PASTE_IMAGE_SIZE_MB` | `10` | positive int, `0` = unlimited | Must be ≤ `MAX_FILE_SIZE_MB`; upload validator rejects larger payloads before any DB write. |
+| `paste_allowed_image_types` | `PASTE_ALLOWED_IMAGE_TYPES` | `image/png,image/jpeg,image/webp` | comma-separated MIME list | Enforced server-side; frontend uses the same list via `/api/health`. |
+
+Validation happens in the API layer (upload handler) **before** the
+bytes reach persistence. Any future move to BLOB storage would still
+rely on these guards — they are the contract that prevents us ever
+approaching `SQLITE_MAX_LENGTH`.
+
+### 5.4 Services
+
+### 5.5 API
+
+### 5.6 Frontend — feature module
+
+### 5.7 Cross-cutting (feature flags, i18n, shared types)
+
+## 6. Alternatives considered
+
+<!--
+At least two genuine alternatives, each with a one-paragraph description
+and the reason it was rejected. "Do nothing" is often a legitimate
+alternative — name it if it is. Reviewers use this section to sanity-check
+that the recommended design was a choice and not the first thing that
+came to mind.
+
+If one of the alternatives represents a significant architectural fork
+(e.g. introducing a new service, replacing a library), spawn an ADR under
+`docs/architecture/adrs/` and link it in §12 — the design doc captures the
+local decision, the ADR captures the cross-cutting one.
+-->
+
+### Alternative A — <name>
+
+- **Summary:**
+- **Why not:**
+
+### Alternative B — <name>
+
+- **Summary:**
+- **Why not:**
+
+## 7. API & data contract
+
+<!--
+Make the wire contract explicit — this is what the frontend, e2e tests,
+and any external consumer will code against.
+
+### Endpoints
+| Method | Path | Request | Response | Breaking? |
+|--------|------|---------|----------|-----------|
+|        |      |         |          |           |
+
+Remember:
+  - API serialization is camelCase (Pydantic `alias_generator`).
+  - Backend internals stay snake_case.
+  - `pages_json` is the documented exception — it carries raw
+    `dataclasses.asdict()` output (snake_case).
+  - Health endpoint (`/api/health`) may need new fields if this design adds
+    a feature flag.
+
+### Persistence schema
+```sql
+-- ALTER TABLE / CREATE TABLE statements, with reasoning
+```
+
+### Env vars / config
+
+All new knobs must land in `README.md` and `docs/deployment/*` (same
+tables that already list `MAX_FILE_SIZE_MB`, `UPLOAD_DIR`, etc.).
+
+| Name | Default | Allowed | Notes |
+|------|---------|---------|-------|
+| `MAX_PASTE_IMAGE_SIZE_MB` | `10` | positive int (`0` = unlimited) | Guards app-level payload size; must stay ≤ `MAX_FILE_SIZE_MB` to avoid double-gating confusion. |
+| `PASTE_ALLOWED_IMAGE_TYPES` | `image/png,image/jpeg,image/webp` | comma-separated MIME list | Surfaced to the frontend via `/api/health` so the paste handler can reject client-side. |
+
+SQLite ceilings (`SQLITE_MAX_LENGTH` = 1 GB, `SQLITE_MAX_SQL_LENGTH` =
+1 MB) are **not** env-configurable — they are compile-time properties
+of the bundled engine. Document them in `docs/deployment/*` as
+background so operators understand why the app-level limits exist.
+
+### Breaking changes
+Enumerate anything a consumer must change. If there are none, say so
+explicitly — "additive only" is a useful commitment.
+-->
+
+## 8. Risks & mitigations
+
+<!--
+One row per non-trivial risk. Map each to an audit dimension so the
+release-gate audit has a clear hook:
+
+| Risk | Audit dimension | Likelihood | Impact | How we notice | Mitigation / rollback |
+|------|-----------------|-----------|--------|---------------|------------------------|
+|      | Security        |           |        |               |                        |
+|      | Performance     |           |        |               |                        |
+|      | Decoupling      |           |        |               |                        |
+
+Common families to scan for:
+  - **Hexagonal Architecture:** cross-layer imports, leaking HTTP into domain, adapter bypassing its port
+  - **Security:** rate limiter bypass, path traversal on uploads, SSRF via
+    the remote converter, unauthenticated data exposure
+  - **Performance:** synchronous work on the FastAPI event loop,
+    unbounded queries, new work inside `MAX_CONCURRENT_ANALYSES` budget
+  - **Tests:** coverage gap on a critical path
+  - **Documentation:** missing README / env var / i18n entry
+
+A design with "no risks identified" is a design that has not been read
+carefully.
+-->
+
+| Risk | Audit dimension | Likelihood | Impact | How we notice | Mitigation / rollback |
+|------|-----------------|------------|--------|---------------|------------------------|
+|      |                 |            |        |               |                        |
+
+## 9. Testing strategy
+
+<!--
+How this design will be verified. Be specific — name files / suites.
+
+### Backend — pytest (`document-parser/tests/`)
+  - Unit: per-layer (`tests/domain/`, `tests/persistence/`, `tests/services/`)
+  - Integration: services wired with real aiosqlite + real adapters
+  - Architecture tests (if applicable): enforce import boundaries
+
+### Frontend — Vitest (`frontend/src/**/*.test.ts`)
+  - Stores: actions / getters / mocked API
+  - Pure helpers (e.g. `bboxScaling.ts`-style modules): deterministic
+  - Components only when behavior is non-trivial; do not test markup
+
+### E2E — Karate UI (`e2e/`)
+  - Use `data-e2e` selectors — never CSS classes (see e2e/CONVENTIONS.md)
+  - `retry()` / `waitFor()` — never `Thread.sleep()` / `delay()`
+  - Setup via API, verify via UI, cleanup via API
+  - Tag appropriately: `@critical` / `@ui` / `@smoke` / `@regression` / `@e2e`
+  - **Never Playwright** — Karate is the tool here.
+
+### Manual QA
+Steps the reviewer can run locally (`docker-compose.dev.yml` up, scenario
+to reproduce). Keep it short — if the manual list is long, automate more.
+
+### Performance / load
+Required when the design claims a latency / throughput / memory property,
+or touches the conversion hot path.
+-->
+
+## 10. Rollout & observability
+
+<!--
+How this change gets to production safely.
+
+### Release branch
+Which `release/X.Y.Z` is the target? Any coordination with a parallel
+release (e.g. R&D branch)?
+
+### Feature flag / staged rollout
+Does the change hide behind a flag surfaced via `/api/health`? If so, what
+flips the flag, and what is the default? HF Space deployments often need
+`deploymentMode === 'huggingface'` gating.
+
+### Observability
+  - Logs to add / extend (structured, low-cardinality keys)
+  - Metrics / counters (if added — call out any new Prometheus names)
+  - New error modes to watch for in `analysis_jobs.status = FAILED`
+
+### Rollback plan
+The revert that is safe to apply at any time:
+  - Which migration is reversible? Which is not?
+  - Which env var flip disables the feature without a redeploy?
+  - Any data cleanup needed after rollback?
+
+Link to the existing release / ops playbooks:
+  - Deployment: `docs/release/*` (also surfaced via `/release:deploy`)
+  - Rollback: also surfaced via `/release:rollback`
+  - Incident: `docs/operations/*` (also surfaced via `/ops:incident`)
+-->
+
+## 11. Open questions
+
+<!--
+Things the author explicitly does not know yet, phrased as questions the
+reviewer can answer or redirect. Empty is allowed once the design is
+Accepted — during Draft / In review, this section is where the honest
+uncertainty lives. Resolve or delete each entry before shipping.
+-->
+
+- ...
+- ...
+
+## 12. References
+
+<!--
+Links to everything a future reader would want.
+-->
+
+- **Issue:** https://github.com/scub-france/Docling-Studio/issues/195
+- **Related PRs / commits:**
+- **ADRs:** <ADR-NNN or "none planned">
+- **Project docs:**
+  - Architecture: `docs/architecture.md`
+  - Coding standards: `docs/architecture/coding-standards.md`
+  - ADR guide / template: `docs/architecture/adr-guide.md`, `docs/architecture/adr-template.md`
+  - Audit master: `docs/audit/master.md`
+  - E2E conventions: `e2e/CONVENTIONS.md`
+- **External:** <specs, upstream issues, dashboards, third-party docs>

docs/design/neo4j-integration.md

@@ -0,0 +1,435 @@
+# Neo4j integration — Docling-Studio v0.5.0
+
+Design doc for Neo4j integration targeting release 0.5.0.
+Target: Hackernoon hackathon demo (Neo4j partner).
+
+---
+
+## 1. Context and goals
+
+### Already in Docling-Studio
+- Ingestion pipeline: Docling parser → chunking (HybridChunker) → embedding → OpenSearch (vector index)
+- Vue 3 + FastAPI UI
+- Debug view to inspect/edit chunks before retrieval
+- Docker compose with existing services
+
+### What we add in v0.5.0
+- Neo4j as **graph-native storage** of the document structure
+- A new ingestion layer that stores the DoclingDocument tree faithfully as a graph
+- Minimal UI to visualize the graph (demo value to the judges)
+- Compose pipeline with Neo4j
+
+### Why graph-native (hackathon positioning)
+> Most document AI tools store parsed content as flat chunks in a vector DB.
+> Docling-Studio v0.5 introduces a graph-native storage layer on top of Neo4j,
+> preserving the full hierarchical structure of documents as first-class citizens.
+> This unlocks hybrid retrieval, agentic navigation, and structural debugging —
+> impossible with chunk-only stores.
+
+### Out of scope for v0.5.0 (roadmap mention only)
+- EnrichmentWriter (entities / summaries / keywords via docling-agent) — v0.6.0
+- Agent reasoning trace viewer — v0.6.0
+- RAG hybrid (graph traversal + vector) — v0.7.0
+- Document versioning — v0.7.0+
+
+---
+
+## 2. Architectural principles
+
+### Port & adapter, with nuance
+
+**Write side**: one `Writer` port, **composable stages** (not alternative adapters).
+Pipelines A and B are additive, not exclusive.
+
+```
+  CORE (always)          Pipeline A (RAG)       Pipeline B (agent-ready, v0.6+)
+  ┌─────────────┐        ┌────────────────┐    ┌───────────────────┐
+  │ TreeWriter  │ ─────▶ │ ChunkWriter    │    │ EnrichmentWriter  │
+  │             │        │ (existing      │    │ (via docling-     │
+  │             │        │  OpenSearch +  │    │  agent, v0.6+)    │
+  │             │        │  adds chunks   │    │                   │
+  │             │        │  to Neo4j)     │    │                   │
+  └─────────────┘        └────────────────┘    └───────────────────┘
+```
+
+```python
+# docling_studio/ingestion/pipeline.py
+class Writer(Protocol):
+    def write(self, doc: DoclingDocument, ctx: IngestionContext) -> None: ...
+
+# Explicit composition per use case
+def build_pipeline(config: PipelineConfig) -> list[Writer]:
+    writers = [TreeWriter(neo4j_driver)]
+    if config.rag_enabled:
+        writers.append(ChunkWriter(neo4j_driver, chunker, embedder, opensearch))
+    if config.enrichment_enabled:  # v0.6.0+
+        writers.append(EnrichmentWriter(neo4j_driver, docling_agent))
+    return writers
+```
+
+**Read side**: two distinct ports (same Neo4j backend, different queries).
+
+```python
+class RAGRetrievalPort(Protocol):
+    def search(self, query: str, k: int) -> list[Chunk]: ...
+    def similar(self, chunk_id: str, k: int) -> list[Chunk]: ...
+
+class TreeNavigationPort(Protocol):  # v0.6.0+
+    def get_outline(self, doc_id: str) -> Tree: ...
+    def read_node(self, ref: str) -> Element: ...
+    def list_children(self, ref: str) -> list[Element]: ...
+    def walk(self, ref: str, depth: int) -> SubTree: ...
+```
+
+---
+
+## 3. Neo4j schema
+
+### Constraints & indexes (created at boot)
+
+```cypher
+// Uniqueness
+CREATE CONSTRAINT document_id IF NOT EXISTS
+  FOR (d:Document) REQUIRE d.id IS UNIQUE;
+
+CREATE CONSTRAINT element_composite IF NOT EXISTS
+  FOR (e:Element) REQUIRE (e.doc_id, e.self_ref) IS UNIQUE;
+
+CREATE CONSTRAINT page_composite IF NOT EXISTS
+  FOR (p:Page) REQUIRE (p.doc_id, p.page_no) IS UNIQUE;
+
+CREATE CONSTRAINT chunk_id IF NOT EXISTS
+  FOR (c:Chunk) REQUIRE c.id IS UNIQUE;
+
+// Full-text index (element text search)
+CREATE FULLTEXT INDEX element_text IF NOT EXISTS
+  FOR (e:Element) ON EACH [e.text];
+
+// Simple indexes for per-doc queries
+CREATE INDEX element_doc IF NOT EXISTS FOR (e:Element) ON (e.doc_id);
+CREATE INDEX chunk_doc  IF NOT EXISTS FOR (c:Chunk)   ON (c.doc_id);
+```
+
+### Data model
+
+```cypher
+// Root document
+(:Document {
+  id: string,                    // UUID or PDF hash
+  title: string,
+  source_uri: string,            // path or S3
+  ingested_at: datetime,
+  docling_version: string,
+  stages_applied: list<string>,  // ["tree", "chunks"] etc.
+  last_tree_write: datetime,
+  last_chunk_write: datetime,
+  tenant_id: string              // simple multi-tenancy
+})
+
+// All tree elements (shared :Element label + specific label)
+(:Element:SectionHeader {doc_id, self_ref, text, level, prov_page, prov_bbox})
+(:Element:Paragraph    {doc_id, self_ref, text, prov_page, prov_bbox})
+(:Element:Table        {doc_id, self_ref, caption, cells_json, prov_page, prov_bbox})
+(:Element:Figure       {doc_id, self_ref, caption, image_uri, prov_page, prov_bbox})
+(:Element:ListItem     {doc_id, self_ref, text, marker, prov_page, prov_bbox})
+(:Element:Formula      {doc_id, self_ref, latex, text, prov_page, prov_bbox})
+
+// Page for layout provenance
+(:Page {doc_id, page_no, width, height})
+
+// Chunks (Pipeline A)
+(:Chunk {
+  id, doc_id,
+  text,
+  chunk_index,
+  embedding_ref,   // id in OpenSearch (no inline duplication)
+  token_count
+})
+```
+
+### Relations
+
+```cypher
+// Hierarchical structure
+(:Document)-[:HAS_ROOT]->(:Element)
+(:Element)-[:PARENT_OF {order: int}]->(:Element)  // order preserves sequence
+(:Element)-[:NEXT]->(:Element)                     // DFS pre-order reading
+
+// Layout
+(:Element)-[:ON_PAGE]->(:Page)
+
+// Pipeline A (chunking)
+(:Document)-[:HAS_CHUNK]->(:Chunk)
+(:Chunk)-[:DERIVED_FROM]->(:Element)  // back-reference; a chunk can span multiple elements
+```
+
+### Decisions
+
+| Decision | Choice | Rationale |
+|----------|-------|---------------|
+| Element composite key | `(doc_id, self_ref)` | self_ref not unique across docs |
+| Multi-tenancy | `tenant_id` property on Document | Simple, filterable, migrable to multi-db later |
+| Table cells | `cells_json` property | v0.5 KISS. May model `(Table)-[:HAS_CELL]->(Cell)` in v0.6+ |
+| Reading order | `[:NEXT]` chain + `{order}` on `PARENT_OF` | Both views useful |
+| Versioning | None (replace strategy on re-upload) | v0.5 KISS |
+| APOC | Not required | Pure Cypher is sufficient for v0.5 |
+
+### Re-ingestion strategy
+
+```cypher
+// Before ingesting, wipe existing
+MATCH (d:Document {id: $doc_id})
+OPTIONAL MATCH (d)-[:HAS_ROOT|HAS_CHUNK]->()
+DETACH DELETE d
+// Then re-walk cleanly
+```
+
+---
+
+## 4. Implementation plan (3 days)
+
+### Day 1 — Infra + schema
+
+- [ ] Add `neo4j` service to `docker-compose.yml` (`neo4j:5.15-community`, persistent volume, healthcheck)
+- [ ] Add env vars (`NEO4J_URI`, `NEO4J_USER`, `NEO4J_PASSWORD`) to `.env.example`
+- [ ] Create module `docling_studio/storage/neo4j/`:
+  - `driver.py` — neo4j-python driver wrapper (connection pool, context manager)
+  - `schema.py` — idempotent `bootstrap_schema()` (CREATE CONSTRAINT / INDEX at startup)
+  - `__init__.py` with exports
+- [ ] Hook `bootstrap_schema()` in FastAPI startup
+- [ ] Basic integration tests:
+  - Driver connection
+  - Schema bootstrap (idempotence verified)
+  - Simple round-trip: write Document, read Document, delete Document
+
+**Deliverable:** docker compose boots with healthy Neo4j, schema in place at init.
+
+### Day 2 — TreeWriter (write + read)
+
+- [ ] `storage/neo4j/tree_writer.py` — `DoclingDocument → Neo4j` walker
+  - `write_document(doc, tenant_id, driver)` in a transaction
+  - DFS pre-order, batched `MERGE` for perf
+  - Pages first, then Elements, then `PARENT_OF` / `NEXT` / `ON_PAGE` relations
+  - Dynamic labels based on `node.label` (`SectionHeader`, `Paragraph`, …)
+- [ ] `storage/neo4j/tree_reader.py` — inverse walker `Neo4j → DoclingDocument`
+  - `read_document(doc_id, driver) -> DoclingDocument`
+  - Loads all Elements + Pages, rebuilds the Pydantic structure
+  - Prerequisite for v0.6 (feeding docling-agent from Neo4j)
+- [ ] Integrate into existing ingestion pipeline:
+  - Add TreeWriter as first stage of `IngestionPipeline`
+  - `neo4j_enabled: bool` config toggle
+- [ ] Round-trip tests:
+  - 3–4 varied PDFs (academic, invoice, report)
+  - Assertion: `doc_original == read_document(write_document(doc_original))`
+  - Beware dates, bbox floats (tolerance)
+
+**Deliverable:** A PDF uploaded to Docling-Studio is fully present in Neo4j and rebuildable.
+
+### Day 3 — UI + ChunkWriter + packaging
+
+- [ ] `storage/neo4j/chunk_writer.py`:
+  - After existing chunking, push each Chunk to Neo4j
+  - Create `(:Chunk)-[:DERIVED_FROM]->(:Element)` via source element `self_ref`
+  - Do NOT duplicate embeddings (stay in OpenSearch, keep `embedding_ref`)
+- [ ] Frontend: new "Graph view" tab in debug panel
+  - Vue component with `cytoscape` (lighter, better layout API — see [ADR-001](../architecture/adrs/ADR-001-graph-visualization-library.md))
+  - FastAPI endpoint `/api/documents/{doc_id}/graph` returns full nodes + edges for the document, **capped at 200 pages** (HTTP 413 beyond; pagination deferred to v0.6). The endpoint must include a `truncated: bool` flag and `node_count` / `edge_count` in the response envelope so the UI can warn the user cleanly.
+  - View: vertical tree, colors per node type, click-to-zoom, hover details
+- [ ] Per-document "Graph-ready" / "RAG-ready" badge in list
+- [ ] README update:
+  - "Graph storage with Neo4j" section
+  - Schema diagram (Mermaid or image)
+  - 2–3 Cypher examples like "find all paragraphs under section X that mention Y"
+  - Neo4j badge in features list
+- [ ] (bonus if time) "Query explorer" dev tab for live demo: Cypher editor + results
+
+**Deliverable:** release 0.5.0 with Neo4j visible, functional, documented.
+
+---
+
+## 5. Proposed code structure
+
+```
+docling_studio/
+├── storage/
+│   ├── neo4j/
+│   │   ├── __init__.py
+│   │   ├── driver.py           # connection management
+│   │   ├── schema.py           # bootstrap_schema()
+│   │   ├── tree_writer.py      # DoclingDocument -> Neo4j
+│   │   ├── tree_reader.py      # Neo4j -> DoclingDocument
+│   │   ├── chunk_writer.py     # Chunks -> Neo4j
+│   │   └── queries.py          # shared Cypher queries
+│   ├── opensearch/             # (existing)
+│   └── ports.py                # Writer, RAGRetrievalPort protocols
+├── ingestion/
+│   └── pipeline.py             # IngestionPipeline composing Writers
+├── api/
+│   └── graph.py                # /api/documents/{id}/graph
+└── frontend/
+    └── components/
+        └── GraphView.vue       # cytoscape + graph API fetch
+```
+
+---
+
+## 6. Docker compose (added excerpt)
+
+```yaml
+services:
+  neo4j:
+    image: neo4j:5.15-community
+    environment:
+      NEO4J_AUTH: ${NEO4J_USER:-neo4j}/${NEO4J_PASSWORD:-changeme}
+      NEO4J_PLUGINS: '["apoc"]'
+      NEO4J_server_memory_heap_initial__size: 512m
+      NEO4J_server_memory_heap_max__size: 1g
+    ports:
+      - "7474:7474"   # Browser UI (demo)
+      - "7687:7687"   # Bolt protocol
+    volumes:
+      - neo4j_data:/data
+      - neo4j_logs:/logs
+    healthcheck:
+      test: ["CMD-SHELL", "cypher-shell -u neo4j -p $${NEO4J_PASSWORD:-changeme} 'RETURN 1' || exit 1"]
+      interval: 10s
+      timeout: 5s
+      retries: 10
+
+  docling-studio-backend:
+    depends_on:
+      neo4j:
+        condition: service_healthy
+    environment:
+      NEO4J_URI: bolt://neo4j:7687
+      NEO4J_USER: neo4j
+      NEO4J_PASSWORD: ${NEO4J_PASSWORD:-changeme}
+
+volumes:
+  neo4j_data:
+  neo4j_logs:
+```
+
+---
+
+## 7. Tests
+
+### Unit tests
+- `tests/storage/neo4j/test_schema.py` — bootstrap is idempotent
+- `tests/storage/neo4j/test_tree_writer.py` — round-trip on synthetic DoclingDocument
+- `tests/storage/neo4j/test_chunk_writer.py` — chunks written with correct `DERIVED_FROM`
+
+### Integration tests
+- `tests/integration/test_ingestion_pipeline.py` — full pipeline on a real PDF
+- PDF fixtures: 1 academic (complex heading hierarchy), 1 invoice (tables), 1 report (lists)
+
+### E2E (bonus)
+- Upload PDF via UI → check structure in Neo4j Browser
+
+---
+
+## 8. Open decisions to settle before coding
+
+1. **Neo4j edition**: Community (free) or AuraDB (managed) ?
+   - Rec: Community in Docker for v0.5.0 dev/demo. AuraDB mentioned as prod option.
+
+2. **Chunks: duplicate embeddings in Neo4j or OpenSearch ref ?**
+   - Rec: OpenSearch ref (avoid duplication; OpenSearch remains source of truth for vectors). In v0.6+, consider native Neo4j vector index.
+
+3. **Graph view UI: cytoscape or vis-network ?**
+   - Decided: **Cytoscape.js** — see [ADR-001](../architecture/adrs/ADR-001-graph-visualization-library.md) for the full analysis.
+
+4. **Graph endpoint: return full doc or paginate ?**
+   - Decided: full doc for v0.5, **hard cap at 200 pages**. Beyond the cap, the endpoint returns HTTP 413 with a `truncated: true` flag; the UI shows "Graph too large to render — reduce scope". Pagination ships in v0.6.
+
+5. **Error strategy**: if Neo4j is down at ingestion, fail or degrade gracefully ?
+   - Rec: **fail fast** for v0.5 (avoid silent inconsistencies). `neo4j_required: bool` config option.
+
+---
+
+## 9. Hooks for later (v0.6.0+ — don't implement but prepare)
+
+**EnrichmentWriter (v0.6)** — will need:
+- The reader (Neo4j → DoclingDocument) to re-materialize the doc, feed docling-agent, re-patch enrichments
+- A stage addable to `IngestionPipeline` without touching other stages
+- An `:Entity` label (not created in v0.5 but schema-compatible)
+
+**Agent reasoning trace viewer (v0.6)** — will need:
+- An event stream (WebSocket) that v0.5 already prepares via the reactive UI
+- A node_ref ↔ Element correlation in Neo4j (our composite `self_ref` key is enough)
+
+**TreeNavigationPort (v0.7)** — will need:
+- Optimized Cypher queries for descendant/ancestor walk (indexes already provisioned)
+
+---
+
+## 10. v0.5.0 success criteria
+
+**Must have:**
+- [ ] A PDF uploaded to Docling-Studio is in Neo4j with structure preserved
+- [ ] Neo4j Browser shows the graph and is manually explorable
+- [ ] A graph visual in the Docling-Studio UI works
+- [ ] `docker compose up` works zero-config
+- [ ] README mentions Neo4j and describes the schema
+
+**Nice to have (decreasing priority):**
+- [ ] Graph-ready / RAG-ready badge per doc
+- [ ] Live query explorer in the UI
+- [ ] 2–3 example queries in README that do something impossible with vector-only
+
+**For the hackathon (post-release):**
+- [ ] 60s video: upload PDF → structure in Neo4j → cross-doc query impossible in vector-only
+- [ ] HackerNoon post explaining "graph-native documents" positioning
+- [ ] Explicit Neo4j partnership mention
+
+---
+
+## 11. Fundamental architectural decisions recap
+
+| Question | Answer |
+|----------|---------|
+| Is Neo4j source of truth or cache ? | **Source of truth** for structure. OpenSearch remains source of truth for embeddings. |
+| Does chunking go away ? | No, v0.5.0 keeps existing chunking. "Chunkless" is Pipeline B, v0.6+. |
+| Can it be toggled per doc ? | Yes — `stages_applied` on Document + pipeline config |
+| What about OpenSearch ? | Stays, stores vectors. Neo4j tracks `(:Chunk)-[:DERIVED_FROM]->(:Element)` links. |
+| Multi-tenancy ? | `tenant_id` property on Document, Cypher filter |
+| Versioning ? | None for v0.5.0 — replace strategy on re-upload |
+
+---
+
+## Appendix — Demo queries
+
+### Query 1 — All "Methods" sections across documents
+```cypher
+MATCH (d:Document)-[:HAS_ROOT]->(:Element)-[:PARENT_OF*]->(s:SectionHeader)
+WHERE toLower(s.text) CONTAINS 'method'
+RETURN d.title, s.text, s.level
+```
+
+### Query 2 — Context of a chunk (parent + siblings)
+```cypher
+MATCH (c:Chunk {id: $chunk_id})-[:DERIVED_FROM]->(e:Element)
+MATCH (e)<-[:PARENT_OF]-(parent:Element)
+MATCH (parent)-[:PARENT_OF]->(sibling:Element)
+RETURN parent, collect(sibling) AS siblings
+```
+
+### Query 3 — All tables from a document type
+```cypher
+MATCH (d:Document)-[:HAS_ROOT]->(:Element)-[:PARENT_OF*]->(t:Table)
+WHERE d.source_uri CONTAINS 'invoices/'
+RETURN d.title, t.caption, t.cells_json
+```
+
+### Query 4 — Direct children of a section (ordered)
+```cypher
+MATCH (s:Element {doc_id: $doc_id, self_ref: $section_ref})
+MATCH (s)-[pof:PARENT_OF]->(child)
+RETURN child
+ORDER BY pof.order
+```
+
+---
+
+*Single reference doc for Neo4j v0.5.0 implementation.
+Read this first in the implementation thread.*

docs/design/reasoning-trace.md

@@ -0,0 +1,253 @@
+# Reasoning Trace Viewer — Docling-Studio v0.6.0 (R&D preview)
+
+Design doc for the `docling-agent` reasoning trace viewer.
+Targeted release: **v0.6.0** (R&D branched from `release/0.5.0` in parallel to the
+0.5 build, so the Neo4j foundation can be leveraged without blocking the hackathon deliverable).
+
+Positioning one-liner:
+> Studio becomes the **reference viewer for any `docling-agent` run** — not another
+> chatbot. The PDF is the debug surface.
+
+---
+
+## 1. Context
+
+### Upstream trigger
+Peter Staar (IBM) suggested surfacing the LLM reasoning trace as `docling-agent`
+walks a `DoclingDocument` outline (chunkless RAG, new IBM repo).
+
+### `docling-agent` in one paragraph
+`DoclingRAGAgent(model_id, tools, max_iterations=5).run(task, sources=[doc])` returns
+a `RAGResult` with `answer`, `converged`, and `iterations: list[RAGIteration]`.
+Each `RAGIteration` carries: `iteration`, `section_ref` (JSON-pointer, e.g. `#/texts/3`),
+`reason`, `can_answer`, `response`, `section_text_length`. No bbox — must be resolved
+through `DoclingDocument.<items>[i].prov[0].bbox` + `page_no`. Runs on Mellea
+(Ollama / OpenAI / HF / WatsonX / LiteLLM / Bedrock). Observability is stdout logs only.
+
+### What Studio already brings
+- **Neo4j graph of the document** (0.5.0, just landed) — every `Element` is keyed by
+  `(doc_id, self_ref)`, Cytoscape node id is `elem::${self_ref}`. **This is the
+  killer enabler.** `RAGIteration.section_ref → node` is a string concat, no resolver.
+- `GraphView.vue` (Cytoscape + dagre) already handles styles via selectors
+  (`selector: 'edge[type = "NEXT"]'`, `selector: 'node[kind = "section"]'`) — adding
+  a `visited` class + `REASONING_NEXT` synthetic edge type is ~20 LOC of style.
+- `analysis_jobs.document_json` in SQLite → DoclingDocument available for the sidecar
+  runner (no PDF re-conversion). Not used by the viewer itself.
+
+### Personas
+- **v1 (this plan)**: dev / integrator of `docling-agent` debugging a run that went wrong.
+- **v2 (roadmap)**: live runner with synchronized demo UX.
+- v3+ (non-goals here): business analyst for semantic navigation, batch QA.
+
+---
+
+## 2. Scope split — **debug first, demo second**
+
+Rendering surface pivoted: **the trace is drawn on the Neo4j graph**, not on the PDF.
+See §1. This kills the whole bbox-resolution stack from v1.
+
+| Phase | Value | Runtime deps | Surface |
+|---|---|---|---|
+| **v1 — Debug (this plan)** | Import externally-produced `RAGResult` JSON, overlay trace on the existing GraphView | **None** server-side. Pure frontend. | GraphView: visited nodes highlighted in order, synthetic `REASONING_NEXT` edges |
+| **v2 — Demo (follow-up)** | Run the agent live against a loaded document | Ollama + Mellea + `docling-agent` (new opt-dep group `rag`) | Same GraphView + SSE streaming of iterations, staggered reveal |
+
+Building v1 first de-risks the **graph-trace UX** on real runs (produced by the
+R&D sidecar — see `experiments/reasoning-trace/`) before wiring the live runner.
+Code shared between v1 and v2 is the GraphView overlay itself — 100 % reused.
+
+**Prerequisite for v1**: the target document must have been processed through the
+"Maintain" step (Neo4j pipeline). Otherwise the graph is empty and the trace has
+nowhere to render — surface an explicit "Run the Maintain step first" empty state.
+
+---
+
+## 3. v1 — Debug mode (frontend-only)
+
+### 3.1 No backend changes in v1
+
+The GraphView already loads nodes keyed by `self_ref` via `GET /api/documents/{doc_id}/graph`.
+Iteration `section_ref` → Cytoscape node id is `` `elem::${section_ref}` `` — a client-side
+string concat. Nothing to compute server-side.
+
+Consequences:
+- No new router, no new service, no new pydantic model, no new migration.
+- No dependency on `docling-agent` in `document-parser/requirements.txt`.
+- `RAGResult` JSON (as produced by `experiments/reasoning-trace/`) is consumed
+  entirely by the frontend.
+
+### 3.2 Frontend — feature folder
+
+New `frontend/src/features/reasoning/`:
+
+```
+reasoning/
+├── store/reasoningStore.ts         # Pinia: trace, activeIteration, importDialogOpen
+├── ui/
+│   ├── ReasoningPanel.vue          # Side panel: query, answer, iteration list
+│   ├── IterationCard.vue           # Single iteration row (reason + can_answer badge)
+│   ├── ImportTraceDialog.vue       # Drag-drop / paste RAGResult JSON
+│   └── GraphReasoningOverlay.ts    # NOT a component — a plugin that decorates cy
+└── types.ts                        # RAGIteration, RAGResult mirror types
+```
+
+### 3.3 Graph overlay — how it's drawn
+
+`GraphReasoningOverlay` takes the existing `cy` Cytoscape instance (exposed from
+`GraphView.vue` via `defineExpose`) and:
+
+1. For each `iteration[i].section_ref`, find node `` `elem::${section_ref}` ``. If
+   missing, tag as `resolution_status: "not_in_graph"` and show a warning in the panel
+   (common cause: doc not processed through Maintain, or agent returned a ref that
+   points at a non-Element node).
+2. Add class `visited` + data attribute `visitOrder: i+1` on matched nodes.
+3. Insert **synthetic edges** between successive visited nodes with `type: "REASONING_NEXT"`
+   and `data: { order: i }`. These edges are UI-only, never written to Neo4j.
+4. On import, fit viewport to the visited subgraph (`cy.fit(cy.$('.visited'), 80)`).
+5. On iteration card click → `cy.$(`#elem::${ref}`).flashClass('pulse', 800)` +
+   centered pan.
+
+Cytoscape styles (append to the existing stylesheet array in `GraphView.vue`):
+
+```js
+{ selector: 'node.visited',
+  style: { 'border-color': '#EA580C', 'border-width': 3, 'overlay-opacity': 0 } },
+{ selector: 'node.visited[visitOrder]',
+  style: { label: 'data(visitOrder)', 'text-valign': 'top',
+           'text-background-color': '#EA580C', 'text-background-opacity': 1,
+           'color': '#FFFFFF', 'font-weight': 700 } },
+{ selector: 'edge[type = "REASONING_NEXT"]',
+  style: { 'line-color': '#EA580C', 'target-arrow-color': '#EA580C',
+           'target-arrow-shape': 'triangle', 'curve-style': 'bezier',
+           width: 2, 'z-index': 99 } },
+```
+
+Color ramp: single warm color (`#EA580C`) for v1. Gradient cold→warm is v2 polish.
+
+### 3.4 Integration points
+
+- `StudioPage.vue` → "Maintain" tab gains an **"Import reasoning trace"** action
+  (don't add a 3rd mode — the viz lives inside the graph view, not a new workspace).
+- `GraphView.vue` → add `defineExpose({ cy })` + a `<slot name="overlay" :cy="cy"/>`
+  that the parent can populate with `<ReasoningPanel>`.
+- `ReasoningPanel` appears as a right rail when a trace is loaded; collapsible.
+
+### 3.5 Empty / error states
+
+- **Graph empty for this doc** → "Run the Maintain step first. Neo4j has no graph for
+  this document yet." (the Maintain button is literally next to it.)
+- **All `section_ref`s unresolved in graph** → "None of the visited sections exist in
+  the graph. The agent may have been run against a different document, or the doc was
+  re-analyzed since. Re-run Maintain or re-run the agent."
+- **Some resolved, some not** → show trace with the missing ones greyed out in the panel.
+
+### 3.6 Tests
+
+No backend tests in v1 (no backend code).
+
+Frontend (Vitest):
+- `reasoningStore.test.ts` — import trace, active iteration transitions, reset on doc change.
+- `graphReasoningOverlay.test.ts` — given a mock `cy` (`cytoscape({ headless: true })`)
+  with a known node set, verify `visited` class applied to the right ids and the
+  correct synthetic edges added.
+- `ReasoningPanel.test.ts` — empty / loaded / partial-resolution states.
+
+### 3.7 Out of scope for v1
+- Live agent runner (v2).
+- Multi-doc queries — reject import if `RAGResult` was produced against `len(sources) > 1`.
+- Phrase-level attribution — `docling-agent` doesn't emit it.
+- Persisting traces in Neo4j — see §7.
+- PDF highlighting — dropped from v1. Could come back as v2.5 if demand exists.
+
+---
+
+## 4. File inventory (v1)
+
+**New — R&D sidecar** (already scaffolded on this branch)
+- `experiments/reasoning-trace/inspect_doc.py` — self-contained `uv run` script.
+- `experiments/reasoning-trace/README.md`
+- `experiments/reasoning-trace/.gitignore`
+
+**New — frontend**
+- `frontend/src/features/reasoning/**` (see §3.2)
+- Vitest siblings under `**/*.test.ts`
+
+**Touched**
+- `frontend/src/features/analysis/ui/GraphView.vue` — `defineExpose({ cy })` +
+  `<slot name="overlay">` + 3 new style selectors.
+- `frontend/src/pages/StudioPage.vue` — "Import reasoning trace" action in the
+  Maintain tab rail.
+
+**Untouched**
+- Entire `document-parser/` backend — no new router, service, schema, or dep.
+- `pyproject.toml` / `requirements.txt` — **no new runtime dep in v1**.
+- Neo4j schema — synthetic edges are client-side Cytoscape only.
+- OpenSearch / ingestion — untouched.
+- SQLite schema — no migration.
+
+---
+
+## 5. Risks & mitigations
+
+| Risk | Mitigation |
+|---|---|
+| `RAGResult` schema drifts in `docling-agent` | `schema_version` discriminator; strict pydantic; one canonical fixture from Peter pinned in CI. |
+| `section_ref` variants (`#/texts/3` vs `#/body/texts/3`) | Normalize in parser; regex test matrix. |
+| Synthetic groups without `prov` | Documented child-walk fallback + `resolved_via_child` status surfaced in UI. |
+| Large `RAGResult` (hundreds of iterations) | Hard-cap `iterations` at 50 in v1 (Peter's agent uses `max_iterations=5` by default) — return 413 above. |
+| `document_json` blob large (some docs > 5 MB) | `analysis_repo` already handles it; but **do not** log the blob. Add redaction test. |
+| Section ref not in graph (doc not through Maintain, or re-analyzed) | Explicit empty-state in `ReasoningPanel` with a link to the Maintain tab. Partial resolution shown as grey in the trace list. |
+| Feature creeping into 0.5.0 | This branch targets **v0.6.0**. Do not merge into `release/0.5.0`. Rebase onto the next release branch when cut. |
+
+---
+
+## 6. Spec anchoring
+
+Pin the `RAGResult` shape to **docling-agent commit SHA at the time of v1 merge** in
+a short ADR `docs/architecture/adrs/ADR-002-rag-result-schema.md`. The schema is
+upstream, unversioned, and will move — this doc freezes the contract Studio imports.
+
+---
+
+## 7. v2 preview — demo mode (not in this plan)
+
+Kept here to constrain v1 interfaces so nothing needs rewriting:
+- `POST /api/rag/answer` — server-side runner. Accepts `{doc_id, question, model_id}`.
+  Streams iterations via SSE. Frontend consumes the stream with the same
+  `GraphReasoningOverlay` used by v1 import — iterations appear one by one with
+  staggered reveal (~400 ms) as the SSE stream drips them in.
+- Ollama wired through `Mellea` — new optional dep group `rag`.
+- Persist traces in Neo4j as `(:ReasoningRun {id, query, converged})-[:VISITED {order,
+  reason, can_answer}]->(:Element)` for replay + cross-run analytics. Leverages
+  `TreeWriter` pattern already present. This is where the synthetic UI edges become
+  real graph edges.
+- Cross-run comparison view: overlay multiple runs on the same graph, diff the paths.
+
+---
+
+## 8. Branch & workflow
+
+- Branch: **`feature/reasoning-trace`** off `origin/release/0.5.0`.
+- Merge target: **next release branch (`release/0.6.0`)** once cut — *not* `0.5.0`.
+- Until then: live on the feature branch; rebase onto `release/0.5.0` periodically to
+  absorb Neo4j fixes.
+- Issues: one umbrella + one per §4 subsystem (resolver, endpoint, UI panel, overlay,
+  import dialog, tests). Commit with `Closes #NNN` per project convention.
+- PR: opened against `release/0.6.0` when available; draft in the meantime.
+
+---
+
+## 9. Open questions (answered by the sidecar first run)
+
+1. Are emitted `section_ref`s reachable as `elem::${ref}` in the Neo4j graph built
+   by `TreeWriter`? I.e. is the `self_ref` the agent sees the same `self_ref` we
+   wrote to the graph? (Expected yes — both come from the same `DoclingDocument` —
+   but the sidecar on a real doc from SQLite will confirm in one run.)
+2. Hit rate of the agent: with `max_iterations=5` and `granite4:micro-h`, does it
+   converge, and how many sections does it actually visit? Determines if the overlay
+   ever has more than 1–2 marked nodes (and whether `REASONING_NEXT` edges are worth
+   the effort vs just node markers).
+3. Quality of `iteration.reason` — is it substantive enough to show in the panel, or
+   LLM filler we should hide? Sidecar output will tell.
+4. Fallback when no section headers exist (`RAGResult(iterations=[], converged=True,
+   answer=<full md>)` — see rag.py): what does the panel show? Probably a degraded
+   "no trace available, full-doc answer" state.

document-parser/api/analyses.py

@@ -155,7 +155,7 @@ async def delete_chunk(job_id: str, chunk_index: int, service: ServiceDep) -> li
     ]
 
 
-@router.delete("/{job_id}", status_code=204)
+@router.delete("/{job_id}", status_code=204, response_model=None)
 async def delete_analysis(job_id: str, service: ServiceDep) -> None:
     """Delete an analysis job."""
     deleted = await service.delete(job_id)

document-parser/api/documents.py

@@ -85,7 +85,7 @@ async def get_document(doc_id: str, service: ServiceDep) -> DocumentResponse:
     return _to_response(doc)
 
 
-@router.delete("/{doc_id}", status_code=204)
+@router.delete("/{doc_id}", status_code=204, response_model=None)
 async def delete_document(doc_id: str, service: ServiceDep) -> None:
     """Delete a document and its file."""
     deleted = await service.delete(doc_id)

document-parser/api/graph.py

@@ -0,0 +1,123 @@
+"""Graph API — returns a cytoscape-shaped view of the document structure.
+
+Two endpoints:
+- `/graph` — read from Neo4j. Rich graph (elements + chunks + pages + merges).
+  Requires the Maintain step (IngestionPipeline) to have run for the document.
+- `/reasoning-graph` — built on-the-fly from the SQLite `document_json` blob.
+  No Neo4j dependency. Lighter graph (no chunks) but enough to render the
+  reasoning-trace overlay on top of `GraphView`.
+"""
+
+from __future__ import annotations
+
+import logging
+
+from fastapi import APIRouter, HTTPException, Request
+from pydantic import BaseModel
+
+from infra.docling_graph import build_graph_payload
+from infra.neo4j import fetch_graph
+
+logger = logging.getLogger(__name__)
+router = APIRouter(prefix="/api/documents", tags=["graph"])
+
+MAX_PAGES = 200
+
+
+class GraphNode(BaseModel):
+    id: str
+    group: str
+    label: str | None = None
+
+    model_config = {"extra": "allow"}
+
+
+class GraphEdge(BaseModel):
+    id: str
+    source: str
+    target: str
+    type: str
+    order: int | None = None
+
+
+class GraphResponse(BaseModel):
+    doc_id: str
+    nodes: list[GraphNode]
+    edges: list[GraphEdge]
+    node_count: int
+    edge_count: int
+    truncated: bool
+    page_count: int
+
+
+@router.get("/{doc_id}/graph", response_model=GraphResponse)
+async def get_document_graph(doc_id: str, request: Request) -> GraphResponse:
+    neo = getattr(request.app.state, "neo4j", None)
+    if neo is None:
+        raise HTTPException(status_code=503, detail="Neo4j is not configured")
+
+    payload = await fetch_graph(neo, doc_id, max_pages=MAX_PAGES)
+    if payload is None:
+        raise HTTPException(status_code=404, detail=f"No graph for document {doc_id}")
+    if payload.truncated:
+        raise HTTPException(
+            status_code=413,
+            detail=(
+                f"Graph too large: document has {payload.page_count} pages "
+                f"(cap {MAX_PAGES}). Pagination ships in v0.6."
+            ),
+        )
+
+    return GraphResponse(
+        doc_id=payload.doc_id,
+        nodes=[GraphNode(**n) for n in payload.nodes],
+        edges=[GraphEdge(**e) for e in payload.edges],
+        node_count=payload.node_count,
+        edge_count=payload.edge_count,
+        truncated=payload.truncated,
+        page_count=payload.page_count,
+    )
+
+
+@router.get("/{doc_id}/reasoning-graph", response_model=GraphResponse)
+async def get_reasoning_graph(doc_id: str, request: Request) -> GraphResponse:
+    """Graph projection built from SQLite `document_json` — no Neo4j needed.
+
+    Serves the reasoning-trace viewer, which only needs the element/page/edge
+    structure to overlay iterations onto.
+    """
+    analysis_repo = getattr(request.app.state, "analysis_repo", None)
+    if analysis_repo is None:
+        raise HTTPException(status_code=500, detail="AnalysisRepository not wired")
+
+    latest = await analysis_repo.find_latest_completed_by_document(doc_id)
+    if latest is None or not latest.document_json:
+        raise HTTPException(
+            status_code=404,
+            detail=f"No completed analysis with document_json for {doc_id}",
+        )
+
+    payload = build_graph_payload(
+        latest.document_json,
+        doc_id=doc_id,
+        title=latest.document_filename or doc_id,
+        max_pages=MAX_PAGES,
+    )
+    if payload.truncated:
+        raise HTTPException(
+            status_code=413,
+            detail=(
+                f"Graph too large: document has {payload.page_count} pages "
+                f"(cap {MAX_PAGES}). Pagination ships in v0.6."
+            ),
+        )
+
+    return GraphResponse(
+        doc_id=payload.doc_id,
+        nodes=[GraphNode(**n) for n in payload.nodes],
+        edges=[GraphEdge(**e) for e in payload.edges],
+        node_count=payload.node_count,
+        edge_count=payload.edge_count,
+        truncated=payload.truncated,
+        page_count=payload.page_count,
+    )

document-parser/api/ingestion.py

@@ -74,7 +74,7 @@ async def ingest_analysis(
     )
 
 
-@router.delete("/{doc_id}", status_code=204)
+@router.delete("/{doc_id}", status_code=204, response_model=None)
 async def delete_ingested_document(doc_id: str, ingestion: IngestionDep) -> None:
     """Delete all indexed chunks for a document."""
     await ingestion.delete_document(doc_id)

document-parser/api/reasoning.py

@@ -0,0 +1,148 @@
+"""Reasoning API — live `docling-agent` runner (R&D).
+
+`POST /api/documents/:id/rag` invokes `docling-agent`'s Chunkless RAG loop
+against the stored `DoclingDocument` and returns a `RAGResult` in the same
+shape the v1 import dialog already consumes — so the frontend overlay code
+is fully reused.
+
+Constraints (docling-agent v0.1.0):
+- Backend is hard-wired to Ollama (`setup_local_session` in
+  `docling_agent/agent_models.py`). Set `OLLAMA_HOST` + `RAG_MODEL_ID` in the
+  environment. No OpenAI/WatsonX path without forking upstream.
+- We call the private `_rag_loop` because `DoclingRAGAgent.run()` wraps the
+  answer in a synthetic `DoclingDocument` and never returns the iteration
+  trace. This is brittle — track upstream for a public hook.
+- Sync blocking call offloaded to a thread so we don't stall the event loop.
+  No streaming at this step (see design doc §7 for v2 SSE plan).
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import os
+
+from fastapi import APIRouter, HTTPException, Request
+from pydantic import BaseModel
+
+from infra.settings import settings
+
+logger = logging.getLogger(__name__)
+router = APIRouter(prefix="/api/documents", tags=["reasoning"])
+
+
+class RagRunRequest(BaseModel):
+    query: str
+    # Optional per-run override; falls back to settings.rag_model_id.
+    model_id: str | None = None
+
+
+class RagIterationResponse(BaseModel):
+    iteration: int
+    section_ref: str
+    reason: str
+    section_text_length: int
+    can_answer: bool
+    response: str
+
+
+class RagResultResponse(BaseModel):
+    answer: str
+    iterations: list[RagIterationResponse]
+    converged: bool
+
+
+@router.post("/{doc_id}/rag", response_model=RagResultResponse)
+async def run_rag(doc_id: str, body: RagRunRequest, request: Request) -> RagResultResponse:
+    if not settings.rag_enabled:
+        raise HTTPException(status_code=503, detail="Live reasoning disabled (RAG_ENABLED=false)")
+
+    if not body.query.strip():
+        raise HTTPException(status_code=400, detail="Query must not be empty")
+
+    analysis_repo = getattr(request.app.state, "analysis_repo", None)
+    if analysis_repo is None:
+        raise HTTPException(status_code=500, detail="AnalysisRepository not wired")
+
+    latest = await analysis_repo.find_latest_completed_by_document(doc_id)
+    if latest is None or not latest.document_json:
+        raise HTTPException(
+            status_code=404,
+            detail=f"No completed analysis with document_json for {doc_id}",
+        )
+
+    # Lazy-import docling-agent so the backend boots even if the dep isn't
+    # installed (R&D group). If missing, return 503 with a clear install hint.
+    try:
+        from docling_agent.agents import DoclingRAGAgent
+        from docling_core.types.doc.document import DoclingDocument
+        from mellea.backends.model_ids import ModelIdentifier
+    except ImportError as e:
+        raise HTTPException(
+            status_code=503,
+            detail=f"docling-agent not installed: {e}. `pip install docling-agent mellea`.",
+        ) from e
+
+    # Ollama client reads OLLAMA_HOST at request time; set it per-call so the
+    # configured host takes effect without needing to restart the server.
+    os.environ["OLLAMA_HOST"] = settings.ollama_host
+    raw_model_id = body.model_id or settings.rag_model_id
+    # `DoclingRAGAgent` (pydantic) validates `model_id` strictly against the
+    # `ModelIdentifier` dataclass from Mellea. A raw string like "gpt-oss:20b"
+    # is rejected even though the Ollama backend itself would accept one.
+    # Wrap on the Ollama axis; add other axes here if we ever fork upstream to
+    # support non-Ollama backends.
+    model_id = ModelIdentifier(ollama_name=raw_model_id)
+
+    try:
+        doc = DoclingDocument.model_validate_json(latest.document_json)
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=f"Failed to parse document_json: {e}") from e
+
+    agent = DoclingRAGAgent(model_id=model_id, tools=[])
+    logger.info(
+        "RAG run: doc_id=%s model_id=%s ollama_host=%s query=%r",
+        doc_id,
+        model_id,
+        settings.ollama_host,
+        body.query[:120],
+    )
+
+    try:
+        # `_rag_loop` is a synchronous LLM-heavy call (N * model latency). Run
+        # it in a worker thread so concurrent requests don't block the loop.
+        result = await asyncio.to_thread(agent._rag_loop, query=body.query, doc=doc)
+    except IndexError as e:
+        # Known docling-agent bug: `_attempt_answer` / `_select_section` call
+        # `find_json_dicts(answer.value)[0]` without checking for an empty
+        # list. When the model can't produce a parseable JSON after 3
+        # rejection-sampling retries + 3 `select_from_failure` retries, the
+        # list is empty and the `[0]` crashes. It's model-dependent (some
+        # questions + some models trip it, others don't).
+        #
+        # Report as 502 Bad Gateway — the upstream LLM couldn't produce a
+        # usable response, not our fault — with a message the UI can show
+        # to the user so they pick another model or rephrase.
+        logger.warning(
+            "docling-agent produced no parseable JSON for doc=%s model=%s query=%r",
+            doc_id,
+            raw_model_id,
+            body.query[:120],
+        )
+        raise HTTPException(
+            status_code=502,
+            detail=(
+                f"The model '{raw_model_id}' couldn't produce a parseable "
+                "answer after retries. Try a different model (e.g. mistral-small3.2) "
+                "or rephrase the question."
+            ),
+        ) from e
+    except Exception as e:
+        logger.exception("RAG loop failed for doc %s", doc_id)
+        raise HTTPException(status_code=500, detail=f"RAG loop failed: {e}") from e
+
+    return RagResultResponse(
+        answer=result.answer,
+        iterations=[RagIterationResponse(**it.model_dump()) for it in result.iterations],
+        converged=result.converged,
+    )

document-parser/api/schemas.py

@@ -34,7 +34,13 @@ class HealthResponse(_CamelModel):
     database: str
     max_page_count: int | None = None
     max_file_size_mb: int | None = None
+    max_paste_image_size_mb: int | None = None
+    paste_allowed_image_types: list[str] = Field(default_factory=list)
     ingestion_available: bool = False
+    # True when the live-reasoning runner (docling-agent + Ollama) is
+    # available: RAG_ENABLED=true AND deps importable. Doesn't imply Ollama
+    # itself is reachable — that's checked per-call.
+    rag_available: bool = False
 
 
 class DocumentResponse(_CamelModel):

document-parser/domain/value_objects.py

@@ -8,6 +8,10 @@ from __future__ import annotations
 
 from dataclasses import dataclass, field
 
+# US Letter page dimensions (points) — fallback when page size is unknown
+DEFAULT_PAGE_WIDTH: float = 612.0
+DEFAULT_PAGE_HEIGHT: float = 792.0
+
 
 @dataclass(frozen=True)
 class PageElement:
@@ -15,6 +19,11 @@ class PageElement:
     bbox: list[float]
     content: str
     level: int = 0
+    # Docling `self_ref` ("#/texts/12", "#/tables/3", …). Empty for items
+    # that don't have one (rare — defensive default). Lets callers correlate
+    # a rendered bbox with the corresponding node in the graph without
+    # resorting to fuzzy bbox matching.
+    self_ref: str = ""
 
 
 @dataclass(frozen=True)
@@ -71,6 +80,14 @@ class ChunkBbox:
     bbox: list[float]  # [left, top, right, bottom] in TOPLEFT origin
 
 
+@dataclass(frozen=True)
+class ChunkDocItem:
+    """Source element referenced by a chunk. Enables Neo4j DERIVED_FROM edges."""
+
+    self_ref: str
+    label: str
+
+
 @dataclass(frozen=True)
 class ChunkResult:
     text: str
@@ -78,3 +95,4 @@ class ChunkResult:
     source_page: int | None = None
     token_count: int = 0
     bboxes: list[ChunkBbox] = field(default_factory=list)
+    doc_items: list[ChunkDocItem] = field(default_factory=list)

document-parser/infra/docling_graph.py

@@ -0,0 +1,178 @@
+"""Build a Cytoscape-shaped graph payload straight from a serialized
+`DoclingDocument` (i.e. the `document_json` blob stored in SQLite).
+
+Mirrors `infra.neo4j.queries.fetch_graph` so the frontend can reuse the same
+`GraphView` component — the only intentional difference is the absence of
+Chunk nodes / HAS_CHUNK / DERIVED_FROM edges, since chunks are a product of
+the Maintain step and don't exist in `document_json` alone.
+
+Used by the reasoning-trace viewer, which needs the structural graph to
+overlay iterations onto but does NOT need (and should not require) Neo4j.
+"""
+
+from __future__ import annotations
+
+import json
+from itertools import pairwise
+from typing import Any
+
+from infra.docling_tree import (
+    build_collapse_index,
+    dfs_order,
+    element_label,
+    is_inline_group,
+    iter_items,
+    iter_pages,
+    iter_provs,
+    parent_ref,
+)
+from infra.neo4j.queries import GraphPayload
+
+
+def _element_node(
+    doc_id: str,
+    item: dict[str, Any],
+    provs: list[dict[str, Any]],
+    *,
+    text_override: str | None = None,
+) -> dict[str, Any]:
+    first_page = provs[0].get("page_no") if provs else None
+    raw_text = text_override if text_override is not None else (item.get("text") or "")
+    return {
+        "id": f"elem::{item.get('self_ref')}",
+        "group": "element",
+        "label": element_label(item.get("label") or ""),
+        "docling_label": (item.get("label") or "").lower(),
+        "self_ref": item.get("self_ref"),
+        "text": raw_text[:200],
+        "prov_page": first_page,
+        "provs": provs,
+        "level": item.get("level"),
+        "doc_id": doc_id,
+    }
+
+
+def _page_node(doc_id: str, page: dict[str, Any]) -> dict[str, Any]:
+    return {
+        "id": f"page::{page.get('page_no')}",
+        "group": "page",
+        "page_no": page.get("page_no"),
+        "width": page.get("width"),
+        "height": page.get("height"),
+        "doc_id": doc_id,
+    }
+
+
+def _edge(source: str, target: str, edge_type: str, *, order: int | None = None) -> dict[str, Any]:
+    return {
+        "id": f"{edge_type}::{source}::{target}",
+        "source": source,
+        "target": target,
+        "type": edge_type,
+        "order": order,
+    }
+
+
+def build_graph_payload(
+    document_json: str,
+    *,
+    doc_id: str,
+    title: str | None = None,
+    max_pages: int = 200,
+) -> GraphPayload:
+    """Build a `GraphPayload` equivalent to `fetch_graph(neo4j, doc_id)` from
+    the raw `DoclingDocument` JSON.
+
+    Returns `truncated=True` with empty node/edge lists beyond `max_pages`, so
+    the caller can mirror the Neo4j endpoint's 413 behavior.
+    """
+    doc_data = json.loads(document_json)
+
+    pages_raw = list(iter_pages(doc_data))
+    page_count = len(pages_raw)
+    if page_count > max_pages:
+        return GraphPayload(
+            doc_id=doc_id,
+            nodes=[],
+            edges=[],
+            node_count=0,
+            edge_count=0,
+            truncated=True,
+            page_count=page_count,
+        )
+
+    nodes: list[dict[str, Any]] = []
+    edges: list[dict[str, Any]] = []
+
+    doc_node_id = f"doc::{doc_id}"
+    nodes.append(
+        {
+            "id": doc_node_id,
+            "group": "document",
+            "doc_id": doc_id,
+            "title": title,
+            # `stages_applied` is a Neo4j-only artifact; keep the key present
+            # for shape parity but leave it empty since SQLite doesn't track it.
+            "stages_applied": [],
+        }
+    )
+
+    # Page nodes.
+    for p in pages_raw:
+        nodes.append(_page_node(doc_id, p))
+
+    # Issue #197: collapse Docling noise — InlineGroup style runs and the
+    # internal text labels Docling extracts from pictures/charts.
+    skip_refs, inline_meta = build_collapse_index(doc_data)
+
+    # Element nodes + collect parent/body metadata for edges below. The
+    # `element_idx` mirrors TreeWriter's `enumerate(elements)` so PARENT_OF
+    # carries the same `order` the Neo4j projection does.
+    by_ref: dict[str, dict[str, Any]] = {}
+    element_idx = 0
+    for _, item in iter_items(doc_data):
+        ref = item.get("self_ref")
+        if not ref or ref in skip_refs:
+            continue
+        by_ref[ref] = item
+        if is_inline_group(item):
+            meta = inline_meta.get(ref, {"text": "", "provs": []})
+            provs = meta["provs"]
+            text_override: str | None = meta["text"]
+        else:
+            provs = iter_provs(item)
+            text_override = None
+        nodes.append(_element_node(doc_id, item, provs, text_override=text_override))
+
+        pref = parent_ref(item)
+        if pref == "#/body":
+            edges.append(_edge(doc_node_id, f"elem::{ref}", "HAS_ROOT"))
+        elif pref:
+            edges.append(_edge(f"elem::{pref}", f"elem::{ref}", "PARENT_OF", order=element_idx))
+
+        # ON_PAGE, dedup'd per (element, page) — matches the Neo4j query's
+        # DISTINCT projection through Provenance.
+        seen_pages: set[int] = set()
+        for prov in provs:
+            page_no = prov.get("page_no")
+            if page_no is None or page_no in seen_pages:
+                continue
+            seen_pages.add(page_no)
+            edges.append(_edge(f"elem::{ref}", f"page::{page_no}", "ON_PAGE"))
+
+        element_idx += 1
+
+    # NEXT chain (DFS pre-order from body), inline-group children skipped.
+    for a, b in pairwise(dfs_order(doc_data, skip_refs)):
+        if a in by_ref and b in by_ref:
+            edges.append(_edge(f"elem::{a}", f"elem::{b}", "NEXT"))
+
+    return GraphPayload(
+        doc_id=doc_id,
+        nodes=nodes,
+        edges=edges,
+        node_count=len(nodes),
+        edge_count=len(edges),
+        truncated=False,
+        page_count=page_count,
+    )

document-parser/infra/docling_tree.py

@@ -0,0 +1,276 @@
+"""Pure helpers over a serialized `DoclingDocument` dict.
+
+No I/O, no Neo4j. Shared between:
+- `infra.neo4j.tree_writer` — persists the tree into Neo4j during the Maintain
+  step (IngestionPipeline).
+- `infra.docling_graph` — builds an in-memory `GraphPayload` from the SQLite
+  `document_json` blob for the reasoning-trace viewer.
+
+Keep this module the single source of truth for how we read Docling's own
+structure, so the two consumers can't drift.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Iterator
+from typing import Any
+
+# Docling label -> specific Neo4j/Cytoscape label. Every element carries the
+# generic :Element tag too. Kept 1:1 with docling-core's label taxonomy so the
+# projection is a faithful mirror of the DoclingDocument.
+LABEL_MAP: dict[str, str] = {
+    "section_header": "SectionHeader",
+    "title": "SectionHeader",
+    "paragraph": "Paragraph",
+    "text": "Paragraph",
+    "list_item": "ListItem",
+    "list": "List",  # distinct from :ListItem — a list is a container
+    "inline": "Paragraph",  # see issue #197 — collapsed into one paragraph node
+    "table": "Table",
+    "picture": "Figure",
+    "formula": "Formula",
+    "code": "Code",
+    "caption": "Caption",
+    "footnote": "Footnote",
+    "page_header": "PageHeader",
+    "page_footer": "PageFooter",
+    "key_value_area": "KeyValueArea",
+    "form_area": "FormArea",
+    "document_index": "DocumentIndex",
+}
+DEFAULT_LABEL = "TextElement"
+
+
+def element_label(docling_label: str) -> str:
+    return LABEL_MAP.get(docling_label.lower(), DEFAULT_LABEL)
+
+
+def is_inline_group(item: dict[str, Any]) -> bool:
+    """True iff `item` is a Docling InlineGroup (paragraph of mixed style runs).
+
+    Docling represents an inline-styled paragraph as one entry in `groups[]`
+    (label `inline`) plus N entries in `texts[]` (label `text`), one per style
+    run. We collapse them into a single Paragraph projection — see #197.
+    """
+    return (item.get("label") or "").lower() == "inline"
+
+
+def is_picture(item: dict[str, Any]) -> bool:
+    """True iff `item` is a Docling PictureItem (figure or chart).
+
+    A `picture` keeps its node in the graph (it IS the figure), but its
+    `children` — internal text labels extracted from a flowchart, diagram,
+    chart axis labels — are noise for graph readability and are skipped.
+    Captions live in a separate `captions` field on the picture, not in
+    `children`, so they are unaffected by this skip.
+    """
+    return (item.get("label") or "").lower() in {"picture", "chart"}
+
+
+def iter_items(doc_data: dict[str, Any]) -> Iterator[tuple[str, dict[str, Any]]]:
+    """Yield every item from texts/tables/pictures/groups with its source list key."""
+    for key in ("texts", "tables", "pictures", "groups"):
+        for item in doc_data.get(key, []) or []:
+            yield key, item
+
+
+def parent_ref(item: dict[str, Any]) -> str | None:
+    parent = item.get("parent")
+    if isinstance(parent, dict):
+        return parent.get("$ref") or parent.get("cref")
+    return None
+
+
+def iter_provs(item: dict[str, Any]) -> list[dict[str, Any]]:
+    """Flatten a Docling item's `prov[]` into a list of dict rows.
+
+    A single item may have multiple provs when it spans page breaks or appears
+    more than once in the layout. The returned dicts carry the original index
+    under `order` so sequence is preserved.
+    """
+    provs = item.get("prov") or []
+    rows: list[dict[str, Any]] = []
+    for idx, p in enumerate(provs):
+        bbox = p.get("bbox")
+        l_, t_, r_, b_ = 0.0, 0.0, 0.0, 0.0
+        if isinstance(bbox, dict):
+            l_ = float(bbox.get("l", 0.0) or 0.0)
+            t_ = float(bbox.get("t", 0.0) or 0.0)
+            r_ = float(bbox.get("r", 0.0) or 0.0)
+            b_ = float(bbox.get("b", 0.0) or 0.0)
+        elif isinstance(bbox, (list, tuple)) and len(bbox) >= 4:
+            l_, t_, r_, b_ = (float(x) for x in bbox[:4])
+        coord_origin = (bbox.get("coord_origin") if isinstance(bbox, dict) else None) or "TOPLEFT"
+        charspan = p.get("charspan") or []
+        rows.append(
+            {
+                "order": idx,
+                "page_no": p.get("page_no"),
+                "bbox_l": l_,
+                "bbox_t": t_,
+                "bbox_r": r_,
+                "bbox_b": b_,
+                "coord_origin": coord_origin,
+                "charspan_start": int(charspan[0]) if len(charspan) >= 1 else None,
+                "charspan_end": int(charspan[1]) if len(charspan) >= 2 else None,
+            }
+        )
+    return rows
+
+
+def dfs_order(doc_data: dict[str, Any], skip_refs: set[str] | None = None) -> list[str]:
+    """Return `self_ref`s in reading order (DFS pre-order from body).
+
+    `skip_refs` (typically the set returned by `build_inline_index`) is omitted
+    from the chain. Inline groups themselves are emitted but the walk does not
+    recurse into their style-run children, so the resulting order references
+    only nodes that survive the InlineGroup collapse.
+    """
+    skip = skip_refs or set()
+    by_ref: dict[str, dict[str, Any]] = {}
+    for _, item in iter_items(doc_data):
+        ref = item.get("self_ref")
+        if ref:
+            by_ref[ref] = item
+    body = doc_data.get("body") or {}
+    order: list[str] = []
+
+    def walk(children: list[dict[str, Any]] | None) -> None:
+        if not children:
+            return
+        for ch in children:
+            ref = ch.get("$ref") or ch.get("cref")
+            if not ref or ref in skip:
+                continue
+            order.append(ref)
+            child = by_ref.get(ref)
+            if child and not is_inline_group(child):
+                walk(child.get("children"))
+
+    walk(body.get("children"))
+    return order
+
+
+def build_collapse_index(
+    doc_data: dict[str, Any],
+) -> tuple[set[str], dict[str, dict[str, Any]]]:
+    """Pre-compute graph-projection collapses for a serialized DoclingDocument.
+
+    Two cases produce noise nodes if mirrored 1:1 — see issue #197:
+
+    1. **InlineGroup** — Docling emits one `groups[]` entry (label `inline`)
+       plus N `texts[]` style runs. We collapse the children into the group,
+       which is then projected as a single `:Paragraph` with concatenated
+       text and the union of children's provs.
+    2. **Picture / Chart** — internal text labels extracted from flowcharts,
+       diagrams or chart axes hang off the picture's `children`. The picture
+       node itself stays, but its descendants are skipped so the graph isn't
+       drowned in dozens of tiny labels.
+
+    Returns `(skip_refs, inline_meta)`:
+
+    - `skip_refs`: every `self_ref` to drop from element / edge projections.
+    - `inline_meta[group_ref]`: `{"text": str, "provs": list[dict]}` —
+      override values for the inline group projection. Pictures don't have
+      an entry here; they keep their own text/prov.
+    """
+    by_ref: dict[str, dict[str, Any]] = {}
+    for _, item in iter_items(doc_data):
+        ref = item.get("self_ref")
+        if ref:
+            by_ref[ref] = item
+
+    skip_refs: set[str] = set()
+    inline_meta: dict[str, dict[str, Any]] = {}
+
+    for item in by_ref.values():
+        ref = item.get("self_ref") or ""
+        if not ref:
+            continue
+        if is_inline_group(item):
+            text_parts, provs = _collect_inline_descendants(ref, by_ref, skip_refs)
+            # Re-index prov order so the resulting :Provenance nodes are 0..N-1
+            # contiguous instead of carrying each child's individual indices.
+            for idx, prov in enumerate(provs):
+                prov["order"] = idx
+            inline_meta[ref] = {
+                "text": " ".join(text_parts),
+                "provs": provs,
+            }
+        elif is_picture(item):
+            _collect_descendants(ref, by_ref, skip_refs)
+
+    return skip_refs, inline_meta
+
+
+def _collect_descendants(
+    root_ref: str,
+    by_ref: dict[str, dict[str, Any]],
+    skip_refs: set[str],
+) -> None:
+    """DFS `root_ref`'s subtree and add every descendant to `skip_refs`.
+
+    Used for picture children — we just want them dropped, not aggregated.
+    """
+
+    def walk(ref: str) -> None:
+        item = by_ref.get(ref)
+        if item is None:
+            return
+        for ch in item.get("children") or []:
+            child_ref = ch.get("$ref") or ch.get("cref")
+            if not child_ref or child_ref in skip_refs:
+                continue
+            skip_refs.add(child_ref)
+            walk(child_ref)
+
+    walk(root_ref)
+
+
+def _collect_inline_descendants(
+    group_ref: str,
+    by_ref: dict[str, dict[str, Any]],
+    skip_refs: set[str],
+) -> tuple[list[str], list[dict[str, Any]]]:
+    """DFS an inline group's subtree, returning its text parts and provs in
+    document order. `skip_refs` is mutated with every visited descendant."""
+    text_parts: list[str] = []
+    provs: list[dict[str, Any]] = []
+
+    def walk(ref: str) -> None:
+        item = by_ref.get(ref)
+        if item is None:
+            return
+        for ch in item.get("children") or []:
+            child_ref = ch.get("$ref") or ch.get("cref")
+            if not child_ref or child_ref in skip_refs:
+                continue
+            skip_refs.add(child_ref)
+            child = by_ref.get(child_ref)
+            if child is None:
+                continue
+            if is_inline_group(child):
+                walk(child_ref)
+                continue
+            text = child.get("text") or ""
+            if text:
+                text_parts.append(text)
+            provs.extend(iter_provs(child))
+
+    walk(group_ref)
+    return text_parts, provs
+
+
+def iter_pages(doc_data: dict[str, Any]) -> Iterator[dict[str, Any]]:
+    """Yield page dicts with `page_no`, `width`, `height` from the `pages` map."""
+    for page_no_str, page_obj in (doc_data.get("pages") or {}).items():
+        try:
+            page_no = int(page_no_str)
+        except (TypeError, ValueError):
+            continue
+        size = (page_obj or {}).get("size") or {}
+        yield {
+            "page_no": page_no,
+            "width": size.get("width"),
+            "height": size.get("height"),
+        }

document-parser/infra/local_chunker.py

@@ -15,7 +15,7 @@ from docling_core.transforms.chunker import HierarchicalChunker
 from docling_core.transforms.chunker.hybrid_chunker import HybridChunker
 from docling_core.types.doc.document import DoclingDocument
 
-from domain.value_objects import ChunkBbox, ChunkingOptions, ChunkResult
+from domain.value_objects import ChunkBbox, ChunkDocItem, ChunkingOptions, ChunkResult
 from infra.bbox import EMPTY_BBOX, to_topleft_list
 
 logger = logging.getLogger(__name__)
@@ -39,9 +39,18 @@ def _chunk_sync(document_json: str, options: ChunkingOptions) -> list[ChunkResul
         source_page = None
         token_count = 0
         bboxes: list[ChunkBbox] = []
+        doc_items: list[ChunkDocItem] = []
 
         if hasattr(chunk, "meta") and chunk.meta and chunk.meta.doc_items:
             for doc_item in chunk.meta.doc_items:
+                ref = getattr(doc_item, "self_ref", None)
+                if ref:
+                    doc_items.append(
+                        ChunkDocItem(
+                            self_ref=ref,
+                            label=str(getattr(doc_item, "label", "") or ""),
+                        )
+                    )
                 if not hasattr(doc_item, "prov") or not doc_item.prov:
                     continue
                 for prov in doc_item.prov:
@@ -67,6 +76,7 @@ def _chunk_sync(document_json: str, options: ChunkingOptions) -> list[ChunkResul
                 source_page=source_page,
                 token_count=token_count,
                 bboxes=bboxes,
+                doc_items=doc_items,
             )
         )
 

document-parser/infra/local_converter.py

@@ -36,6 +36,8 @@ from docling_core.types.doc import (
 )
 
 from domain.value_objects import (
+    DEFAULT_PAGE_HEIGHT,
+    DEFAULT_PAGE_WIDTH,
     ConversionOptions,
     ConversionResult,
     PageDetail,
@@ -50,10 +52,6 @@ logger = logging.getLogger(__name__)
 # Uses a timeout to prevent a frozen conversion from blocking all others.
 _converter_lock = threading.Lock()
 
-# US Letter page dimensions (points) — fallback when page size is unknown
-_DEFAULT_PAGE_WIDTH = 612.0
-_DEFAULT_PAGE_HEIGHT = 792.0
-
 # Default converter (lazy-init on first request)
 _default_converter: DoclingConverter | None = None
 
@@ -175,11 +173,11 @@ def _process_content_item(
                 logger.warning(
                     "Page %d not found in document metadata — using US Letter fallback (%sx%s pt)",
                     page_no,
-                    _DEFAULT_PAGE_WIDTH,
+                    DEFAULT_PAGE_WIDTH,
-                    _DEFAULT_PAGE_HEIGHT,
+                    DEFAULT_PAGE_HEIGHT,
                 )
                 pages[page_no] = PageDetail(
-                    page_number=page_no, width=_DEFAULT_PAGE_WIDTH, height=_DEFAULT_PAGE_HEIGHT
+                    page_number=page_no, width=DEFAULT_PAGE_WIDTH, height=DEFAULT_PAGE_HEIGHT
                 )
 
             page_height = pages[page_no].height
@@ -196,7 +194,13 @@ def _process_content_item(
                     content = item.export_to_markdown()
 
             pages[page_no].elements.append(
-                PageElement(type=element_type, bbox=bbox, content=content, level=level)
+                PageElement(
+                    type=element_type,
+                    bbox=bbox,
+                    content=content,
+                    level=level,
+                    self_ref=getattr(item, "self_ref", "") or "",
+                )
             )
         except (AttributeError, KeyError, TypeError, ValueError):
             logger.warning(
@@ -248,10 +252,10 @@ def _convert_sync(
         pages_detail = [
             PageDetail(
                 page_number=i + 1,
-                width=doc.pages[i + 1].size.width if (i + 1) in doc.pages else _DEFAULT_PAGE_WIDTH,
+                width=doc.pages[i + 1].size.width if (i + 1) in doc.pages else DEFAULT_PAGE_WIDTH,
                 height=doc.pages[i + 1].size.height
                 if (i + 1) in doc.pages
-                else _DEFAULT_PAGE_HEIGHT,
+                else DEFAULT_PAGE_HEIGHT,
             )
             for i in range(page_count)
         ]

document-parser/infra/neo4j/__init__.py

@@ -0,0 +1,31 @@
+"""Neo4j storage adapter — graph-native document structure.
+
+Provides a thin driver wrapper, idempotent schema bootstrap, and
+walkers between DoclingDocument and the graph model.
+"""
+
+from infra.neo4j.chunk_writer import ChunkWriteResult, write_chunks
+from infra.neo4j.driver import Neo4jDriver, close_driver, get_driver
+from infra.neo4j.queries import fetch_graph
+from infra.neo4j.schema import bootstrap_schema
+from infra.neo4j.tree_reader import (
+    delete_document,
+    document_exists,
+    read_document_json,
+)
+from infra.neo4j.tree_writer import TreeWriteResult, write_document
+
+__all__ = [
+    "ChunkWriteResult",
+    "Neo4jDriver",
+    "TreeWriteResult",
+    "bootstrap_schema",
+    "close_driver",
+    "delete_document",
+    "document_exists",
+    "fetch_graph",
+    "get_driver",
+    "read_document_json",
+    "write_chunks",
+    "write_document",
+]

document-parser/infra/neo4j/chunk_writer.py

@@ -0,0 +1,134 @@
+"""ChunkWriter — push chunk nodes and DERIVED_FROM edges to Neo4j.
+
+Embeddings stay in OpenSearch. Each :Chunk node carries a chunk_index so the
+OpenSearch entry can be retrieved via (doc_id, chunk_index). The
+`embedding_ref` property is reserved for a future vector-store id (not used
+in v0.5 — OpenSearch indexes by doc_id+chunk_index already).
+
+When chunks carry `doc_items` provenance (list of `self_ref` strings), we
+create `(:Chunk)-[:DERIVED_FROM]->(:Element)` links so that queries can go
+from a chunk back to its source elements. Chunks without doc_items get no
+back-links but are still persisted.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Any
+
+if TYPE_CHECKING:
+    from infra.neo4j.driver import Neo4jDriver
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class ChunkWriteResult:
+    doc_id: str
+    chunks_written: int
+    derived_from_edges: int
+
+
+def _chunk_id(doc_id: str, index: int) -> str:
+    return f"{doc_id}::chunk::{index}"
+
+
+async def write_chunks(
+    neo: Neo4jDriver,
+    *,
+    doc_id: str,
+    chunks_json: str,
+) -> ChunkWriteResult:
+    """Persist chunks for `doc_id`. Wipes prior chunks first (idempotent)."""
+    chunks: list[dict[str, Any]] = json.loads(chunks_json)
+    active = [c for c in chunks if not c.get("deleted")]
+
+    chunk_rows: list[dict[str, Any]] = []
+    derived_rows: list[dict[str, Any]] = []
+    for idx, c in enumerate(active):
+        cid = _chunk_id(doc_id, idx)
+        chunk_rows.append(
+            {
+                "id": cid,
+                "doc_id": doc_id,
+                "text": c.get("text") or "",
+                "chunk_index": idx,
+                "token_count": c.get("tokenCount") or 0,
+                "embedding_ref": "",
+            }
+        )
+        for item in c.get("docItems") or []:
+            ref = item.get("selfRef") if isinstance(item, dict) else None
+            if ref:
+                derived_rows.append({"chunk_id": cid, "doc_id": doc_id, "self_ref": ref})
+
+    async with (
+        neo.driver.session(database=neo.database) as session,
+        await session.begin_transaction() as tx,
+    ):
+        # Replace existing chunks.
+        await tx.run(
+            """
+                MATCH (d:Document {id: $doc_id})-[:HAS_CHUNK]->(c:Chunk)
+                DETACH DELETE c
+                """,
+            doc_id=doc_id,
+        )
+        await tx.run("MATCH (c:Chunk {doc_id: $doc_id}) DETACH DELETE c", doc_id=doc_id)
+
+        if chunk_rows:
+            await tx.run(
+                """
+                    MATCH (d:Document {id: $doc_id})
+                    UNWIND $rows AS r
+                    CREATE (c:Chunk {
+                      id: r.id,
+                      doc_id: r.doc_id,
+                      text: r.text,
+                      chunk_index: r.chunk_index,
+                      token_count: r.token_count,
+                      embedding_ref: r.embedding_ref
+                    })
+                    MERGE (d)-[:HAS_CHUNK]->(c)
+                    """,
+                doc_id=doc_id,
+                rows=chunk_rows,
+            )
+
+        if derived_rows:
+            await tx.run(
+                """
+                    UNWIND $rows AS r
+                    MATCH (c:Chunk {id: r.chunk_id})
+                    MATCH (e:Element {doc_id: r.doc_id, self_ref: r.self_ref})
+                    MERGE (c)-[:DERIVED_FROM]->(e)
+                    """,
+                rows=derived_rows,
+            )
+
+        # Flag the Document with the new stage.
+        await tx.run(
+            """
+                MATCH (d:Document {id: $doc_id})
+                SET d.stages_applied = [s IN coalesce(d.stages_applied, []) WHERE s <> 'chunks']
+                                       + ['chunks'],
+                    d.last_chunk_write = datetime()
+                """,
+            doc_id=doc_id,
+        )
+
+        await tx.commit()
+
+    logger.info(
+        "Neo4j: wrote %d chunks (%d DERIVED_FROM) for doc %s",
+        len(chunk_rows),
+        len(derived_rows),
+        doc_id,
+    )
+    return ChunkWriteResult(
+        doc_id=doc_id,
+        chunks_written=len(chunk_rows),
+        derived_from_edges=len(derived_rows),
+    )

document-parser/infra/neo4j/driver.py

@@ -0,0 +1,48 @@
+"""Async Neo4j driver wrapper.
+
+Owns a single `AsyncDriver` per process. Callers acquire it via
+`get_driver()` and must call `close_driver()` at shutdown.
+"""
+
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass
+
+from neo4j import AsyncDriver, AsyncGraphDatabase
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass(frozen=True)
+class Neo4jDriver:
+    driver: AsyncDriver
+    database: str = "neo4j"
+
+
+_instance: Neo4jDriver | None = None
+
+
+async def get_driver(uri: str, user: str, password: str, database: str = "neo4j") -> Neo4jDriver:
+    """Return the process-wide driver, creating it on first call.
+
+    Verifies connectivity once at creation — raises if the server is unreachable.
+    """
+    global _instance
+    if _instance is not None:
+        return _instance
+
+    driver = AsyncGraphDatabase.driver(uri, auth=(user, password))
+    await driver.verify_connectivity()
+    logger.info("Neo4j driver connected to %s (db=%s)", uri, database)
+    _instance = Neo4jDriver(driver=driver, database=database)
+    return _instance
+
+
+async def close_driver() -> None:
+    global _instance
+    if _instance is None:
+        return
+    await _instance.driver.close()
+    _instance = None
+    logger.info("Neo4j driver closed")

document-parser/infra/neo4j/queries.py

@@ -0,0 +1,269 @@
+"""Reusable Cypher queries — kept out of the API layer for reuse + testing."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Any
+
+if TYPE_CHECKING:
+    from infra.neo4j.driver import Neo4jDriver
+
+
+@dataclass
+class GraphPayload:
+    doc_id: str
+    nodes: list[dict[str, Any]]
+    edges: list[dict[str, Any]]
+    node_count: int
+    edge_count: int
+    truncated: bool
+    page_count: int
+
+
+# Full graph for one doc: Document + Elements + Pages + Chunks and their edges.
+# Each node/edge type is collected inside its own CALL {} subquery so every
+# block contributes a single row — avoids the cartesian product that chained
+# OPTIONAL MATCH on 6+ edge types would produce (hangs on multi-page docs).
+# See: https://neo4j.com/developer/kb/using-subqueries-to-control-the-scope-of-aggregations/
+#
+# Provenance nodes (post-v0.6 refactor) are NOT returned as top-level graph
+# nodes — they're metadata of their owning Element. We aggregate them inline
+# per element, and derive a dedup'd ON_PAGE edge set from them.
+_FETCH_GRAPH = """
+MATCH (d:Document {id: $doc_id})
+CALL {
+  WITH d
+  MATCH (e:Element {doc_id: d.id})
+  OPTIONAL MATCH (e)-[hp:HAS_PROV]->(pv:Provenance)
+  WITH e, pv ORDER BY hp.order
+  WITH e,
+    collect(
+      CASE WHEN pv IS NULL THEN NULL ELSE {
+        order: pv.prov_order,
+        page_no: pv.page_no,
+        bbox_l: pv.bbox_l, bbox_t: pv.bbox_t,
+        bbox_r: pv.bbox_r, bbox_b: pv.bbox_b,
+        coord_origin: pv.coord_origin,
+        charspan_start: pv.charspan_start,
+        charspan_end: pv.charspan_end
+      } END
+    ) AS all_provs
+  RETURN collect({element: e, provs: [p IN all_provs WHERE p IS NOT NULL]}) AS elements
+}
+CALL { WITH d MATCH (p:Page {doc_id: d.id})    RETURN collect(p) AS pages }
+CALL { WITH d MATCH (c:Chunk {doc_id: d.id})   RETURN collect(c) AS chunks }
+CALL {
+  WITH d
+  MATCH (pe:Element {doc_id: d.id})-[r:PARENT_OF]->(ce:Element)
+  RETURN collect({from: pe.self_ref, to: ce.self_ref, order: r.order, type: 'PARENT_OF'}) AS parent_edges
+}
+CALL {
+  WITH d
+  MATCH (a:Element {doc_id: d.id})-[:NEXT]->(b:Element)
+  RETURN collect({from: a.self_ref, to: b.self_ref, type: 'NEXT'}) AS next_edges
+}
+CALL {
+  WITH d
+  // ON_PAGE is stored on Provenance since v0.6; surface it at the Element
+  // level (dedup'd per Element/Page pair) for the Cytoscape viz.
+  MATCH (er:Element {doc_id: d.id})-[:HAS_PROV]->(:Provenance)-[:ON_PAGE]->(pr:Page)
+  WITH DISTINCT er, pr
+  RETURN collect({from: er.self_ref, to: pr.page_no, type: 'ON_PAGE'}) AS on_page_edges
+}
+CALL {
+  WITH d
+  MATCH (d)-[:HAS_ROOT]->(rr:Element)
+  RETURN collect({from: d.id, to: rr.self_ref, type: 'HAS_ROOT'}) AS has_root_edges
+}
+CALL {
+  WITH d
+  MATCH (d)-[:HAS_CHUNK]->(rc:Chunk)
+  RETURN collect({from: d.id, to: rc.id, type: 'HAS_CHUNK'}) AS has_chunk_edges
+}
+CALL {
+  WITH d
+  MATCH (cc:Chunk {doc_id: d.id})-[:DERIVED_FROM]->(ee:Element)
+  RETURN collect({from: cc.id, to: ee.self_ref, type: 'DERIVED_FROM'}) AS derived_from_edges
+}
+RETURN d AS document, elements, pages, chunks,
+       parent_edges, next_edges, on_page_edges,
+       has_root_edges, has_chunk_edges, derived_from_edges
+"""
+
+
+def _element_node(
+    doc_id: str, e: dict[str, Any], provs: list[dict[str, Any]] | None = None
+) -> dict[str, Any]:
+    # Determine the specific element label: Neo4j returns it via labels(e) on the
+    # driver side; when we project nodes via RETURN, the driver wraps them as Node
+    # objects, so we convert below.
+    first_page: int | None = None
+    if provs:
+        # Convenience: the first provenance's page — the old `prov_page` property,
+        # useful for label rendering in Cytoscape. Full list is in `provs`.
+        first_page = provs[0].get("page_no")
+    return {
+        "id": f"elem::{e.get('self_ref')}",
+        "group": "element",
+        "docling_label": e.get("docling_label"),
+        "self_ref": e.get("self_ref"),
+        "text": (e.get("text") or "")[:200],
+        "prov_page": first_page,
+        "provs": provs or [],
+        "level": e.get("level"),
+        "doc_id": doc_id,
+    }
+
+
+def _page_node(doc_id: str, p: dict[str, Any]) -> dict[str, Any]:
+    return {
+        "id": f"page::{p.get('page_no')}",
+        "group": "page",
+        "page_no": p.get("page_no"),
+        "width": p.get("width"),
+        "height": p.get("height"),
+        "doc_id": doc_id,
+    }
+
+
+def _chunk_node(p: dict[str, Any]) -> dict[str, Any]:
+    return {
+        "id": f"chunk::{p.get('id')}",
+        "group": "chunk",
+        "chunk_index": p.get("chunk_index"),
+        "text": (p.get("text") or "")[:200],
+        "token_count": p.get("token_count"),
+    }
+
+
+def _edge_id(from_id: str, to_id: str, edge_type: str) -> str:
+    return f"{edge_type}::{from_id}::{to_id}"
+
+
+async def fetch_graph(
+    neo: Neo4jDriver,
+    doc_id: str,
+    *,
+    max_pages: int = 200,
+) -> GraphPayload | None:
+    """Return the full graph for a document, or None if the document is unknown.
+
+    Enforces the page cap from design §8.4: beyond `max_pages`, returns a
+    `truncated=True` payload with empty node/edge lists so the caller can
+    surface a clean error (HTTP 413) to the UI.
+    """
+    async with neo.driver.session(database=neo.database) as session:
+        page_count_result = await session.run(
+            "MATCH (p:Page {doc_id: $doc_id}) RETURN count(p) AS n",
+            doc_id=doc_id,
+        )
+        pc_record = await page_count_result.single()
+        if pc_record is None:
+            return None
+        page_count = int(pc_record["n"])
+
+        exists_result = await session.run(
+            "MATCH (d:Document {id: $doc_id}) RETURN count(d) AS n",
+            doc_id=doc_id,
+        )
+        exists_record = await exists_result.single()
+        if not exists_record or exists_record["n"] == 0:
+            return None
+
+        if page_count > max_pages:
+            return GraphPayload(
+                doc_id=doc_id,
+                nodes=[],
+                edges=[],
+                node_count=0,
+                edge_count=0,
+                truncated=True,
+                page_count=page_count,
+            )
+
+        result = await session.run(_FETCH_GRAPH, doc_id=doc_id)
+        record = await result.single()
+
+    nodes: list[dict[str, Any]] = []
+    edges: list[dict[str, Any]] = []
+
+    if record is None:
+        return None
+
+    # Document node.
+    doc_node = record["document"]
+    if doc_node is not None:
+        nodes.append(
+            {
+                "id": f"doc::{doc_id}",
+                "group": "document",
+                "doc_id": doc_id,
+                "title": doc_node.get("title"),
+                "stages_applied": doc_node.get("stages_applied"),
+            }
+        )
+
+    # Element nodes, keeping the specific label (:SectionHeader, etc.).
+    # Each row is a {element, provs} dict from the CALL above; provs is a list
+    # of per-provenance dicts in original order.
+    for row in record["elements"] or []:
+        if row is None:
+            continue
+        e = row.get("element") if isinstance(row, dict) else None
+        if e is None:
+            continue
+        provs = [p for p in (row.get("provs") or []) if p is not None]
+        labels = [label for label in e.labels if label != "Element"]
+        node = _element_node(doc_id, dict(e), provs=provs)
+        node["label"] = labels[0] if labels else "TextElement"
+        nodes.append(node)
+
+    # Pages.
+    for p in record["pages"] or []:
+        if p is None:
+            continue
+        nodes.append(_page_node(doc_id, dict(p)))
+
+    # Chunks.
+    for c in record["chunks"] or []:
+        if c is None:
+            continue
+        nodes.append(_chunk_node(dict(c)))
+
+    # Edges — filter out rows whose from/to is null (OPTIONAL MATCH can yield them).
+    def _push_element_edge(e: dict[str, Any], from_prefix: str, to_prefix: str) -> None:
+        frm, to = e.get("from"), e.get("to")
+        if frm is None or to is None:
+            return
+        edges.append(
+            {
+                "id": _edge_id(f"{from_prefix}{frm}", f"{to_prefix}{to}", e["type"]),
+                "source": f"{from_prefix}{frm}",
+                "target": f"{to_prefix}{to}",
+                "type": e["type"],
+                "order": e.get("order"),
+            }
+        )
+
+    for e in record["parent_edges"] or []:
+        _push_element_edge(e, "elem::", "elem::")
+    for e in record["next_edges"] or []:
+        _push_element_edge(e, "elem::", "elem::")
+    for e in record["on_page_edges"] or []:
+        _push_element_edge(e, "elem::", "page::")
+    for e in record["has_root_edges"] or []:
+        _push_element_edge(e, "doc::", "elem::")
+    for e in record["has_chunk_edges"] or []:
+        _push_element_edge(e, "doc::", "chunk::")
+    for e in record["derived_from_edges"] or []:
+        _push_element_edge(e, "chunk::", "elem::")
+
+    return GraphPayload(
+        doc_id=doc_id,
+        nodes=nodes,
+        edges=edges,
+        node_count=len(nodes),
+        edge_count=len(edges),
+        truncated=False,
+        page_count=page_count,
+    )

document-parser/infra/neo4j/schema.py

@@ -0,0 +1,56 @@
+"""Idempotent Neo4j schema bootstrap.
+
+Runs at backend startup. All statements use `IF NOT EXISTS`, so calling
+this multiple times is safe — it's the contract integration tests rely on.
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from infra.neo4j.driver import Neo4jDriver
+
+logger = logging.getLogger(__name__)
+
+
+CONSTRAINTS: tuple[str, ...] = (
+    "CREATE CONSTRAINT document_id IF NOT EXISTS FOR (d:Document) REQUIRE d.id IS UNIQUE",
+    "CREATE CONSTRAINT element_composite IF NOT EXISTS "
+    "FOR (e:Element) REQUIRE (e.doc_id, e.self_ref) IS UNIQUE",
+    "CREATE CONSTRAINT page_composite IF NOT EXISTS "
+    "FOR (p:Page) REQUIRE (p.doc_id, p.page_no) IS UNIQUE",
+    "CREATE CONSTRAINT chunk_id IF NOT EXISTS FOR (c:Chunk) REQUIRE c.id IS UNIQUE",
+)
+
+INDEXES: tuple[str, ...] = (
+    "CREATE INDEX element_doc IF NOT EXISTS FOR (e:Element) ON (e.doc_id)",
+    "CREATE INDEX chunk_doc IF NOT EXISTS FOR (c:Chunk) ON (c.doc_id)",
+    # Reasoning tunnel / bbox-highlight: looking up a Provenance by its owner
+    # element is a hot path (one lookup per visited section). Composite index
+    # avoids a full scan of every Provenance in the DB.
+    "CREATE INDEX provenance_element IF NOT EXISTS "
+    "FOR (pv:Provenance) ON (pv.doc_id, pv.element_ref)",
+    "CREATE INDEX provenance_page IF NOT EXISTS FOR (pv:Provenance) ON (pv.doc_id, pv.page_no)",
+)
+
+FULLTEXT_INDEXES: tuple[str, ...] = (
+    "CREATE FULLTEXT INDEX element_text IF NOT EXISTS FOR (e:Element) ON EACH [e.text]",
+)
+
+
+async def bootstrap_schema(neo: Neo4jDriver) -> None:
+    """Create constraints and indexes required by the graph model.
+
+    Idempotent: safe to call on every startup.
+    """
+    async with neo.driver.session(database=neo.database) as session:
+        for stmt in (*CONSTRAINTS, *INDEXES, *FULLTEXT_INDEXES):
+            await session.run(stmt)
+    logger.info(
+        "Neo4j schema bootstrapped (%d constraints, %d indexes, %d fulltext)",
+        len(CONSTRAINTS),
+        len(INDEXES),
+        len(FULLTEXT_INDEXES),
+    )

document-parser/infra/neo4j/tree_reader.py

@@ -0,0 +1,64 @@
+"""TreeReader — fetch a DoclingDocument back from Neo4j.
+
+v0.5.0 implementation relies on the verbatim `document_json` property stored
+on the Document node by TreeWriter. Reconstruction by walking Element nodes
+is deferred to v0.6 (EnrichmentWriter prerequisite), where we may need to
+rebuild the DoclingDocument after enrichments have been patched on graph
+nodes directly.
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from infra.neo4j.driver import Neo4jDriver
+
+logger = logging.getLogger(__name__)
+
+
+async def read_document_json(neo: Neo4jDriver, doc_id: str) -> str | None:
+    """Return the stored DoclingDocument JSON for `doc_id`, or None if absent."""
+    async with neo.driver.session(database=neo.database) as session:
+        result = await session.run(
+            "MATCH (d:Document {id: $doc_id}) RETURN d.document_json AS json",
+            doc_id=doc_id,
+        )
+        record = await result.single()
+    if record is None:
+        return None
+    return record["json"]
+
+
+async def document_exists(neo: Neo4jDriver, doc_id: str) -> bool:
+    async with neo.driver.session(database=neo.database) as session:
+        result = await session.run(
+            "MATCH (d:Document {id: $doc_id}) RETURN count(d) AS n",
+            doc_id=doc_id,
+        )
+        record = await result.single()
+    return bool(record and record["n"] > 0)
+
+
+async def delete_document(neo: Neo4jDriver, doc_id: str) -> int:
+    """Wipe everything related to a doc_id. Returns nodes removed."""
+    async with neo.driver.session(database=neo.database) as session:
+        result = await session.run(
+            """
+            MATCH (d:Document {id: $doc_id})
+            OPTIONAL MATCH (d)-[:HAS_ROOT|HAS_CHUNK*0..]->(n)
+            WITH d, collect(DISTINCT n) AS children
+            DETACH DELETE d
+            WITH children
+            UNWIND children AS c
+            DETACH DELETE c
+            RETURN size(children) + 1 AS removed
+            """,
+            doc_id=doc_id,
+        )
+        record = await result.single()
+        # Also clean up orphan elements and pages tagged with this doc_id.
+        await session.run("MATCH (e:Element {doc_id: $doc_id}) DETACH DELETE e", doc_id=doc_id)
+        await session.run("MATCH (p:Page {doc_id: $doc_id}) DETACH DELETE p", doc_id=doc_id)
+    return int(record["removed"]) if record else 0

document-parser/infra/neo4j/tree_writer.py

@@ -0,0 +1,310 @@
+"""TreeWriter — persist a DoclingDocument as a graph in Neo4j.
+
+v0.5.0 strategy: replace-on-write. For a given doc_id, all existing
+Document/Element/Page/Chunk nodes are wiped before re-ingestion. The full
+serialized `DoclingDocument` JSON is stored as a property on the Document
+node so that `TreeReader` can round-trip it verbatim — reconstruction from
+graph nodes is deferred to v0.6 (see docs/design/neo4j-integration.md §2).
+"""
+
+from __future__ import annotations
+
+import contextlib
+import json
+import logging
+from dataclasses import dataclass
+from datetime import UTC, datetime
+from typing import TYPE_CHECKING, Any
+
+from infra.docling_tree import (
+    build_collapse_index,
+    dfs_order,
+    element_label,
+    is_inline_group,
+    iter_items,
+    iter_pages,
+    iter_provs,
+    parent_ref,
+)
+
+if TYPE_CHECKING:
+    from infra.neo4j.driver import Neo4jDriver
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class TreeWriteResult:
+    doc_id: str
+    elements_written: int
+    pages_written: int
+    provenances_written: int = 0
+
+
+def _element_props(item: dict[str, Any], doc_id: str) -> dict[str, Any]:
+    """Properties stored on the `:Element` node itself.
+
+    Provenance (page + bbox) is NOT here anymore — see `_iter_provs` and the
+    `:Provenance` nodes. Keeping it out of the element matches DoclingDocument's
+    own model (`prov` is a list of objects, not a scalar).
+    """
+    props: dict[str, Any] = {
+        "doc_id": doc_id,
+        "self_ref": item.get("self_ref") or "",
+        "docling_label": (item.get("label") or "").lower(),
+        "text": item.get("text") or "",
+    }
+    # Type-specific extras.
+    if "level" in item:
+        props["level"] = item.get("level")
+    if "caption" in item and isinstance(item.get("caption"), str):
+        props["caption"] = item.get("caption")
+    if item.get("data") and isinstance(item["data"], dict):
+        # Tables carry cell layout under data; stringify to keep the schema flat.
+        with contextlib.suppress(TypeError, ValueError):
+            props["cells_json"] = json.dumps(item["data"])
+    return props
+
+
+async def write_document(
+    neo: Neo4jDriver,
+    *,
+    doc_id: str,
+    filename: str,
+    document_json: str,
+    tenant_id: str = "default",
+    source_uri: str | None = None,
+    docling_version: str | None = None,
+) -> TreeWriteResult:
+    """Persist the full DoclingDocument tree to Neo4j.
+
+    Idempotent: wipes any existing graph for doc_id before writing.
+    Fails fast (exception propagates) if Neo4j is unavailable — per design §8.5.
+    """
+    doc_data = json.loads(document_json)
+    ingested_at = datetime.now(tz=UTC).isoformat()
+
+    # Issue #197: collapse two noise patterns from Docling into the projection.
+    # InlineGroups (paragraph style runs) are merged into a single :Paragraph,
+    # and Pictures' internal text labels (flowchart/diagram/chart annotations)
+    # are dropped. Both produce refs that land in `skip_refs`.
+    skip_refs, inline_meta = build_collapse_index(doc_data)
+
+    elements: list[dict[str, Any]] = []
+    # Parallel list: one row per Provenance — each refers back to its owner
+    # element via `self_ref`, so we can batch MATCH-and-link after both node
+    # sets are created.
+    provenances: list[dict[str, Any]] = []
+    for _, item in iter_items(doc_data):
+        ref = item.get("self_ref")
+        if not ref or ref in skip_refs:
+            continue
+        specific = element_label(item.get("label") or "")
+        props = _element_props(item, doc_id)
+        if is_inline_group(item):
+            meta = inline_meta.get(ref, {"text": "", "provs": []})
+            props["text"] = meta["text"]
+            item_provs = meta["provs"]
+        else:
+            item_provs = iter_provs(item)
+        elements.append(
+            {
+                "specific_label": specific,
+                "parent_ref": parent_ref(item),
+                **props,
+            }
+        )
+        for prov in item_provs:
+            provenances.append({"doc_id": doc_id, "self_ref": ref, **prov})
+
+    pages: list[dict[str, Any]] = [{"doc_id": doc_id, **p} for p in iter_pages(doc_data)]
+
+    reading_order = dfs_order(doc_data, skip_refs)
+
+    async with (
+        neo.driver.session(database=neo.database) as session,
+        await session.begin_transaction() as tx,
+    ):
+        # 1. Wipe existing graph for this doc_id (replace strategy).
+        await tx.run(
+            "MATCH (d:Document {id: $doc_id}) "
+            "OPTIONAL MATCH (d)-[:HAS_ROOT|HAS_CHUNK*0..]->(n) "
+            "DETACH DELETE d, n",
+            doc_id=doc_id,
+        )
+        # Orphan sweep — covers Provenance/Element/Page/Chunk that may linger
+        # from an interrupted write or a pre-refactor schema.
+        await tx.run("MATCH (pv:Provenance {doc_id: $doc_id}) DETACH DELETE pv", doc_id=doc_id)
+        await tx.run("MATCH (e:Element {doc_id: $doc_id}) DETACH DELETE e", doc_id=doc_id)
+        await tx.run("MATCH (p:Page {doc_id: $doc_id}) DETACH DELETE p", doc_id=doc_id)
+
+        # 2. Document node (carries the verbatim JSON for TreeReader).
+        await tx.run(
+            """
+                CREATE (d:Document {
+                  id: $doc_id,
+                  title: $title,
+                  source_uri: $source_uri,
+                  ingested_at: datetime($ingested_at),
+                  docling_version: $docling_version,
+                  stages_applied: ['tree'],
+                  last_tree_write: datetime($ingested_at),
+                  tenant_id: $tenant_id,
+                  document_json: $document_json
+                })
+                """,
+            doc_id=doc_id,
+            title=filename,
+            source_uri=source_uri or "",
+            ingested_at=ingested_at,
+            docling_version=docling_version or "",
+            tenant_id=tenant_id,
+            document_json=document_json,
+        )
+
+        # 3. Page nodes.
+        if pages:
+            await tx.run(
+                "UNWIND $pages AS p "
+                "CREATE (:Page {doc_id: p.doc_id, page_no: p.page_no, "
+                "width: p.width, height: p.height})",
+                pages=pages,
+            )
+
+        # 4. Element nodes — use dynamic :Element:<specific> labels via APOC-free trick.
+        # We split by specific label so the CREATE statement is static (no APOC).
+        by_specific: dict[str, list[dict[str, Any]]] = {}
+        for e in elements:
+            by_specific.setdefault(e["specific_label"], []).append(e)
+        for specific, batch in by_specific.items():
+            await tx.run(
+                f"""
+                    UNWIND $batch AS e
+                    CREATE (n:Element:{specific} {{
+                      doc_id: e.doc_id,
+                      self_ref: e.self_ref,
+                      docling_label: e.docling_label,
+                      text: e.text,
+                      level: e.level,
+                      caption: e.caption,
+                      cells_json: e.cells_json
+                    }})
+                    """,
+                batch=batch,
+            )
+
+        # 5. PARENT_OF relations (tree structure). Order tracked inline.
+        parent_rows = [
+            {
+                "doc_id": doc_id,
+                "parent_ref": e["parent_ref"],
+                "child_ref": e["self_ref"],
+                "order": idx,
+            }
+            for idx, e in enumerate(elements)
+            if e["parent_ref"] and e["parent_ref"] != "#/body"
+        ]
+        if parent_rows:
+            await tx.run(
+                """
+                    UNWIND $rows AS r
+                    MATCH (p:Element {doc_id: r.doc_id, self_ref: r.parent_ref})
+                    MATCH (c:Element {doc_id: r.doc_id, self_ref: r.child_ref})
+                    MERGE (p)-[rel:PARENT_OF]->(c)
+                    SET rel.order = r.order
+                    """,
+                rows=parent_rows,
+            )
+
+        # 6. HAS_ROOT for top-level children of the document body.
+        root_rows = [
+            {"doc_id": doc_id, "child_ref": e["self_ref"]}
+            for e in elements
+            if e["parent_ref"] == "#/body"
+        ]
+        if root_rows:
+            await tx.run(
+                """
+                    UNWIND $rows AS r
+                    MATCH (d:Document {id: r.doc_id})
+                    MATCH (c:Element {doc_id: r.doc_id, self_ref: r.child_ref})
+                    MERGE (d)-[:HAS_ROOT]->(c)
+                    """,
+                rows=root_rows,
+            )
+
+        # 7. Provenance nodes — one per (element, prov-entry) pair. Mirrors
+        # Docling's `item.prov = list[ProvenanceItem]` 1:1 so a single item
+        # that spans page breaks (or appears twice in the layout) keeps every
+        # (page, bbox, charspan) without losing data.
+        if provenances:
+            await tx.run(
+                """
+                    UNWIND $rows AS r
+                    MATCH (e:Element {doc_id: r.doc_id, self_ref: r.self_ref})
+                    CREATE (pv:Provenance {
+                      doc_id: r.doc_id,
+                      element_ref: r.self_ref,
+                      prov_order: r.order,
+                      page_no: r.page_no,
+                      bbox_l: r.bbox_l,
+                      bbox_t: r.bbox_t,
+                      bbox_r: r.bbox_r,
+                      bbox_b: r.bbox_b,
+                      coord_origin: r.coord_origin,
+                      charspan_start: r.charspan_start,
+                      charspan_end: r.charspan_end
+                    })
+                    CREATE (e)-[:HAS_PROV {order: r.order}]->(pv)
+                    """,
+                rows=provenances,
+            )
+            # ON_PAGE now attaches the Provenance to its Page — lets downstream
+            # queries ("what's on page 3?") stay simple without walking through
+            # the Element. A Provenance with no page_no (rare) yields no edge.
+            await tx.run(
+                """
+                    UNWIND $rows AS r
+                    WITH r WHERE r.page_no IS NOT NULL
+                    MATCH (pv:Provenance {
+                      doc_id: r.doc_id,
+                      element_ref: r.self_ref,
+                      prov_order: r.order
+                    })
+                    MATCH (p:Page {doc_id: r.doc_id, page_no: r.page_no})
+                    MERGE (pv)-[:ON_PAGE]->(p)
+                    """,
+                rows=provenances,
+            )
+
+        # 8. NEXT chain in DFS pre-order.
+        if len(reading_order) > 1:
+            pairs = [
+                {"doc_id": doc_id, "a": reading_order[i], "b": reading_order[i + 1]}
+                for i in range(len(reading_order) - 1)
+            ]
+            await tx.run(
+                """
+                    UNWIND $pairs AS p
+                    MATCH (a:Element {doc_id: p.doc_id, self_ref: p.a})
+                    MATCH (b:Element {doc_id: p.doc_id, self_ref: p.b})
+                    MERGE (a)-[:NEXT]->(b)
+                    """,
+                pairs=pairs,
+            )
+
+        await tx.commit()
+
+    logger.info(
+        "Neo4j: wrote doc %s (%d elements, %d pages, %d provenances)",
+        doc_id,
+        len(elements),
+        len(pages),
+        len(provenances),
+    )
+    return TreeWriteResult(
+        doc_id=doc_id,
+        elements_written=len(elements),
+        pages_written=len(pages),
+        provenances_written=len(provenances),
+    )

document-parser/infra/opensearch_store.py

@@ -69,13 +69,14 @@ class OpenSearchStore:
         verify_certs: Whether to verify TLS certificates.
     """
 
-    def __init__(self, url: str, *, verify_certs: bool = False) -> None:
+    def __init__(self, url: str, *, verify_certs: bool = False, default_limit: int = 1000) -> None:
         self._client = AsyncOpenSearch(
             hosts=[url],
             use_ssl=url.startswith("https"),
             verify_certs=verify_certs,
             ssl_show_warn=False,
         )
+        self._default_limit = default_limit
 
     # -- lifecycle -------------------------------------------------------------
 
@@ -147,9 +148,11 @@ class OpenSearchStore:
         index_name: str,
         doc_id: str,
         *,
-        limit: int = 1000,
+        limit: int | None = None,
     ) -> list[SearchResult]:
         """Retrieve all indexed chunks for a document, ordered by chunk_index."""
+        if limit is None:
+            limit = self._default_limit
         resp = await self._client.search(
             index=index_name,
             body={

document-parser/infra/serve_converter.py

@@ -21,6 +21,8 @@ import httpx
 from docling_core.types.doc.base import BoundingBox, CoordOrigin
 
 from domain.value_objects import (
+    DEFAULT_PAGE_HEIGHT,
+    DEFAULT_PAGE_WIDTH,
     ConversionOptions,
     ConversionResult,
     PageDetail,
@@ -31,7 +33,6 @@ from infra.bbox import to_topleft_list
 logger = logging.getLogger(__name__)
 
 _API_PREFIX = "/v1"
-_DEFAULT_TIMEOUT = 600.0
 
 # Docling Serve label → our element type
 _LABEL_MAP = {
@@ -60,7 +61,7 @@ class ServeConverter:
         self,
         base_url: str,
         api_key: str | None = None,
-        timeout: float = _DEFAULT_TIMEOUT,
+        timeout: float = 600.0,
     ):
         self._base_url = base_url.rstrip("/")
         self._api_key = api_key
@@ -95,6 +96,13 @@ class ServeConverter:
                     headers=self._headers(),
                 )
 
+        if response.status_code >= 400:
+            logger.error(
+                "Docling Serve error %d: %s (form_data=%s)",
+                response.status_code,
+                response.text[:500],
+                {k: v for k, v in form_data.items()},
+            )
         response.raise_for_status()
         result_data = response.json()
 
@@ -121,8 +129,12 @@ def _build_form_data(
 ) -> dict[str, str | list[str]]:
     """Build form fields matching Docling Serve's multipart form contract.
 
-    Array fields (to_formats) are sent as lists — httpx encodes them as
+    Serve uses FastAPI's ``Form()`` parsing — list/tuple fields are sent
-    repeated form keys (to_formats=md&to_formats=html&to_formats=json).
+    as **repeated form keys** (httpx encodes Python lists this way
+    automatically: ``to_formats=md&to_formats=html&to_formats=json``).
+
+    Note: ``generate_page_images`` is a PdfPipelineOptions field, NOT a
+    ConvertDocumentsOptions field — sending it causes a 422.
     """
     data: dict[str, str | list[str]] = {
         "to_formats": ["md", "html", "json"],
@@ -134,11 +146,12 @@ def _build_form_data(
         "do_picture_classification": str(options.do_picture_classification).lower(),
         "do_picture_description": str(options.do_picture_description).lower(),
         "include_images": str(options.generate_picture_images).lower(),
-        "generate_page_images": str(options.generate_page_images).lower(),
         "images_scale": str(options.images_scale),
     }
     if page_range is not None:
-        data["page_range"] = f"{page_range[0]}-{page_range[1]}"
+        # Serve expects page_range as two repeated form fields:
+        # page_range=1&page_range=10
+        data["page_range"] = [str(page_range[0]), str(page_range[1])]
     return data
 
 
@@ -192,8 +205,8 @@ def _extract_pages_from_docling_document(doc: dict) -> list[PageDetail]:
         size = page_data.get("size", {})
         pages_dict[page_no] = PageDetail(
             page_number=page_no,
-            width=size.get("width", 612.0),
+            width=size.get("width", DEFAULT_PAGE_WIDTH),
-            height=size.get("height", 792.0),
+            height=size.get("height", DEFAULT_PAGE_HEIGHT),
         )
 
     # Process all element arrays
@@ -220,8 +233,8 @@ def _add_element(item: dict, pages: dict[int, PageDetail]) -> None:
         if page_no not in pages:
             pages[page_no] = PageDetail(
                 page_number=page_no,
-                width=612.0,
+                width=DEFAULT_PAGE_WIDTH,
-                height=792.0,
+                height=DEFAULT_PAGE_HEIGHT,
             )
 
         bbox_data = prov.get("bbox", {})

document-parser/infra/settings.py

@@ -25,9 +25,23 @@ class Settings:
     batch_page_size: int = 0  # 0 = disabled, > 0 = pages per batch
     opensearch_url: str = ""  # empty = disabled
     embedding_url: str = ""  # empty = disabled (e.g. http://localhost:8001)
+    neo4j_uri: str = ""  # empty = disabled (e.g. bolt://neo4j:7687)
+    neo4j_user: str = "neo4j"
+    neo4j_password: str = "changeme"
+    # Live reasoning via docling-agent — off by default (heavy deps, needs an
+    # Ollama host reachable from the backend). Toggle RAG_ENABLED=true + point
+    # OLLAMA_HOST at a running instance (default http://localhost:11434).
+    rag_enabled: bool = False
+    ollama_host: str = "http://localhost:11434"
+    rag_model_id: str = "gpt-oss:20b"  # matches docling-agent's example_05
+    opensearch_default_limit: int = 1000  # max chunks returned by get_chunks
     embedding_dimension: int = 384  # Granite Embedding 30M / all-MiniLM-L6-v2
     upload_dir: str = "./uploads"
     db_path: str = "./data/docling_studio.db"
+    max_paste_image_size_mb: int = 10  # clipboard-paste image limit in MB (0 = unlimited)
+    paste_allowed_image_types: list[str] = field(
+        default_factory=lambda: ["image/png", "image/jpeg", "image/webp"]
+    )
     cors_origins: list[str] = field(
         default_factory=lambda: ["http://localhost:3000", "http://localhost:5173"]
     )
@@ -50,10 +64,20 @@ class Settings:
             errors.append(f"max_file_size must be >= 0 (got {self.max_file_size})")
         if self.max_file_size_mb < 0:
             errors.append(f"max_file_size_mb must be >= 0 (got {self.max_file_size_mb})")
+        if self.max_paste_image_size_mb < 0:
+            errors.append(
+                f"max_paste_image_size_mb must be >= 0 (got {self.max_paste_image_size_mb})"
+            )
+        if not self.paste_allowed_image_types:
+            errors.append("paste_allowed_image_types must not be empty")
         if self.rate_limit_rpm < 0:
             errors.append(f"rate_limit_rpm must be >= 0 (got {self.rate_limit_rpm})")
         if self.batch_page_size < 0:
             errors.append(f"batch_page_size must be >= 0 (got {self.batch_page_size})")
+        if self.opensearch_default_limit < 1:
+            errors.append(
+                f"opensearch_default_limit must be >= 1 (got {self.opensearch_default_limit})"
+            )
         if self.embedding_dimension < 1:
             errors.append(f"embedding_dimension must be >= 1 (got {self.embedding_dimension})")
         if self.default_table_mode not in ("accurate", "fast"):
@@ -79,6 +103,9 @@ class Settings:
     def from_env(cls) -> Settings:
         """Build a Settings instance from environment variables."""
         cors_raw = os.environ.get("CORS_ORIGINS", "http://localhost:3000,http://localhost:5173")
+        paste_types_raw = os.environ.get(
+            "PASTE_ALLOWED_IMAGE_TYPES", "image/png,image/jpeg,image/webp"
+        )
         return cls(
             app_version=os.environ.get("APP_VERSION", "dev"),
             conversion_engine=os.environ.get("CONVERSION_ENGINE", "local"),
@@ -94,12 +121,26 @@ class Settings:
             max_file_size=int(os.environ.get("MAX_FILE_SIZE", "0")),
             max_file_size_mb=int(os.environ.get("MAX_FILE_SIZE_MB", "50")),
             rate_limit_rpm=int(os.environ.get("RATE_LIMIT_RPM", "100")),
-            batch_page_size=int(os.environ.get("BATCH_PAGE_SIZE", "10")),
+            # 0 = batching disabled (matches dataclass default). Batching
+            # preserves memory on very large docs but `merge_results` drops
+            # `document_json`, which breaks the reasoning tunnel. Enable
+            # explicitly (e.g. 50+) for memory-bound deploys.
+            batch_page_size=int(os.environ.get("BATCH_PAGE_SIZE", "0")),
             opensearch_url=os.environ.get("OPENSEARCH_URL", ""),
             embedding_url=os.environ.get("EMBEDDING_URL", ""),
+            neo4j_uri=os.environ.get("NEO4J_URI", ""),
+            neo4j_user=os.environ.get("NEO4J_USER", "neo4j"),
+            neo4j_password=os.environ.get("NEO4J_PASSWORD", "changeme"),
+            rag_enabled=os.environ.get("RAG_ENABLED", "false").lower()
+            in ("1", "true", "yes", "on"),
+            ollama_host=os.environ.get("OLLAMA_HOST", "http://localhost:11434"),
+            rag_model_id=os.environ.get("RAG_MODEL_ID", "gpt-oss:20b"),
+            opensearch_default_limit=int(os.environ.get("OPENSEARCH_DEFAULT_LIMIT", "1000")),
             embedding_dimension=int(os.environ.get("EMBEDDING_DIMENSION", "384")),
             upload_dir=os.environ.get("UPLOAD_DIR", "./uploads"),
             db_path=os.environ.get("DB_PATH", "./data/docling_studio.db"),
+            max_paste_image_size_mb=int(os.environ.get("MAX_PASTE_IMAGE_SIZE_MB", "10")),
+            paste_allowed_image_types=[t.strip() for t in paste_types_raw.split(",") if t.strip()],
             cors_origins=[o.strip() for o in cors_raw.split(",")],
         )
 

document-parser/main.py

@@ -47,6 +47,7 @@ def _build_converter():
         return ServeConverter(
             base_url=settings.docling_serve_url,
             api_key=settings.docling_serve_api_key,
+            timeout=settings.conversion_timeout,
         )
     else:
         from infra.local_converter import LocalConverter
@@ -56,12 +57,15 @@ def _build_converter():
 
 
 def _build_chunker():
-    """Build the chunker adapter — only available in local mode."""
+    """Build the chunker adapter.
-    if settings.conversion_engine == "local":
-        from infra.local_chunker import LocalChunker
 
-        return LocalChunker()
+    Uses LocalChunker in all modes — in remote mode it chunks the
-    return None
+    DoclingDocument JSON returned by Docling Serve, so docling-core
+    (lightweight) is the only local dependency needed.
+    """
+    from infra.local_chunker import LocalChunker
+
+    return LocalChunker()
 
 
 def _build_repos() -> tuple[SqliteDocumentRepository, SqliteAnalysisRepository]:
@@ -71,6 +75,7 @@ def _build_repos() -> tuple[SqliteDocumentRepository, SqliteAnalysisRepository]:
 def _build_analysis_service(
     document_repo: SqliteDocumentRepository,
     analysis_repo: SqliteAnalysisRepository,
+    neo4j_driver=None,
 ) -> AnalysisService:
     converter = _build_converter()
     chunker = _build_chunker()
@@ -86,10 +91,33 @@ def _build_analysis_service(
         conversion_timeout=settings.conversion_timeout,
         max_concurrent=settings.max_concurrent_analyses,
         config=config,
+        neo4j_driver=neo4j_driver,
     )
 
 
-def _build_ingestion_service() -> IngestionService | None:
+async def _init_neo4j():
+    """Initialize the Neo4j driver and bootstrap schema — skip if not configured."""
+    if not settings.neo4j_uri:
+        logger.info("Neo4j disabled (NEO4J_URI not set)")
+        return None
+
+    from infra.neo4j import bootstrap_schema, get_driver
+
+    try:
+        neo = await get_driver(
+            settings.neo4j_uri,
+            settings.neo4j_user,
+            settings.neo4j_password,
+        )
+        await bootstrap_schema(neo)
+        logger.info("Neo4j ready (uri=%s)", settings.neo4j_uri)
+        return neo
+    except Exception:
+        logger.exception("Neo4j init failed — continuing without graph storage")
+        return None
+
+
+def _build_ingestion_service(neo4j_driver=None) -> IngestionService | None:
     """Build the ingestion service — only if embedding + opensearch are configured."""
     if not settings.embedding_url or not settings.opensearch_url:
         logger.info("Ingestion disabled (EMBEDDING_URL or OPENSEARCH_URL not set)")
@@ -99,7 +127,10 @@ def _build_ingestion_service() -> IngestionService | None:
     from infra.opensearch_store import OpenSearchStore
 
     embedding = EmbeddingClient(settings.embedding_url)
-    vector_store = OpenSearchStore(settings.opensearch_url)
+    vector_store = OpenSearchStore(
+        settings.opensearch_url,
+        default_limit=settings.opensearch_default_limit,
+    )
     config = IngestionConfig(
         embedding_dimension=settings.embedding_dimension,
     )
@@ -108,7 +139,7 @@ def _build_ingestion_service() -> IngestionService | None:
         settings.embedding_url,
         settings.opensearch_url,
     )
-    return IngestionService(embedding, vector_store, config)
+    return IngestionService(embedding, vector_store, config, neo4j_driver=neo4j_driver)
 
 
 def _build_document_service(
@@ -136,15 +167,30 @@ def _build_document_service(
 async def lifespan(app: FastAPI) -> AsyncIterator[None]:
     await init_db()
     document_repo, analysis_repo = _build_repos()
-    app.state.analysis_service = _build_analysis_service(document_repo, analysis_repo)
+    # Exposed on app.state so routers that need direct repo access (e.g. the
+    # reasoning-graph endpoint, which reads `document_json` from SQLite to
+    # build the graph without touching Neo4j) can reach them without going
+    # through a service.
+    app.state.analysis_repo = analysis_repo
+    app.state.document_repo = document_repo
+    app.state.neo4j = await _init_neo4j()
+    app.state.analysis_service = _build_analysis_service(
+        document_repo, analysis_repo, neo4j_driver=app.state.neo4j
+    )
     app.state.document_service = _build_document_service(document_repo, analysis_repo)
-    ingestion_service = _build_ingestion_service()
+    ingestion_service = _build_ingestion_service(neo4j_driver=app.state.neo4j)
     app.state.ingestion_service = ingestion_service
     if ingestion_service is not None:
         app.include_router(ingestion_router)
         logger.info("Ingestion router mounted")
     logger.info("Docling Studio backend ready (engine=%s)", settings.conversion_engine)
-    yield
+    try:
+        yield
+    finally:
+        if app.state.neo4j is not None:
+            from infra.neo4j import close_driver
+
+            await close_driver()
 
 
 app = FastAPI(
@@ -170,6 +216,18 @@ if settings.rate_limit_rpm > 0:
 app.include_router(documents_router)
 app.include_router(analyses_router)
 
+# Graph view — mounted regardless; individual requests 503 if Neo4j is absent.
+from api.graph import router as graph_router  # noqa: E402
+
+app.include_router(graph_router)
+
+# Live reasoning (docling-agent runner). Router is mounted unconditionally so
+# the route is introspectable in OpenAPI; the handler itself 503s when
+# `RAG_ENABLED` is off or the deps aren't installed.
+from api.reasoning import router as reasoning_router  # noqa: E402
+
+app.include_router(reasoning_router)
+
 
 @app.get("/api/health", response_model=HealthResponse)
 async def health() -> HealthResponse:
@@ -191,5 +249,23 @@ async def health() -> HealthResponse:
         database=db_status,
         max_page_count=settings.max_page_count if settings.max_page_count > 0 else None,
         max_file_size_mb=settings.max_file_size_mb if settings.max_file_size_mb > 0 else None,
+        max_paste_image_size_mb=(
+            settings.max_paste_image_size_mb if settings.max_paste_image_size_mb > 0 else None
+        ),
+        paste_allowed_image_types=settings.paste_allowed_image_types,
         ingestion_available=getattr(app.state, "ingestion_service", None) is not None,
+        # True when the live-reasoning runner is wired (flag on + deps present).
+        # The actual Ollama reachability is checked lazily at call-time to avoid
+        # blocking health checks on the LLM host.
+        rag_available=settings.rag_enabled and _rag_deps_present(),
     )
+
+
+def _rag_deps_present() -> bool:
+    """Import-check only — does not hit Ollama."""
+    try:
+        import docling_agent.agents  # noqa: F401
+        import mellea  # noqa: F401
+    except ImportError:
+        return False
+    return True

document-parser/persistence/analysis_repo.py

@@ -73,6 +73,23 @@ class SqliteAnalysisRepository:
             row = await cursor.fetchone()
             return _row_to_job(row) if row else None
 
+    async def find_latest_completed_by_document(self, document_id: str) -> AnalysisJob | None:
+        """Latest COMPLETED analysis with a non-null `document_json` for a doc.
+
+        Used by the reasoning-trace tunnel to prime Neo4j from an existing
+        analysis when the graph doesn't yet exist (e.g. analysis ran before
+        Neo4j was wired in).
+        """
+        async with get_connection() as db:
+            cursor = await db.execute(
+                f"{_SELECT_WITH_DOC} WHERE aj.document_id = ? "
+                "AND aj.status = 'COMPLETED' AND aj.document_json IS NOT NULL "
+                "ORDER BY aj.completed_at DESC LIMIT 1",
+                (document_id,),
+            )
+            row = await cursor.fetchone()
+            return _row_to_job(row) if row else None
+
     async def update_status(self, job: AnalysisJob) -> None:
         """Persist all mutable fields of an analysis job (status, results, timestamps)."""
         async with get_connection() as db:

document-parser/persistence/database.py

@@ -9,11 +9,9 @@ from contextlib import asynccontextmanager
 
 import aiosqlite
 
-from infra.settings import settings
-
 logger = logging.getLogger(__name__)
 
-DB_PATH = settings.db_path
+DB_PATH = os.environ.get("DB_PATH", "./data/docling_studio.db")
 
 _SCHEMA = """
 CREATE TABLE IF NOT EXISTS documents (

document-parser/requirements-test.txt

@@ -0,0 +1,4 @@
+-r requirements.txt
+pytest>=8.0.0,<9.0.0
+pytest-asyncio>=0.23.0,<1.0.0
+pytestarch>=2.0.0,<3.0.0

document-parser/requirements.txt

@@ -1,4 +1,4 @@
-docling-core>=2.0.0,<3.0.0
+docling-core[chunking]>=2.0.0,<3.0.0
 fastapi>=0.115.0,<1.0.0
 uvicorn[standard]>=0.32.0,<1.0.0
 python-multipart>=0.0.12
@@ -8,3 +8,8 @@ aiosqlite>=0.20.0,<1.0.0
 httpx>=0.27.0,<1.0.0
 pypdfium2>=4.0.0,<5.0.0
 opensearch-py[async]>=2.6.0,<3.0.0
+neo4j>=5.15.0,<6.0.0
+# R&D reasoning-trace live runner — calls docling-agent's `_rag_loop` over
+# an Ollama backend. Gated server-side by `RAG_ENABLED`; pulls ~60MB of deps.
+docling-agent==0.1.0
+mellea==0.4.2

document-parser/services/analysis_service.py

@@ -44,6 +44,7 @@ def _chunk_to_dict(c: ChunkResult) -> dict:
         "sourcePage": c.source_page,
         "tokenCount": c.token_count,
         "bboxes": [{"page": b.page, "bbox": b.bbox} for b in c.bboxes],
+        "docItems": [{"selfRef": d.self_ref, "label": d.label} for d in c.doc_items],
     }
 
 
@@ -69,6 +70,7 @@ class AnalysisConfig:
 
     default_table_mode: str = "accurate"
     batch_page_size: int = 0
+    neo4j_required: bool = False  # if True, ingestion fails when Neo4j write fails
 
 
 class AnalysisService:
@@ -83,6 +85,7 @@ class AnalysisService:
         conversion_timeout: int = 600,
         max_concurrent: int = _DEFAULT_MAX_CONCURRENT,
         config: AnalysisConfig | None = None,
+        neo4j_driver=None,
     ):
         self._converter = converter
         self._chunker = chunker
@@ -93,6 +96,7 @@ class AnalysisService:
         self._running_tasks: dict[str, asyncio.Task] = {}
         self._background_tasks: set[asyncio.Task] = set()
         self._config = config or AnalysisConfig()
+        self._neo4j = neo4j_driver
 
     async def create(
         self,
@@ -324,11 +328,18 @@ class AnalysisService:
         file_path: str,
         options: ConversionOptions,
     ) -> ConversionResult | None:
-        """Run batched or single conversion. Returns None if the job was deleted mid-batch."""
+        """Run batched or single conversion. Returns None if the job was deleted mid-batch.
+
+        Batching is only used for local mode — it limits memory usage when
+        Docling runs in-process.  In remote mode the Serve instance manages
+        its own resources, and batching would discard document_json (needed
+        for chunking).
+        """
         total_pages = _count_pdf_pages(file_path)
         batch_size = self._config.batch_page_size
+        is_remote = self._is_remote_converter()
 
-        if batch_size > 0 and total_pages > batch_size:
+        if batch_size > 0 and total_pages > batch_size and not is_remote:
             return await self._run_batched_conversion(
                 job_id, file_path, options, total_pages, batch_size
             )
@@ -337,6 +348,15 @@ class AnalysisService:
             timeout=self._conversion_timeout,
         )
 
+    def _is_remote_converter(self) -> bool:
+        """Check if the converter is a remote (Serve) adapter."""
+        try:
+            from infra.serve_converter import ServeConverter
+
+            return isinstance(self._converter, ServeConverter)
+        except ImportError:
+            return False
+
     async def _finalize_analysis(
         self,
         job_id: str,
@@ -370,8 +390,32 @@ class AnalysisService:
         if result.page_count:
             await self._document_repo.update_page_count(job.document_id, result.page_count)
 
+        await self._write_tree_to_neo4j(job, result.document_json)
+
         logger.info("Analysis completed: %s (%d pages)", job_id, result.page_count)
 
+    async def _write_tree_to_neo4j(self, job, document_json: str | None) -> None:
+        """Mirror the DoclingDocument tree into Neo4j if configured.
+
+        Silent no-op when Neo4j isn't wired in. Logs but does not fail the
+        analysis when the write fails, unless `config.neo4j_required` is set.
+        """
+        if self._neo4j is None or not document_json:
+            return
+        try:
+            from infra.neo4j import write_document
+
+            await write_document(
+                self._neo4j,
+                doc_id=job.document_id,
+                filename=job.document_filename or job.document_id,
+                document_json=document_json,
+            )
+        except Exception:
+            logger.exception("Neo4j TreeWriter failed for doc %s", job.document_id)
+            if self._config.neo4j_required:
+                raise
+
     async def _run_analysis_inner(
         self,
         job_id: str,

document-parser/services/ingestion_service.py

@@ -54,10 +54,12 @@ class IngestionService:
         embedding_service: EmbeddingService,
         vector_store: VectorStore,
         config: IngestionConfig | None = None,
+        neo4j_driver=None,
     ) -> None:
         self._embedding = embedding_service
         self._vector_store = vector_store
         self._config = config or IngestionConfig()
+        self._neo4j = neo4j_driver
 
     async def ensure_index(self) -> None:
         """Ensure the vector index exists with the correct mapping."""
@@ -139,6 +141,15 @@ class IngestionService:
         indexed = await self._vector_store.index_chunks(self._config.index_name, indexed_chunks)
         logger.info("Indexed %d/%d chunks for doc %s", indexed, len(indexed_chunks), doc_id)
 
+        # 5. Mirror chunks in Neo4j if configured (with DERIVED_FROM edges).
+        if self._neo4j is not None:
+            try:
+                from infra.neo4j import write_chunks
+
+                await write_chunks(self._neo4j, doc_id=doc_id, chunks_json=chunks_json)
+            except Exception:
+                logger.exception("Neo4j ChunkWriter failed for doc %s", doc_id)
+
         return IngestionResult(
             doc_id=doc_id,
             chunks_indexed=indexed,

document-parser/tests/neo4j/conftest.py

@@ -0,0 +1,40 @@
+"""Shared fixtures for Neo4j integration tests.
+
+These tests are skipped unless a live Neo4j is reachable via NEO4J_TEST_URI
+(defaulting to bolt://localhost:7687). CI spins up `neo4j:5.15-community`
+alongside the job; locally run `docker compose -f docker-compose.dev.yml up neo4j`.
+"""
+
+from __future__ import annotations
+
+import os
+
+import pytest
+
+# Skip the entire module cleanly when the neo4j driver package is absent
+# (e.g. local dev without the dependency installed).
+pytest.importorskip("neo4j")
+
+from infra.neo4j import close_driver, get_driver
+
+
+def _cfg() -> tuple[str, str, str]:
+    return (
+        os.environ.get("NEO4J_TEST_URI", "bolt://localhost:7687"),
+        os.environ.get("NEO4J_TEST_USER", "neo4j"),
+        os.environ.get("NEO4J_TEST_PASSWORD", "changeme"),
+    )
+
+
+@pytest.fixture
+async def neo4j_driver():
+    uri, user, password = _cfg()
+    try:
+        neo = await get_driver(uri, user, password)
+    except Exception as exc:
+        pytest.skip(f"Neo4j not reachable at {uri}: {exc}")
+    # Wipe DB before each test — integration tests assume an empty graph.
+    async with neo.driver.session(database=neo.database) as session:
+        await session.run("MATCH (n) DETACH DELETE n")
+    yield neo
+    await close_driver()

document-parser/tests/neo4j/test_chunk_writer.py

@@ -0,0 +1,113 @@
+"""ChunkWriter creates Chunk nodes + DERIVED_FROM links.
+
+Builds on the tree_writer fixture — writes the tree first so that DERIVED_FROM
+has Elements to link against.
+"""
+
+from __future__ import annotations
+
+import json
+
+from infra.neo4j import fetch_graph, write_chunks, write_document
+from infra.neo4j.schema import bootstrap_schema
+from tests.neo4j.test_tree_writer import FIXTURE
+
+CHUNKS = [
+    {
+        "text": "Introduction. First paragraph on page 1.",
+        "sourcePage": 1,
+        "tokenCount": 8,
+        "docItems": [
+            {"selfRef": "#/texts/0", "label": "section_header"},
+            {"selfRef": "#/texts/1", "label": "paragraph"},
+        ],
+    },
+    {
+        "text": "Continued on page 2.",
+        "sourcePage": 2,
+        "tokenCount": 4,
+        "docItems": [{"selfRef": "#/texts/2", "label": "paragraph"}],
+        "deleted": False,
+    },
+    # soft-deleted chunk: must be ignored
+    {"text": "gone", "deleted": True, "docItems": []},
+]
+
+
+async def test_write_chunks_and_derived_from(neo4j_driver):
+    await bootstrap_schema(neo4j_driver)
+    await write_document(
+        neo4j_driver,
+        doc_id="doc-fixture",
+        filename="fixture.pdf",
+        document_json=json.dumps(FIXTURE),
+    )
+
+    result = await write_chunks(
+        neo4j_driver,
+        doc_id="doc-fixture",
+        chunks_json=json.dumps(CHUNKS),
+    )
+
+    assert result.chunks_written == 2
+    assert result.derived_from_edges == 3
+
+    async with neo4j_driver.driver.session(database=neo4j_driver.database) as s:
+        count = await (
+            await s.run(
+                "MATCH (:Document {id: $id})-[:HAS_CHUNK]->(c:Chunk) RETURN count(c) AS n",
+                id="doc-fixture",
+            )
+        ).single()
+        assert count["n"] == 2
+
+        # First chunk derives from 2 elements, second from 1.
+        for idx, expected in [(0, 2), (1, 1)]:
+            cnt = await (
+                await s.run(
+                    "MATCH (c:Chunk {id: $cid})-[:DERIVED_FROM]->(e:Element) RETURN count(e) AS n",
+                    cid=f"doc-fixture::chunk::{idx}",
+                )
+            ).single()
+            assert cnt["n"] == expected
+
+        stages = await (
+            await s.run(
+                "MATCH (d:Document {id: $id}) RETURN d.stages_applied AS s", id="doc-fixture"
+            )
+        ).single()
+        assert "chunks" in stages["s"]
+
+
+async def test_fetch_graph_returns_full_payload(neo4j_driver):
+    await bootstrap_schema(neo4j_driver)
+    await write_document(
+        neo4j_driver,
+        doc_id="doc-fixture",
+        filename="fixture.pdf",
+        document_json=json.dumps(FIXTURE),
+    )
+    await write_chunks(
+        neo4j_driver,
+        doc_id="doc-fixture",
+        chunks_json=json.dumps(CHUNKS),
+    )
+
+    payload = await fetch_graph(neo4j_driver, "doc-fixture")
+    assert payload is not None
+    assert payload.truncated is False
+    assert payload.page_count == 2
+
+    groups = {n["group"] for n in payload.nodes}
+    assert groups == {"document", "element", "page", "chunk"}
+
+    edge_types = {e["type"] for e in payload.edges}
+    # Every edge kind this fixture can produce should be present.
+    # PARENT_OF is intentionally excluded: all FIXTURE items have
+    # `parent = #/body`, so they're roots (→ HAS_ROOT) with no nested hierarchy.
+    assert {"HAS_ROOT", "NEXT", "ON_PAGE", "HAS_CHUNK", "DERIVED_FROM"} <= edge_types
+
+
+async def test_fetch_graph_missing_doc_returns_none(neo4j_driver):
+    await bootstrap_schema(neo4j_driver)
+    assert await fetch_graph(neo4j_driver, "no-such-doc") is None

document-parser/tests/neo4j/test_document_roundtrip.py

@@ -0,0 +1,32 @@
+"""Minimal Document node round-trip — validates the driver + schema end-to-end."""
+
+from __future__ import annotations
+
+from infra.neo4j.schema import bootstrap_schema
+
+
+async def test_document_write_read_delete(neo4j_driver):
+    await bootstrap_schema(neo4j_driver)
+
+    async with neo4j_driver.driver.session(database=neo4j_driver.database) as session:
+        await session.run(
+            "CREATE (d:Document {id: $id, title: $title, tenant_id: $tenant})",
+            id="doc-42",
+            title="Round-trip fixture",
+            tenant="default",
+        )
+
+        result = await session.run(
+            "MATCH (d:Document {id: $id}) RETURN d.title AS title, d.tenant_id AS tenant",
+            id="doc-42",
+        )
+        record = await result.single()
+        assert record is not None
+        assert record["title"] == "Round-trip fixture"
+        assert record["tenant"] == "default"
+
+        await session.run("MATCH (d:Document {id: $id}) DETACH DELETE d", id="doc-42")
+        gone = await (
+            await session.run("MATCH (d:Document {id: $id}) RETURN d", id="doc-42")
+        ).single()
+        assert gone is None

document-parser/tests/neo4j/test_driver.py

@@ -0,0 +1,10 @@
+"""Neo4j driver connectivity smoke test."""
+
+from __future__ import annotations
+
+
+async def test_driver_connects_and_runs_cypher(neo4j_driver):
+    async with neo4j_driver.driver.session(database=neo4j_driver.database) as session:
+        result = await session.run("RETURN 1 AS x")
+        record = await result.single()
+        assert record["x"] == 1

document-parser/tests/neo4j/test_schema.py

@@ -0,0 +1,38 @@
+"""Schema bootstrap is idempotent and produces the expected constraints/indexes."""
+
+from __future__ import annotations
+
+from infra.neo4j.schema import CONSTRAINTS, FULLTEXT_INDEXES, INDEXES, bootstrap_schema
+
+
+async def _count_schema(neo4j_driver) -> tuple[int, int]:
+    async with neo4j_driver.driver.session(database=neo4j_driver.database) as session:
+        constraints = await (await session.run("SHOW CONSTRAINTS")).data()
+        indexes = await (await session.run("SHOW INDEXES")).data()
+    return len(constraints), len(indexes)
+
+
+async def test_bootstrap_is_idempotent(neo4j_driver):
+    await bootstrap_schema(neo4j_driver)
+    first = await _count_schema(neo4j_driver)
+
+    # Running a second time must not duplicate anything.
+    await bootstrap_schema(neo4j_driver)
+    second = await _count_schema(neo4j_driver)
+
+    assert first == second
+    # Sanity: we created at least what we declared.
+    assert first[0] >= len(CONSTRAINTS)
+    assert first[1] >= len(INDEXES) + len(FULLTEXT_INDEXES)
+
+
+async def test_document_id_is_unique(neo4j_driver):
+    await bootstrap_schema(neo4j_driver)
+    async with neo4j_driver.driver.session(database=neo4j_driver.database) as session:
+        await session.run("CREATE (d:Document {id: 'doc-1', title: 'first'})")
+        with_err: Exception | None = None
+        try:
+            await session.run("CREATE (d:Document {id: 'doc-1', title: 'dup'})")
+        except Exception as exc:
+            with_err = exc
+        assert with_err is not None, "unique constraint on Document.id must reject duplicates"

document-parser/tests/neo4j/test_tree_writer.py

@@ -0,0 +1,469 @@
+"""TreeWriter round-trip + structural sanity checks.
+
+Fixture is a hand-crafted DoclingDocument JSON with: one section containing
+two paragraphs and a table, spanning two pages. Tests verify that the graph
+mirrors the structure (HAS_ROOT, PARENT_OF, ON_PAGE, NEXT) and that
+re-writing the same doc is an idempotent replace.
+"""
+
+from __future__ import annotations
+
+import json
+
+from infra.neo4j import read_document_json, write_document
+from infra.neo4j.schema import bootstrap_schema
+
+FIXTURE = {
+    "name": "fixture.pdf",
+    "pages": {
+        "1": {"page_no": 1, "size": {"width": 595, "height": 842}},
+        "2": {"page_no": 2, "size": {"width": 595, "height": 842}},
+    },
+    "body": {
+        "self_ref": "#/body",
+        "children": [
+            {"$ref": "#/texts/0"},
+            {"$ref": "#/texts/1"},
+            {"$ref": "#/texts/2"},
+            {"$ref": "#/tables/0"},
+        ],
+    },
+    "texts": [
+        {
+            "self_ref": "#/texts/0",
+            "parent": {"$ref": "#/body"},
+            "label": "section_header",
+            "text": "Introduction",
+            "level": 1,
+            "prov": [{"page_no": 1, "bbox": {"l": 10, "t": 10, "r": 100, "b": 30}}],
+        },
+        {
+            "self_ref": "#/texts/1",
+            "parent": {"$ref": "#/body"},
+            "label": "paragraph",
+            "text": "First paragraph on page 1.",
+            "prov": [{"page_no": 1, "bbox": {"l": 10, "t": 40, "r": 500, "b": 80}}],
+        },
+        {
+            "self_ref": "#/texts/2",
+            "parent": {"$ref": "#/body"},
+            "label": "paragraph",
+            "text": "Continued on page 2.",
+            "prov": [{"page_no": 2, "bbox": {"l": 10, "t": 40, "r": 500, "b": 80}}],
+        },
+    ],
+    "tables": [
+        {
+            "self_ref": "#/tables/0",
+            "parent": {"$ref": "#/body"},
+            "label": "table",
+            "text": "",
+            "data": {"num_rows": 2, "num_cols": 2, "grid": [[1, 2], [3, 4]]},
+            "prov": [{"page_no": 2, "bbox": {"l": 10, "t": 90, "r": 500, "b": 200}}],
+        }
+    ],
+    "pictures": [],
+    "groups": [],
+}
+
+
+async def _count(session, cypher: str, **params) -> int:
+    r = await session.run(cypher, **params)
+    rec = await r.single()
+    return int(rec["n"]) if rec else 0
+
+
+async def test_write_creates_expected_structure(neo4j_driver):
+    await bootstrap_schema(neo4j_driver)
+    doc_json = json.dumps(FIXTURE)
+
+    result = await write_document(
+        neo4j_driver,
+        doc_id="doc-fixture",
+        filename="fixture.pdf",
+        document_json=doc_json,
+    )
+
+    assert result.elements_written == 4
+    assert result.pages_written == 2
+    # Every item in FIXTURE has exactly one prov entry → 4 Provenance nodes.
+    assert result.provenances_written == 4
+
+    async with neo4j_driver.driver.session(database=neo4j_driver.database) as s:
+        assert (
+            await _count(
+                s,
+                "MATCH (d:Document {id: $id}) RETURN count(d) AS n",
+                id="doc-fixture",
+            )
+            == 1
+        )
+        assert (
+            await _count(
+                s,
+                "MATCH (:Document {id: $id})-[:HAS_ROOT]->(e:Element) RETURN count(e) AS n",
+                id="doc-fixture",
+            )
+            == 4
+        )
+        assert (
+            await _count(
+                s,
+                "MATCH (e:Element:SectionHeader {doc_id: $id, self_ref: '#/texts/0'}) "
+                "RETURN count(e) AS n",
+                id="doc-fixture",
+            )
+            == 1
+        )
+        assert (
+            await _count(
+                s,
+                "MATCH (e:Element:Table {doc_id: $id}) RETURN count(e) AS n",
+                id="doc-fixture",
+            )
+            == 1
+        )
+        # Reading-order chain: 3 NEXT edges for 4 elements.
+        assert (
+            await _count(
+                s,
+                "MATCH (a:Element {doc_id: $id})-[:NEXT]->(b:Element {doc_id: $id}) "
+                "RETURN count(*) AS n",
+                id="doc-fixture",
+            )
+            == 3
+        )
+        # Post-v0.6: ON_PAGE attaches Provenance to Page, not Element directly.
+        # Traverse through the Provenance node.
+        assert (
+            await _count(
+                s,
+                "MATCH (:Element {doc_id: $id})-[:HAS_PROV]->"
+                "(:Provenance)-[:ON_PAGE]->(:Page {doc_id: $id}) "
+                "RETURN count(*) AS n",
+                id="doc-fixture",
+            )
+            == 4
+        )
+        # Each element has exactly one Provenance here (single-page fixture).
+        assert (
+            await _count(
+                s,
+                "MATCH (e:Element {doc_id: $id})-[:HAS_PROV]->(pv:Provenance) "
+                "RETURN count(pv) AS n",
+                id="doc-fixture",
+            )
+            == 4
+        )
+
+
+async def test_rewrite_is_idempotent_replace(neo4j_driver):
+    await bootstrap_schema(neo4j_driver)
+    doc_json = json.dumps(FIXTURE)
+
+    await write_document(
+        neo4j_driver,
+        doc_id="doc-fixture",
+        filename="fixture.pdf",
+        document_json=doc_json,
+    )
+    # Second write with the same id must not duplicate anything.
+    await write_document(
+        neo4j_driver,
+        doc_id="doc-fixture",
+        filename="fixture.pdf",
+        document_json=doc_json,
+    )
+
+    async with neo4j_driver.driver.session(database=neo4j_driver.database) as s:
+        assert (
+            await _count(s, "MATCH (d:Document {id: $id}) RETURN count(d) AS n", id="doc-fixture")
+            == 1
+        )
+        assert (
+            await _count(
+                s,
+                "MATCH (e:Element {doc_id: $id}) RETURN count(e) AS n",
+                id="doc-fixture",
+            )
+            == 4
+        )
+
+
+async def test_reader_returns_verbatim_json(neo4j_driver):
+    await bootstrap_schema(neo4j_driver)
+    doc_json = json.dumps(FIXTURE, sort_keys=True)
+    await write_document(
+        neo4j_driver,
+        doc_id="doc-fixture",
+        filename="fixture.pdf",
+        document_json=doc_json,
+    )
+
+    read_back = await read_document_json(neo4j_driver, "doc-fixture")
+    assert read_back is not None
+    assert json.loads(read_back) == json.loads(doc_json)
+
+
+async def test_reader_missing_doc_returns_none(neo4j_driver):
+    await bootstrap_schema(neo4j_driver)
+    assert await read_document_json(neo4j_driver, "no-such-doc") is None
+
+
+# Issue #197: Docling emits one InlineGroup per inline-styled paragraph plus N
+# child `text` items (one per style run). Naive 1:1 mirroring blew up section
+# graphs into per-style-run nodes. The writer now collapses an InlineGroup into
+# a single :Paragraph node carrying the concatenated text of its children, and
+# skips those children entirely.
+INLINE_FIXTURE = {
+    "name": "inline.html",
+    "pages": {
+        "1": {"page_no": 1, "size": {"width": 595, "height": 842}},
+    },
+    "body": {
+        "self_ref": "#/body",
+        "children": [
+            {"$ref": "#/texts/0"},
+            {"$ref": "#/groups/0"},
+        ],
+    },
+    "texts": [
+        {
+            "self_ref": "#/texts/0",
+            "parent": {"$ref": "#/body"},
+            "label": "section_header",
+            "text": "Heading",
+            "level": 1,
+            "prov": [{"page_no": 1, "bbox": {"l": 10, "t": 10, "r": 100, "b": 30}}],
+        },
+        {
+            "self_ref": "#/texts/1",
+            "parent": {"$ref": "#/groups/0"},
+            "label": "text",
+            "text": "Hello",
+            "prov": [{"page_no": 1, "bbox": {"l": 10, "t": 40, "r": 50, "b": 60}}],
+        },
+        {
+            "self_ref": "#/texts/2",
+            "parent": {"$ref": "#/groups/0"},
+            "label": "text",
+            "text": "world",
+            "prov": [{"page_no": 1, "bbox": {"l": 55, "t": 40, "r": 100, "b": 60}}],
+        },
+        {
+            "self_ref": "#/texts/3",
+            "parent": {"$ref": "#/groups/0"},
+            "label": "text",
+            "text": "!",
+            "prov": [{"page_no": 1, "bbox": {"l": 105, "t": 40, "r": 110, "b": 60}}],
+        },
+    ],
+    "tables": [],
+    "pictures": [],
+    "groups": [
+        {
+            "self_ref": "#/groups/0",
+            "parent": {"$ref": "#/body"},
+            "label": "inline",
+            "children": [
+                {"$ref": "#/texts/1"},
+                {"$ref": "#/texts/2"},
+                {"$ref": "#/texts/3"},
+            ],
+        },
+    ],
+}
+
+
+async def test_inline_group_collapses_into_single_paragraph(neo4j_driver):
+    await bootstrap_schema(neo4j_driver)
+    doc_json = json.dumps(INLINE_FIXTURE)
+
+    result = await write_document(
+        neo4j_driver,
+        doc_id="doc-inline",
+        filename="inline.html",
+        document_json=doc_json,
+    )
+
+    # Section header + collapsed inline group only — NOT 5 (+3 style runs).
+    assert result.elements_written == 2
+    # The inline group inherits its 3 children's provs (1 each); the section
+    # header has its own prov → 4 Provenance nodes total.
+    assert result.provenances_written == 4
+
+    async with neo4j_driver.driver.session(database=neo4j_driver.database) as s:
+        r = await s.run(
+            "MATCH (e:Element:Paragraph {doc_id: $id, self_ref: '#/groups/0'}) "
+            "RETURN e.text AS text",
+            id="doc-inline",
+        )
+        rec = await r.single()
+        assert rec is not None, "InlineGroup should write a :Paragraph node"
+        assert rec["text"] == "Hello world !"
+
+        for child_ref in ("#/texts/1", "#/texts/2", "#/texts/3"):
+            assert (
+                await _count(
+                    s,
+                    "MATCH (e:Element {doc_id: $id, self_ref: $ref}) RETURN count(e) AS n",
+                    id="doc-inline",
+                    ref=child_ref,
+                )
+                == 0
+            ), f"Style-run {child_ref} should be skipped, not written as a node"
+
+        # Inline group inherits provs from all children (order preserved).
+        assert (
+            await _count(
+                s,
+                "MATCH (e:Element:Paragraph {doc_id: $id, self_ref: '#/groups/0'})"
+                "-[:HAS_PROV]->(pv:Provenance) RETURN count(pv) AS n",
+                id="doc-inline",
+            )
+            == 3
+        )
+
+        # Reading order: section_header → inline-as-paragraph (1 NEXT edge).
+        assert (
+            await _count(
+                s,
+                "MATCH (a:Element {doc_id: $id})-[:NEXT]->(b:Element) RETURN count(*) AS n",
+                id="doc-inline",
+            )
+            == 1
+        )
+
+        assert (
+            await _count(
+                s,
+                "MATCH (:Document {id: $id})-[:HAS_ROOT]->(:Element) RETURN count(*) AS n",
+                id="doc-inline",
+            )
+            == 2
+        )
+
+        # ON_PAGE through Provenance → Page (matches the new schema).
+        assert (
+            await _count(
+                s,
+                "MATCH (:Element {doc_id: $id})-[:HAS_PROV]->"
+                "(:Provenance)-[:ON_PAGE]->(:Page) RETURN count(*) AS n",
+                id="doc-inline",
+            )
+            == 4
+        )
+
+
+# Same issue (#197): a Picture's `children` are internal text labels (flowchart
+# boxes, chart axis labels, diagram callouts) extracted by Docling's layout
+# model. Mirroring them 1:1 drowns the figure in dozens of tiny nodes.
+PICTURE_FIXTURE = {
+    "name": "figure.pdf",
+    "pages": {
+        "1": {"page_no": 1, "size": {"width": 595, "height": 842}},
+    },
+    "body": {
+        "self_ref": "#/body",
+        "children": [
+            {"$ref": "#/texts/0"},
+            {"$ref": "#/pictures/0"},
+        ],
+    },
+    "texts": [
+        {
+            "self_ref": "#/texts/0",
+            "parent": {"$ref": "#/body"},
+            "label": "caption",
+            "text": "Figure 1: Pipeline overview.",
+            "prov": [{"page_no": 1, "bbox": {"l": 0, "t": 700, "r": 500, "b": 720}}],
+        },
+        # Internal labels — children of #/pictures/0.
+        {
+            "self_ref": "#/texts/1",
+            "parent": {"$ref": "#/pictures/0"},
+            "label": "text",
+            "text": "Parse",
+            "prov": [{"page_no": 1, "bbox": {"l": 100, "t": 200, "r": 130, "b": 220}}],
+        },
+        {
+            "self_ref": "#/texts/2",
+            "parent": {"$ref": "#/pictures/0"},
+            "label": "text",
+            "text": "Build",
+            "prov": [{"page_no": 1, "bbox": {"l": 140, "t": 200, "r": 170, "b": 220}}],
+        },
+        {
+            "self_ref": "#/texts/3",
+            "parent": {"$ref": "#/pictures/0"},
+            "label": "text",
+            "text": "Enrich",
+            "prov": [{"page_no": 1, "bbox": {"l": 180, "t": 200, "r": 220, "b": 220}}],
+        },
+    ],
+    "tables": [],
+    "pictures": [
+        {
+            "self_ref": "#/pictures/0",
+            "parent": {"$ref": "#/body"},
+            "label": "picture",
+            "children": [
+                {"$ref": "#/texts/1"},
+                {"$ref": "#/texts/2"},
+                {"$ref": "#/texts/3"},
+            ],
+            "captions": [{"$ref": "#/texts/0"}],
+            "prov": [{"page_no": 1, "bbox": {"l": 90, "t": 100, "r": 510, "b": 600}}],
+        }
+    ],
+    "groups": [],
+}
+
+
+async def test_picture_internal_labels_are_skipped(neo4j_driver):
+    await bootstrap_schema(neo4j_driver)
+    doc_json = json.dumps(PICTURE_FIXTURE)
+
+    result = await write_document(
+        neo4j_driver,
+        doc_id="doc-pic",
+        filename="figure.pdf",
+        document_json=doc_json,
+    )
+
+    # Caption + picture only — the 3 internal labels are skipped.
+    assert result.elements_written == 2
+
+    async with neo4j_driver.driver.session(database=neo4j_driver.database) as s:
+        for child_ref in ("#/texts/1", "#/texts/2", "#/texts/3"):
+            assert (
+                await _count(
+                    s,
+                    "MATCH (e:Element {doc_id: $id, self_ref: $ref}) RETURN count(e) AS n",
+                    id="doc-pic",
+                    ref=child_ref,
+                )
+                == 0
+            ), f"Picture child {child_ref} should be skipped"
+
+        # Picture stays a :Figure node with its own prov.
+        assert (
+            await _count(
+                s,
+                "MATCH (e:Element:Figure {doc_id: $id, self_ref: '#/pictures/0'}) "
+                "RETURN count(e) AS n",
+                id="doc-pic",
+            )
+            == 1
+        )
+
+        # No PARENT_OF from the picture to its dropped children.
+        assert (
+            await _count(
+                s,
+                "MATCH (:Element {doc_id: $id, self_ref: '#/pictures/0'})-[:PARENT_OF]->"
+                "(:Element) RETURN count(*) AS n",
+                id="doc-pic",
+            )
+            == 0
+        )

document-parser/tests/test_architecture.py

@@ -0,0 +1,208 @@
+"""Hexagonal architecture tests — enforce layer dependency rules.
+
+Uses pytestarch for inter-layer dependency rules and ast-based import
+scanning for external (third-party) dependency constraints.
+
+Rules enforced:
+- domain   -> no import from api, services, infra, persistence
+- services -> no import from api, infra, persistence
+- api      -> no import from infra, persistence
+- infra    -> no import from api, services
+- persistence -> no import from api, services, infra
+- domain   -> no import of fastapi, sqlalchemy, httpx, opensearchpy
+- services -> no import of fastapi
+"""
+
+from __future__ import annotations
+
+import ast
+from pathlib import Path
+
+import pytest
+from pytestarch import Rule, get_evaluable_architecture
+
+# ---------------------------------------------------------------------------
+# pytestarch evaluable (project root = document-parser/)
+# ---------------------------------------------------------------------------
+
+_PROJECT_ROOT = Path(__file__).resolve().parent.parent
+
+# pytestarch uses the directory name as module prefix when given absolute paths.
+# We use the directory name to build qualified module references.
+_PREFIX = _PROJECT_ROOT.name  # "document-parser"
+
+_evaluable = get_evaluable_architecture(str(_PROJECT_ROOT), str(_PROJECT_ROOT))
+
+
+def _mod(layer: str) -> str:
+    """Return the fully-qualified pytestarch module name for a layer."""
+    return f"{_PREFIX}.{layer}"
+
+
+# ---------------------------------------------------------------------------
+# Helper: collect top-level imports from all .py files in a package
+# ---------------------------------------------------------------------------
+
+
+def _collect_imports(package: str) -> set[str]:
+    """Return the set of top-level module names imported by *package*."""
+    pkg_path = Path(_PROJECT_ROOT) / package
+    imports: set[str] = set()
+    for py_file in pkg_path.rglob("*.py"):
+        tree = ast.parse(py_file.read_text(), filename=str(py_file))
+        for node in ast.walk(tree):
+            if isinstance(node, ast.Import):
+                for alias in node.names:
+                    imports.add(alias.name.split(".")[0])
+            elif isinstance(node, ast.ImportFrom) and node.module:
+                imports.add(node.module.split(".")[0])
+    return imports
+
+
+# ---------------------------------------------------------------------------
+# Inter-layer dependency rules (pytestarch)
+# ---------------------------------------------------------------------------
+
+
+class TestDomainLayerIsolation:
+    """domain must not depend on any other layer."""
+
+    @pytest.mark.parametrize("forbidden", ["api", "services", "infra", "persistence"])
+    def test_domain_does_not_import(self, forbidden: str):
+        rule = (
+            Rule()
+            .modules_that()
+            .are_sub_modules_of(_mod("domain"))
+            .should_not()
+            .import_modules_that()
+            .are_sub_modules_of(_mod(forbidden))
+        )
+        rule.assert_applies(_evaluable)
+
+
+class TestServicesLayerIsolation:
+    """services may import domain only."""
+
+    @pytest.mark.parametrize("forbidden", ["api", "infra", "persistence"])
+    def test_services_does_not_import(self, forbidden: str):
+        rule = (
+            Rule()
+            .modules_that()
+            .are_sub_modules_of(_mod("services"))
+            .should_not()
+            .import_modules_that()
+            .are_sub_modules_of(_mod(forbidden))
+        )
+        rule.assert_applies(_evaluable)
+
+
+class TestApiLayerIsolation:
+    """api may import services and domain, but not infra or persistence."""
+
+    @pytest.mark.parametrize("forbidden", ["infra", "persistence"])
+    def test_api_does_not_import(self, forbidden: str):
+        rule = (
+            Rule()
+            .modules_that()
+            .are_sub_modules_of(_mod("api"))
+            .should_not()
+            .import_modules_that()
+            .are_sub_modules_of(_mod(forbidden))
+        )
+        rule.assert_applies(_evaluable)
+
+
+class TestInfraLayerIsolation:
+    """infra may import domain (ports), but not api or services."""
+
+    @pytest.mark.parametrize("forbidden", ["api", "services"])
+    def test_infra_does_not_import(self, forbidden: str):
+        rule = (
+            Rule()
+            .modules_that()
+            .are_sub_modules_of(_mod("infra"))
+            .should_not()
+            .import_modules_that()
+            .are_sub_modules_of(_mod(forbidden))
+        )
+        rule.assert_applies(_evaluable)
+
+
+class TestPersistenceLayerIsolation:
+    """persistence may import domain, but not api, services, or infra."""
+
+    @pytest.mark.parametrize("forbidden", ["api", "services", "infra"])
+    def test_persistence_does_not_import(self, forbidden: str):
+        rule = (
+            Rule()
+            .modules_that()
+            .are_sub_modules_of(_mod("persistence"))
+            .should_not()
+            .import_modules_that()
+            .are_sub_modules_of(_mod(forbidden))
+        )
+        rule.assert_applies(_evaluable)
+
+
+# ---------------------------------------------------------------------------
+# External dependency rules (ast-based)
+# ---------------------------------------------------------------------------
+
+_DOMAIN_FORBIDDEN_EXTERNALS = {"fastapi", "sqlalchemy", "httpx", "opensearchpy"}
+_SERVICES_FORBIDDEN_EXTERNALS = {"fastapi"}
+
+
+class TestDomainExternalDependencies:
+    """domain must not import infrastructure-specific third-party libraries."""
+
+    @pytest.mark.parametrize("lib", sorted(_DOMAIN_FORBIDDEN_EXTERNALS))
+    def test_domain_does_not_import_external(self, lib: str):
+        imports = _collect_imports("domain")
+        assert lib not in imports, f"domain imports forbidden external library '{lib}'"
+
+
+class TestServicesExternalDependencies:
+    """services must not import web-framework libraries."""
+
+    @pytest.mark.parametrize("lib", sorted(_SERVICES_FORBIDDEN_EXTERNALS))
+    def test_services_does_not_import_external(self, lib: str):
+        imports = _collect_imports("services")
+        assert lib not in imports, f"services imports forbidden external library '{lib}'"
+
+
+# ---------------------------------------------------------------------------
+# Convention: ports live exclusively in domain.ports
+# ---------------------------------------------------------------------------
+
+
+class TestPortConvention:
+    """Protocol definitions (ports) must live in domain.ports only."""
+
+    def test_no_protocol_outside_domain_ports(self):
+        """No Protocol subclass should be defined outside domain/ports.py."""
+        ports_file = Path(_PROJECT_ROOT) / "domain" / "ports.py"
+        for py_file in Path(_PROJECT_ROOT).rglob("*.py"):
+            if py_file == ports_file:
+                continue
+            # Skip test files and __pycache__
+            if "tests" in py_file.parts or "__pycache__" in py_file.parts:
+                continue
+            tree = ast.parse(py_file.read_text(), filename=str(py_file))
+            for node in ast.walk(tree):
+                if isinstance(node, ast.ClassDef):
+                    for base in node.bases:
+                        base_name = _get_name(base)
+                        if base_name == "Protocol":
+                            pytest.fail(
+                                f"Protocol '{node.name}' defined in {py_file.relative_to(_PROJECT_ROOT)}"
+                                f" — ports must live in domain/ports.py"
+                            )
+
+
+def _get_name(node: ast.expr) -> str:
+    """Extract a simple name from an AST expression node."""
+    if isinstance(node, ast.Name):
+        return node.id
+    if isinstance(node, ast.Attribute):
+        return node.attr
+    return ""

document-parser/tests/test_chunking.py

@@ -66,6 +66,7 @@ class TestChunkResult:
             "source_page": 1,
             "token_count": 10,
             "bboxes": [],
+            "doc_items": [],
         }
 
 
@@ -465,3 +466,74 @@ class TestRechunkEndpoint:
             },
         )
         assert resp.status_code == 422
+
+
+# ---------------------------------------------------------------------------
+# Remote chunking path — hybrid local chunking from Serve's document_json
+# ---------------------------------------------------------------------------
+
+
+class TestRemoteChunkingPath:
+    """Verify that chunking works on document_json produced by Serve (remote mode)."""
+
+    @pytest.mark.asyncio
+    async def test_rechunk_with_serve_document_json(self):
+        """AnalysisService.rechunk() works with a LocalChunker even in remote mode."""
+        from infra.local_chunker import LocalChunker
+        from services.analysis_service import AnalysisService
+
+        chunker = LocalChunker()
+        analysis_repo = AsyncMock()
+        document_repo = AsyncMock()
+        converter = AsyncMock()  # ServeConverter mock — not used for rechunking
+
+        service = AnalysisService(
+            converter=converter,
+            analysis_repo=analysis_repo,
+            document_repo=document_repo,
+            chunker=chunker,
+        )
+
+        # Simulate a completed job with document_json from Serve
+        job = AnalysisJob(id="j-remote", document_id="d1")
+        job.mark_running()
+        job.mark_completed(
+            markdown="# Title\nParagraph text here.",
+            html="<h1>Title</h1><p>Paragraph text here.</p>",
+            pages_json="[]",
+            document_json=json.dumps(
+                {
+                    "schema_name": "DoclingDocument",
+                    "version": "1.0.0",
+                    "name": "test",
+                    "origin": {
+                        "mimetype": "application/pdf",
+                        "filename": "test.pdf",
+                        "binary_hash": 0,
+                    },
+                    "furniture": {
+                        "self_ref": "#/furniture",
+                        "children": [],
+                        "content_layer": "furniture",
+                    },
+                    "body": {"self_ref": "#/body", "children": [], "content_layer": "body"},
+                    "groups": [],
+                    "texts": [],
+                    "pictures": [],
+                    "tables": [],
+                    "key_value_items": [],
+                    "form_items": [],
+                    "pages": {},
+                }
+            ),
+        )
+        analysis_repo.find_by_id = AsyncMock(return_value=job)
+        analysis_repo.update_chunks = AsyncMock(return_value=True)
+
+        chunks = await service.rechunk(
+            "j-remote",
+            {"chunker_type": "hybrid", "max_tokens": 512},
+        )
+
+        assert isinstance(chunks, list)
+        analysis_repo.update_chunks.assert_called_once()

document-parser/tests/test_docling_graph.py

@@ -0,0 +1,398 @@
+"""Tests for `infra.docling_graph.build_graph_payload`.
+
+The fixture mirrors the DoclingDocument shape used in
+`tests/neo4j/test_tree_writer.py` so any structural drift between the two
+consumers (TreeWriter -> Neo4j, builder -> SQLite reasoning-graph) surfaces
+immediately.
+"""
+
+from __future__ import annotations
+
+import json
+
+from infra.docling_graph import build_graph_payload
+
+FIXTURE = {
+    "name": "fixture.pdf",
+    "pages": {
+        "1": {"page_no": 1, "size": {"width": 595, "height": 842}},
+        "2": {"page_no": 2, "size": {"width": 595, "height": 842}},
+    },
+    "body": {
+        "self_ref": "#/body",
+        "children": [
+            {"$ref": "#/texts/0"},
+            {"$ref": "#/texts/1"},
+            {"$ref": "#/texts/2"},
+            {"$ref": "#/tables/0"},
+        ],
+    },
+    "texts": [
+        {
+            "self_ref": "#/texts/0",
+            "parent": {"$ref": "#/body"},
+            "label": "section_header",
+            "text": "Introduction",
+            "level": 1,
+            "prov": [{"page_no": 1, "bbox": {"l": 10, "t": 10, "r": 100, "b": 30}}],
+        },
+        {
+            "self_ref": "#/texts/1",
+            "parent": {"$ref": "#/body"},
+            "label": "paragraph",
+            "text": "First paragraph on page 1.",
+            "prov": [{"page_no": 1, "bbox": {"l": 10, "t": 40, "r": 500, "b": 80}}],
+        },
+        {
+            "self_ref": "#/texts/2",
+            "parent": {"$ref": "#/body"},
+            "label": "paragraph",
+            "text": "Continued on page 2.",
+            "prov": [{"page_no": 2, "bbox": {"l": 10, "t": 40, "r": 500, "b": 80}}],
+        },
+    ],
+    "tables": [
+        {
+            "self_ref": "#/tables/0",
+            "parent": {"$ref": "#/body"},
+            "label": "table",
+            "text": "",
+            "data": {"num_rows": 2, "num_cols": 2},
+            "prov": [{"page_no": 2, "bbox": {"l": 10, "t": 90, "r": 500, "b": 200}}],
+        }
+    ],
+    "pictures": [],
+    "groups": [],
+}
+
+
+def _ids(items, group=None):
+    return [i["id"] for i in items if group is None or i.get("group") == group]
+
+
+def _types(edges):
+    return [e["type"] for e in edges]
+
+
+def test_builds_document_page_and_element_nodes():
+    payload = build_graph_payload(json.dumps(FIXTURE), doc_id="doc-fixture", title="fixture.pdf")
+    assert payload.doc_id == "doc-fixture"
+    assert payload.page_count == 2
+    assert payload.truncated is False
+
+    assert _ids(payload.nodes, group="document") == ["doc::doc-fixture"]
+    assert set(_ids(payload.nodes, group="page")) == {"page::1", "page::2"}
+    assert set(_ids(payload.nodes, group="element")) == {
+        "elem::#/texts/0",
+        "elem::#/texts/1",
+        "elem::#/texts/2",
+        "elem::#/tables/0",
+    }
+
+
+def test_element_keeps_docling_label_and_specific_label():
+    payload = build_graph_payload(json.dumps(FIXTURE), doc_id="doc-fixture")
+    by_id = {n["id"]: n for n in payload.nodes}
+    section = by_id["elem::#/texts/0"]
+    para = by_id["elem::#/texts/1"]
+    table = by_id["elem::#/tables/0"]
+
+    # Specific label mirrors TreeWriter's `_LABEL_MAP`.
+    assert section["label"] == "SectionHeader"
+    assert para["label"] == "Paragraph"
+    assert table["label"] == "Table"
+    # Docling's original label is preserved (lowercased) for filtering.
+    assert section["docling_label"] == "section_header"
+    assert para["docling_label"] == "paragraph"
+
+
+def test_provs_are_carried_on_element_nodes():
+    payload = build_graph_payload(json.dumps(FIXTURE), doc_id="doc-fixture")
+    by_id = {n["id"]: n for n in payload.nodes}
+    para = by_id["elem::#/texts/1"]
+    assert para["prov_page"] == 1
+    assert len(para["provs"]) == 1
+    assert para["provs"][0]["bbox_l"] == 10.0
+    assert para["provs"][0]["page_no"] == 1
+
+
+def test_has_root_parent_next_and_on_page_edges():
+    payload = build_graph_payload(json.dumps(FIXTURE), doc_id="doc-fixture")
+    types = _types(payload.edges)
+    # 4 top-level children => 4 HAS_ROOT.
+    assert types.count("HAS_ROOT") == 4
+    # 4 elements in reading order => 3 NEXT edges.
+    assert types.count("NEXT") == 3
+    # 4 elements each on 1 page => 4 ON_PAGE edges.
+    assert types.count("ON_PAGE") == 4
+    # No PARENT_OF in this flat fixture (all parents are #/body).
+    assert types.count("PARENT_OF") == 0
+
+
+def test_on_page_edges_point_to_correct_pages():
+    payload = build_graph_payload(json.dumps(FIXTURE), doc_id="doc-fixture")
+    on_page = [(e["source"], e["target"]) for e in payload.edges if e["type"] == "ON_PAGE"]
+    assert ("elem::#/texts/0", "page::1") in on_page
+    assert ("elem::#/texts/2", "page::2") in on_page
+    assert ("elem::#/tables/0", "page::2") in on_page
+
+
+def test_on_page_dedups_when_element_has_multiple_provs_same_page():
+    # Paragraph with two provs on the same page — we expect ONE ON_PAGE edge.
+    fixture = {
+        **FIXTURE,
+        "texts": [
+            {
+                "self_ref": "#/texts/0",
+                "parent": {"$ref": "#/body"},
+                "label": "paragraph",
+                "text": "Split across two provs same page",
+                "prov": [
+                    {"page_no": 1, "bbox": {"l": 0, "t": 0, "r": 10, "b": 10}},
+                    {"page_no": 1, "bbox": {"l": 0, "t": 20, "r": 10, "b": 30}},
+                ],
+            },
+        ],
+        "tables": [],
+        "body": {"self_ref": "#/body", "children": [{"$ref": "#/texts/0"}]},
+    }
+    payload = build_graph_payload(json.dumps(fixture), doc_id="doc-split")
+    on_page = [e for e in payload.edges if e["type"] == "ON_PAGE"]
+    assert len(on_page) == 1
+
+
+def test_parent_of_edges_when_items_are_nested():
+    fixture = {
+        "pages": {"1": {"page_no": 1, "size": {"width": 595, "height": 842}}},
+        "body": {"self_ref": "#/body", "children": [{"$ref": "#/texts/0"}]},
+        "texts": [
+            {
+                "self_ref": "#/texts/0",
+                "parent": {"$ref": "#/body"},
+                "label": "section_header",
+                "text": "Chapter",
+                "level": 1,
+                "children": [{"$ref": "#/texts/1"}],
+                "prov": [{"page_no": 1, "bbox": {"l": 0, "t": 0, "r": 10, "b": 10}}],
+            },
+            {
+                "self_ref": "#/texts/1",
+                "parent": {"$ref": "#/texts/0"},
+                "label": "paragraph",
+                "text": "Body",
+                "prov": [{"page_no": 1, "bbox": {"l": 0, "t": 20, "r": 10, "b": 30}}],
+            },
+        ],
+        "tables": [],
+        "pictures": [],
+        "groups": [],
+    }
+    payload = build_graph_payload(json.dumps(fixture), doc_id="doc-nested")
+    parents = [(e["source"], e["target"]) for e in payload.edges if e["type"] == "PARENT_OF"]
+    assert parents == [("elem::#/texts/0", "elem::#/texts/1")]
+    roots = [e for e in payload.edges if e["type"] == "HAS_ROOT"]
+    assert len(roots) == 1  # only the section is a direct child of body
+    # NEXT follows DFS order: #/texts/0 -> #/texts/1
+    nexts = [(e["source"], e["target"]) for e in payload.edges if e["type"] == "NEXT"]
+    assert nexts == [("elem::#/texts/0", "elem::#/texts/1")]
+
+
+def test_truncated_when_page_count_exceeds_cap():
+    fixture = {
+        **FIXTURE,
+        "pages": {str(i): {"page_no": i, "size": {"width": 1, "height": 1}} for i in range(1, 12)},
+    }
+    payload = build_graph_payload(json.dumps(fixture), doc_id="doc-big", max_pages=10)
+    assert payload.truncated is True
+    assert payload.page_count == 11
+    assert payload.nodes == []
+    assert payload.edges == []
+
+
+def test_title_is_surfaced_on_document_node():
+    payload = build_graph_payload(json.dumps(FIXTURE), doc_id="doc-fixture", title="My Doc.pdf")
+    doc_node = next(n for n in payload.nodes if n["group"] == "document")
+    assert doc_node["title"] == "My Doc.pdf"
+
+
+def test_inline_group_collapses_into_single_paragraph():
+    """Issue #197: an InlineGroup + N child `text` items must yield ONE
+    Paragraph node carrying the joined text and the union of children's provs."""
+    fixture = {
+        "pages": {"1": {"page_no": 1, "size": {"width": 595, "height": 842}}},
+        "body": {
+            "self_ref": "#/body",
+            "children": [{"$ref": "#/texts/0"}, {"$ref": "#/groups/0"}],
+        },
+        "texts": [
+            {
+                "self_ref": "#/texts/0",
+                "parent": {"$ref": "#/body"},
+                "label": "section_header",
+                "text": "Heading",
+                "level": 1,
+                "prov": [{"page_no": 1, "bbox": {"l": 10, "t": 10, "r": 100, "b": 30}}],
+            },
+            {
+                "self_ref": "#/texts/1",
+                "parent": {"$ref": "#/groups/0"},
+                "label": "text",
+                "text": "Hello",
+                "prov": [{"page_no": 1, "bbox": {"l": 10, "t": 40, "r": 50, "b": 60}}],
+            },
+            {
+                "self_ref": "#/texts/2",
+                "parent": {"$ref": "#/groups/0"},
+                "label": "text",
+                "text": "world",
+                "prov": [{"page_no": 1, "bbox": {"l": 55, "t": 40, "r": 100, "b": 60}}],
+            },
+            {
+                "self_ref": "#/texts/3",
+                "parent": {"$ref": "#/groups/0"},
+                "label": "text",
+                "text": "!",
+                "prov": [{"page_no": 1, "bbox": {"l": 105, "t": 40, "r": 110, "b": 60}}],
+            },
+        ],
+        "tables": [],
+        "pictures": [],
+        "groups": [
+            {
+                "self_ref": "#/groups/0",
+                "parent": {"$ref": "#/body"},
+                "label": "inline",
+                "children": [
+                    {"$ref": "#/texts/1"},
+                    {"$ref": "#/texts/2"},
+                    {"$ref": "#/texts/3"},
+                ],
+            },
+        ],
+    }
+    payload = build_graph_payload(json.dumps(fixture), doc_id="doc-inline")
+
+    # Section header + collapsed inline group only — NOT 5 (+3 style runs).
+    elements = [n for n in payload.nodes if n.get("group") == "element"]
+    assert {e["self_ref"] for e in elements} == {"#/texts/0", "#/groups/0"}
+
+    inline = next(n for n in elements if n["self_ref"] == "#/groups/0")
+    assert inline["label"] == "Paragraph"
+    assert inline["text"] == "Hello world !"
+    # Provs union: 3 children x 1 prov each = 3 provs on the collapsed node.
+    assert len(inline["provs"]) == 3
+    assert all(p.get("page_no") == 1 for p in inline["provs"])
+
+    # NEXT follows the post-collapse reading order — section_header → inline only.
+    nexts = [(e["source"], e["target"]) for e in payload.edges if e["type"] == "NEXT"]
+    assert nexts == [("elem::#/texts/0", "elem::#/groups/0")]
+
+    # ON_PAGE edges still wire the surviving elements to the page (one per
+    # element, deduped by `seen_pages`).
+    on_page = [e for e in payload.edges if e["type"] == "ON_PAGE"]
+    assert sorted(e["source"] for e in on_page) == ["elem::#/groups/0", "elem::#/texts/0"]
+
+
+def test_picture_internal_labels_are_skipped():
+    """Issue #197: a Picture's `children` (flowchart / diagram text labels
+    extracted by Docling's layout model) must not become standalone graph
+    nodes — they drown the figure in dozens of tiny labels. The picture
+    itself stays, and a caption (separate `parent` chain on body) is
+    unaffected."""
+    fixture = {
+        "pages": {"1": {"page_no": 1, "size": {"width": 595, "height": 842}}},
+        "body": {
+            "self_ref": "#/body",
+            "children": [{"$ref": "#/texts/0"}, {"$ref": "#/pictures/0"}],
+        },
+        "texts": [
+            # Caption — separate item under body, referenced from the picture's
+            # `captions` field (not its `children`). Survives the collapse.
+            {
+                "self_ref": "#/texts/0",
+                "parent": {"$ref": "#/body"},
+                "label": "caption",
+                "text": "Figure 1: Sketch of Docling's pipelines.",
+                "prov": [{"page_no": 1, "bbox": {"l": 0, "t": 700, "r": 500, "b": 720}}],
+            },
+            # Internal labels of the figure (flowchart boxes). Live in
+            # `texts[]` with `parent` pointing at the picture.
+            *[
+                {
+                    "self_ref": f"#/texts/{i}",
+                    "parent": {"$ref": "#/pictures/0"},
+                    "label": "text",
+                    "text": label,
+                    "prov": [
+                        {
+                            "page_no": 1,
+                            "bbox": {"l": 100 + i * 30, "t": 200, "r": 130 + i * 30, "b": 220},
+                        }
+                    ],
+                }
+                for i, label in enumerate(
+                    ["Parse", "Build", "Enrich", "Assemble", "Document"], start=1
+                )
+            ],
+        ],
+        "tables": [],
+        "pictures": [
+            {
+                "self_ref": "#/pictures/0",
+                "parent": {"$ref": "#/body"},
+                "label": "picture",
+                "children": [
+                    {"$ref": "#/texts/1"},
+                    {"$ref": "#/texts/2"},
+                    {"$ref": "#/texts/3"},
+                    {"$ref": "#/texts/4"},
+                    {"$ref": "#/texts/5"},
+                ],
+                "captions": [{"$ref": "#/texts/0"}],
+                "prov": [{"page_no": 1, "bbox": {"l": 90, "t": 100, "r": 510, "b": 600}}],
+            }
+        ],
+        "groups": [],
+    }
+    payload = build_graph_payload(json.dumps(fixture), doc_id="doc-pic")
+
+    elements = [n for n in payload.nodes if n.get("group") == "element"]
+    refs = {e["self_ref"] for e in elements}
+    # Picture + caption only — the 5 internal labels must be dropped.
+    assert refs == {"#/pictures/0", "#/texts/0"}
+
+    # Picture keeps its own label / text / prov (no inline-style override).
+    pic = next(e for e in elements if e["self_ref"] == "#/pictures/0")
+    assert pic["label"] == "Figure"
+    assert pic["docling_label"] == "picture"
+
+    # No PARENT_OF edges from the picture to its (skipped) children.
+    parent_edges = [e for e in payload.edges if e["type"] == "PARENT_OF"]
+    assert all(e["source"] != "elem::#/pictures/0" for e in parent_edges)
+
+    # NEXT chain only includes surviving elements.
+    next_targets = [e["target"] for e in payload.edges if e["type"] == "NEXT"]
+    assert all(t in {"elem::#/texts/0", "elem::#/pictures/0"} for t in next_targets)
+
+
+def test_element_text_is_capped_at_200_chars():
+    long = "x" * 500
+    fixture = {
+        "pages": {"1": {"page_no": 1, "size": {"width": 1, "height": 1}}},
+        "body": {"self_ref": "#/body", "children": [{"$ref": "#/texts/0"}]},
+        "texts": [
+            {
+                "self_ref": "#/texts/0",
+                "parent": {"$ref": "#/body"},
+                "label": "paragraph",
+                "text": long,
+                "prov": [{"page_no": 1}],
+            }
+        ],
+        "tables": [],
+        "pictures": [],
+        "groups": [],
+    }
+    payload = build_graph_payload(json.dumps(fixture), doc_id="doc-long")
+    para = next(n for n in payload.nodes if n.get("self_ref") == "#/texts/0")
+    assert len(para["text"]) == 200

document-parser/tests/test_graph_api.py

@@ -0,0 +1,114 @@
+"""Tests for `api.graph` — the `/graph` (Neo4j) and `/reasoning-graph`
+(SQLite) endpoints. Neo4j itself is not exercised here; `/graph` is covered
+by the integration tests under `tests/neo4j/`. This file focuses on the
+SQLite-backed reasoning endpoint and the error paths.
+"""
+
+from __future__ import annotations
+
+import json
+from unittest.mock import AsyncMock
+
+import pytest
+from fastapi import FastAPI
+from fastapi.testclient import TestClient
+
+from api.graph import router
+from domain.models import AnalysisJob
+
+FIXTURE = {
+    "pages": {"1": {"page_no": 1, "size": {"width": 595, "height": 842}}},
+    "body": {"self_ref": "#/body", "children": [{"$ref": "#/texts/0"}]},
+    "texts": [
+        {
+            "self_ref": "#/texts/0",
+            "parent": {"$ref": "#/body"},
+            "label": "section_header",
+            "text": "Hello",
+            "level": 1,
+            "prov": [{"page_no": 1, "bbox": {"l": 0, "t": 0, "r": 10, "b": 10}}],
+        }
+    ],
+    "tables": [],
+    "pictures": [],
+    "groups": [],
+}
+
+
+def _job_with_doc_json() -> AnalysisJob:
+    job = AnalysisJob(document_id="doc-1")
+    job.document_filename = "hello.pdf"
+    job.mark_running()
+    job.mark_completed(
+        markdown="# Hello",
+        html="<h1>Hello</h1>",
+        pages_json="[]",
+        document_json=json.dumps(FIXTURE),
+        chunks_json="[]",
+    )
+    return job
+
+
+@pytest.fixture
+def mock_analysis_repo() -> AsyncMock:
+    repo = AsyncMock()
+    repo.find_latest_completed_by_document.return_value = _job_with_doc_json()
+    return repo
+
+
+@pytest.fixture
+def client(mock_analysis_repo: AsyncMock) -> TestClient:
+    app = FastAPI()
+    app.include_router(router)
+    app.state.analysis_repo = mock_analysis_repo
+    app.state.neo4j = None  # /reasoning-graph must not need Neo4j
+    return TestClient(app)
+
+
+class TestReasoningGraph:
+    def test_returns_payload_built_from_sqlite_json(self, client: TestClient) -> None:
+        resp = client.get("/api/documents/doc-1/reasoning-graph")
+        assert resp.status_code == 200
+        data = resp.json()
+        assert data["doc_id"] == "doc-1"
+        assert data["page_count"] == 1
+        assert data["truncated"] is False
+
+        groups = {n["group"] for n in data["nodes"]}
+        assert groups == {"document", "page", "element"}
+
+        edge_types = {e["type"] for e in data["edges"]}
+        # HAS_ROOT + ON_PAGE expected; NEXT absent (single element so no chain).
+        assert edge_types == {"HAS_ROOT", "ON_PAGE"}
+
+    def test_404_when_no_completed_analysis(
+        self, client: TestClient, mock_analysis_repo: AsyncMock
+    ) -> None:
+        mock_analysis_repo.find_latest_completed_by_document.return_value = None
+        resp = client.get("/api/documents/doc-1/reasoning-graph")
+        assert resp.status_code == 404
+
+    def test_404_when_analysis_has_no_document_json(
+        self, client: TestClient, mock_analysis_repo: AsyncMock
+    ) -> None:
+        job = AnalysisJob(document_id="doc-1")
+        job.mark_running()
+        job.mark_completed(
+            markdown="", html="", pages_json="[]", document_json=None, chunks_json="[]"
+        )
+        mock_analysis_repo.find_latest_completed_by_document.return_value = job
+        resp = client.get("/api/documents/doc-1/reasoning-graph")
+        assert resp.status_code == 404
+
+    def test_does_not_need_neo4j(self, client: TestClient) -> None:
+        # `app.state.neo4j = None` and the endpoint still serves — proves the
+        # reasoning graph is fully decoupled from the Neo4j provider.
+        resp = client.get("/api/documents/doc-1/reasoning-graph")
+        assert resp.status_code == 200
+
+
+class TestPrimeEndpointRemoved:
+    def test_graph_prime_endpoint_is_gone(self, client: TestClient) -> None:
+        # Guardrail — if someone reintroduces /graph/prime we want a failing test.
+        resp = client.post("/api/documents/doc-1/graph/prime")
+        assert resp.status_code in (404, 405)

document-parser/tests/test_reasoning_api.py

@@ -0,0 +1,261 @@
+"""Tests for `api.reasoning` — the live `docling-agent` RAG runner endpoint.
+
+docling-agent + mellea are NOT installed in the CI test env (heavy deps).
+The endpoint does a lazy import inside the handler; we stub the modules via
+`sys.modules` injection so the tests cover the real code path without
+bringing in Ollama, mellea, or LLM clients.
+"""
+
+from __future__ import annotations
+
+import sys
+import types
+from dataclasses import replace
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+from fastapi import FastAPI
+from fastapi.testclient import TestClient
+
+from api import reasoning as reasoning_module
+from api.reasoning import router
+from domain.models import AnalysisJob
+
+
+def _patched_settings(monkeypatch, **overrides):
+    """Replace `api.reasoning.settings` with a frozen dataclass copy carrying
+    the given overrides. `Settings` is frozen, so attribute-level monkeypatch
+    doesn't work — we swap the whole instance on the module.
+    """
+    new_settings = replace(reasoning_module.settings, **overrides)
+    monkeypatch.setattr(reasoning_module, "settings", new_settings)
+    return new_settings
+
+
+def _job_with_doc_json() -> AnalysisJob:
+    job = AnalysisJob(document_id="doc-1")
+    job.document_filename = "hello.pdf"
+    job.mark_running()
+    job.mark_completed(
+        markdown="# Hello",
+        html="<h1>Hello</h1>",
+        pages_json="[]",
+        # Minimal placeholder — the test stubs `DoclingDocument.model_validate_json`
+        # so the content doesn't need to be a real DoclingDocument.
+        document_json='{"stub": true}',
+        chunks_json="[]",
+    )
+    return job
+
+
+@pytest.fixture
+def mock_analysis_repo() -> AsyncMock:
+    repo = AsyncMock()
+    repo.find_latest_completed_by_document.return_value = _job_with_doc_json()
+    return repo
+
+
+@pytest.fixture
+def stub_docling_agent(monkeypatch):
+    """Inject fake `docling_agent.agents` + `docling_core.types.doc.document`
+    modules so the endpoint's lazy imports resolve to our stubs.
+
+    Returns the `DoclingRAGAgent` stub class so tests can assert on its calls
+    / configure its `_rag_loop` return value.
+    """
+    fake_result = MagicMock()
+    fake_result.answer = "stub answer"
+    fake_result.converged = True
+    fake_result.iterations = [
+        MagicMock(
+            model_dump=lambda: {
+                "iteration": 1,
+                "section_ref": "#/texts/0",
+                "reason": "looks relevant",
+                "section_text_length": 42,
+                "can_answer": True,
+                "response": "stub answer",
+            }
+        )
+    ]
+
+    agent_instance = MagicMock()
+    agent_instance._rag_loop.return_value = fake_result
+    agent_class = MagicMock(return_value=agent_instance)
+
+    fake_agents_mod = types.ModuleType("docling_agent.agents")
+    fake_agents_mod.DoclingRAGAgent = agent_class
+    fake_root_mod = types.ModuleType("docling_agent")
+    fake_root_mod.agents = fake_agents_mod
+
+    fake_doc_class = MagicMock()
+    fake_doc_class.model_validate_json = MagicMock(return_value="fake-doc-instance")
+    fake_doc_mod = types.ModuleType("docling_core.types.doc.document")
+    fake_doc_mod.DoclingDocument = fake_doc_class
+
+    # Stub `mellea.backends.model_ids.ModelIdentifier` — the endpoint wraps
+    # the string model_id in this dataclass before handing to DoclingRAGAgent.
+    # Identity-like: stores the kwargs so tests can assert on `ollama_name`.
+    def fake_model_identifier(**kwargs):
+        m = MagicMock()
+        m.ollama_name = kwargs.get("ollama_name")
+        m.openai_name = kwargs.get("openai_name")
+        return m
+
+    fake_model_ids_mod = types.ModuleType("mellea.backends.model_ids")
+    fake_model_ids_mod.ModelIdentifier = fake_model_identifier
+    fake_backends_mod = types.ModuleType("mellea.backends")
+    fake_backends_mod.model_ids = fake_model_ids_mod
+    fake_mellea_mod = types.ModuleType("mellea")
+    fake_mellea_mod.backends = fake_backends_mod
+
+    monkeypatch.setitem(sys.modules, "docling_agent", fake_root_mod)
+    monkeypatch.setitem(sys.modules, "docling_agent.agents", fake_agents_mod)
+    monkeypatch.setitem(sys.modules, "docling_core.types.doc.document", fake_doc_mod)
+    monkeypatch.setitem(sys.modules, "mellea", fake_mellea_mod)
+    monkeypatch.setitem(sys.modules, "mellea.backends", fake_backends_mod)
+    monkeypatch.setitem(sys.modules, "mellea.backends.model_ids", fake_model_ids_mod)
+
+    return agent_class, agent_instance, fake_result
+
+
+@pytest.fixture
+def client(mock_analysis_repo: AsyncMock) -> TestClient:
+    app = FastAPI()
+    app.include_router(router)
+    app.state.analysis_repo = mock_analysis_repo
+    return TestClient(app)
+
+
+class TestRagDisabled:
+    def test_503_when_flag_off(self, client: TestClient, monkeypatch) -> None:
+        _patched_settings(monkeypatch, rag_enabled=False)
+        resp = client.post("/api/documents/doc-1/rag", json={"query": "Q"})
+        assert resp.status_code == 503
+        assert "RAG_ENABLED" in resp.json()["detail"]
+
+
+class TestRagValidation:
+    def test_400_when_query_empty(self, client: TestClient, monkeypatch) -> None:
+        _patched_settings(monkeypatch, rag_enabled=True)
+        resp = client.post("/api/documents/doc-1/rag", json={"query": "   "})
+        assert resp.status_code == 400
+
+    def test_404_when_no_completed_analysis(
+        self, client: TestClient, mock_analysis_repo: AsyncMock, monkeypatch
+    ) -> None:
+        _patched_settings(monkeypatch, rag_enabled=True)
+        mock_analysis_repo.find_latest_completed_by_document.return_value = None
+        resp = client.post("/api/documents/doc-1/rag", json={"query": "Q"})
+        assert resp.status_code == 404
+
+
+class TestRagSuccess:
+    def test_returns_rag_result_shape(
+        self, client: TestClient, stub_docling_agent, monkeypatch
+    ) -> None:
+        _patched_settings(monkeypatch, rag_enabled=True)
+        _agent_class, _agent_instance, _fake_result = stub_docling_agent
+
+        resp = client.post("/api/documents/doc-1/rag", json={"query": "What is this?"})
+        assert resp.status_code == 200
+        data = resp.json()
+        assert data["answer"] == "stub answer"
+        assert data["converged"] is True
+        assert len(data["iterations"]) == 1
+        it = data["iterations"][0]
+        assert it["iteration"] == 1
+        assert it["section_ref"] == "#/texts/0"
+        assert it["can_answer"] is True
+
+    def test_calls_rag_loop_with_query_and_doc(
+        self, client: TestClient, stub_docling_agent, monkeypatch
+    ) -> None:
+        _patched_settings(monkeypatch, rag_enabled=True)
+        _agent_class, agent_instance, _ = stub_docling_agent
+
+        client.post("/api/documents/doc-1/rag", json={"query": "Hello?"})
+
+        agent_instance._rag_loop.assert_called_once()
+        kwargs = agent_instance._rag_loop.call_args.kwargs
+        assert kwargs["query"] == "Hello?"
+        # The stub returns the string "fake-doc-instance" from model_validate_json
+        # and we pass it straight through to `doc=`.
+        assert kwargs["doc"] == "fake-doc-instance"
+
+    def test_uses_default_model_id_when_not_overridden(
+        self, client: TestClient, stub_docling_agent, monkeypatch
+    ) -> None:
+        _patched_settings(monkeypatch, rag_enabled=True, rag_model_id="custom-model:7b")
+        agent_class, _, _ = stub_docling_agent
+
+        client.post("/api/documents/doc-1/rag", json={"query": "Q"})
+
+        agent_class.assert_called_once()
+        # model_id is wrapped in a ModelIdentifier(ollama_name=...) dataclass
+        # before reaching the agent — the stub exposes the field for assertion.
+        passed = agent_class.call_args.kwargs["model_id"]
+        assert passed.ollama_name == "custom-model:7b"
+
+    def test_per_request_model_id_override_wins(
+        self, client: TestClient, stub_docling_agent, monkeypatch
+    ) -> None:
+        _patched_settings(monkeypatch, rag_enabled=True, rag_model_id="default:7b")
+        agent_class, _, _ = stub_docling_agent
+
+        client.post("/api/documents/doc-1/rag", json={"query": "Q", "model_id": "override:13b"})
+
+        passed = agent_class.call_args.kwargs["model_id"]
+        assert passed.ollama_name == "override:13b"
+
+    def test_sets_ollama_host_env_from_settings(
+        self, client: TestClient, stub_docling_agent, monkeypatch
+    ) -> None:
+        import os
+
+        _patched_settings(monkeypatch, rag_enabled=True, ollama_host="http://ollama:11434")
+
+        client.post("/api/documents/doc-1/rag", json={"query": "Q"})
+        assert os.environ["OLLAMA_HOST"] == "http://ollama:11434"
+
+
+class TestRagDepsMissing:
+    def test_503_when_docling_agent_not_installed(self, client: TestClient, monkeypatch) -> None:
+        _patched_settings(monkeypatch, rag_enabled=True)
+        # Simulate the import failing: remove any stub and ensure the name
+        # resolves to a module that raises on attribute access.
+        monkeypatch.setitem(sys.modules, "docling_agent.agents", None)
+
+        resp = client.post("/api/documents/doc-1/rag", json={"query": "Q"})
+        assert resp.status_code == 503
+        assert "docling-agent" in resp.json()["detail"]
+
+
+class TestRagUpstreamFailure:
+    def test_502_when_docling_agent_raises_indexerror(
+        self, client: TestClient, stub_docling_agent, monkeypatch
+    ) -> None:
+        """Known docling-agent bug: `find_json_dicts(answer.value)[0]` raises
+        `IndexError` when the model fails to produce parseable JSON after
+        retries. Our endpoint must surface a 502 with a human-readable
+        message, not a 500 stack trace."""
+        _patched_settings(monkeypatch, rag_enabled=True, rag_model_id="granite4:micro-h")
+        _agent_class, agent_instance, _ = stub_docling_agent
+        agent_instance._rag_loop.side_effect = IndexError("list index out of range")
+
+        resp = client.post("/api/documents/doc-1/rag", json={"query": "Quelle tarification ?"})
+        assert resp.status_code == 502
+        detail = resp.json()["detail"]
+        assert "granite4:micro-h" in detail
+        assert "parseable" in detail or "rephrase" in detail
+
+    def test_500_for_other_unexpected_errors(
+        self, client: TestClient, stub_docling_agent, monkeypatch
+    ) -> None:
+        _patched_settings(monkeypatch, rag_enabled=True)
+        _agent_class, agent_instance, _ = stub_docling_agent
+        agent_instance._rag_loop.side_effect = RuntimeError("Ollama unreachable")
+
+        resp = client.post("/api/documents/doc-1/rag", json={"query": "Q"})
+        assert resp.status_code == 500
+        assert "Ollama unreachable" in resp.json()["detail"]

document-parser/tests/test_repos.py

@@ -153,6 +153,49 @@ class TestAnalysisRepo:
         deleted = await analysis_repo.delete("nonexistent")
         assert deleted is False
 
+    async def test_find_latest_completed_by_document(self, document_repo, analysis_repo):
+        """Reasoning tunnel helper: latest COMPLETED analysis with document_json."""
+        await self._insert_doc(document_repo)
+
+        # Each job must be insert()'d before update_status can touch it.
+        # Scenarios: pending (excluded — not COMPLETED), old completed without
+        # document_json (excluded — NULL json), recent completed with
+        # document_json (the one we want), running (excluded).
+        pending = AnalysisJob(id="job-pending", document_id="doc-1")
+        await analysis_repo.insert(pending)
+
+        old_completed = AnalysisJob(id="job-old", document_id="doc-1")
+        await analysis_repo.insert(old_completed)
+        old_completed.mark_running()
+        old_completed.mark_completed(markdown="", html="", pages_json="[]")
+        await analysis_repo.update_status(old_completed)
+
+        latest = AnalysisJob(id="job-latest", document_id="doc-1")
+        await analysis_repo.insert(latest)
+        latest.mark_running()
+        latest.mark_completed(
+            markdown="md",
+            html="<p/>",
+            pages_json="[]",
+            document_json='{"body":{"children":[]},"texts":[]}',
+        )
+        await analysis_repo.update_status(latest)
+
+        running = AnalysisJob(id="job-running", document_id="doc-1")
+        await analysis_repo.insert(running)
+        running.mark_running()
+        await analysis_repo.update_status(running)
+
+        found = await analysis_repo.find_latest_completed_by_document("doc-1")
+        assert found is not None
+        assert found.id == "job-latest"
+        assert found.document_json is not None
+
+    async def test_find_latest_completed_by_document_none(self, document_repo, analysis_repo):
+        await self._insert_doc(document_repo)
+        found = await analysis_repo.find_latest_completed_by_document("doc-1")
+        assert found is None
+
     async def test_delete_by_document(self, document_repo, analysis_repo):
         await self._insert_doc(document_repo)
         for i in range(3):

document-parser/tests/test_serve_converter.py

@@ -39,7 +39,6 @@ class TestBuildFormData:
         assert data["do_picture_classification"] == "false"
         assert data["do_picture_description"] == "false"
         assert data["include_images"] == "false"
-        assert data["generate_page_images"] == "false"
         assert data["images_scale"] == "1.0"
         assert set(data["to_formats"]) == {"md", "html", "json"}
 
@@ -56,9 +55,14 @@ class TestBuildFormData:
         assert data["images_scale"] == "2.0"
         assert data["include_images"] == "true"
 
-    def test_page_range_included_when_set(self):
+    def test_no_generate_page_images_field(self):
+        """generate_page_images is a PdfPipelineOptions field, not a Serve field."""
+        data = _build_form_data(ConversionOptions())
+        assert "generate_page_images" not in data
+
+    def test_page_range_as_repeated_fields(self):
         data = _build_form_data(ConversionOptions(), page_range=(11, 20))
-        assert data["page_range"] == "11-20"
+        assert data["page_range"] == ["11", "20"]
 
     def test_page_range_absent_when_none(self):
         data = _build_form_data(ConversionOptions())
@@ -402,11 +406,12 @@ class TestServeConverterConvert:
         assert len(result.pages[0].elements) == 1
         assert result.pages[0].elements[0].type == "title"
 
-        # Verify form fields sent as dict with list for repeated keys
+        # Verify form fields sent correctly
         call_kwargs = mock_client.post.call_args
         sent_data = call_kwargs.kwargs.get("data", {})
         assert sent_data["do_ocr"] == "true"
         assert set(sent_data["to_formats"]) == {"md", "html", "json"}
+        assert "generate_page_images" not in sent_data
 
     @pytest.mark.asyncio
     async def test_http_error_raises(self, tmp_path):
@@ -414,6 +419,8 @@ class TestServeConverterConvert:
         test_file.write_bytes(b"%PDF-1.4 fake content")
 
         mock_response = MagicMock()
+        mock_response.status_code = 500
+        mock_response.text = "Internal Server Error"
         mock_response.raise_for_status.side_effect = httpx.HTTPStatusError(
             "Server Error",
             request=MagicMock(),
@@ -510,3 +517,17 @@ class TestConverterWiring:
             converter = _build_converter()
         assert isinstance(converter, ServeConverter)
         assert converter._api_key == "my-key"
+
+    def test_remote_engine_builds_chunker(self):
+        """Chunker must be available in remote mode (hybrid local chunking)."""
+        from infra.local_chunker import LocalChunker
+        from infra.settings import Settings
+
+        with patch(
+            "main.settings",
+            Settings(conversion_engine="remote", docling_serve_url="http://serve:5001"),
+        ):
+            from main import _build_chunker
+
+            chunker = _build_chunker()
+        assert isinstance(chunker, LocalChunker)

document-parser/tests/test_settings.py

@@ -18,7 +18,10 @@ class TestSettingsDefaults:
         assert s.lock_timeout == 300
         assert s.max_page_count == 0
         assert s.max_file_size_mb == 50
+        assert s.max_paste_image_size_mb == 10
+        assert s.paste_allowed_image_types == ["image/png", "image/jpeg", "image/webp"]
         assert s.batch_page_size == 0
+        assert s.opensearch_default_limit == 1000
         assert s.upload_dir == "./uploads"
         assert s.db_path == "./data/docling_studio.db"
         assert "http://localhost:3000" in s.cors_origins
@@ -75,6 +78,18 @@ class TestSettingsValidation:
         with pytest.raises(ValueError, match="max_file_size_mb must be >= 0"):
             Settings(max_file_size_mb=-1)
 
+    def test_negative_max_paste_image_size_mb_rejected(self):
+        import pytest
+
+        with pytest.raises(ValueError, match="max_paste_image_size_mb must be >= 0"):
+            Settings(max_paste_image_size_mb=-1)
+
+    def test_empty_paste_allowed_image_types_rejected(self):
+        import pytest
+
+        with pytest.raises(ValueError, match="paste_allowed_image_types must not be empty"):
+            Settings(paste_allowed_image_types=[])
+
     def test_zero_max_file_size_mb_accepted(self):
         s = Settings(max_file_size_mb=0)
         assert s.max_file_size_mb == 0
@@ -103,6 +118,12 @@ class TestSettingsValidation:
         with pytest.raises(ValueError, match="lock_timeout must be > 0"):
             Settings(lock_timeout=0)
 
+    def test_zero_opensearch_default_limit_rejected(self):
+        import pytest
+
+        with pytest.raises(ValueError, match="opensearch_default_limit must be >= 1"):
+            Settings(opensearch_default_limit=0)
+
     def test_invalid_table_mode_rejected(self):
         import pytest
 
@@ -146,6 +167,7 @@ class TestSettingsFromEnv:
         monkeypatch.setenv("MAX_PAGE_COUNT", "20")
         monkeypatch.setenv("MAX_FILE_SIZE_MB", "100")
         monkeypatch.setenv("BATCH_PAGE_SIZE", "15")
+        monkeypatch.setenv("OPENSEARCH_DEFAULT_LIMIT", "500")
         monkeypatch.setenv("UPLOAD_DIR", "/data/uploads")
         monkeypatch.setenv("DB_PATH", "/data/test.db")
         monkeypatch.setenv("CORS_ORIGINS", "http://a.com, http://b.com")
@@ -163,6 +185,7 @@ class TestSettingsFromEnv:
         assert s.max_page_count == 20
         assert s.max_file_size_mb == 100
         assert s.batch_page_size == 15
+        assert s.opensearch_default_limit == 500
         assert s.upload_dir == "/data/uploads"
         assert s.db_path == "/data/test.db"
         assert s.cors_origins == ["http://a.com", "http://b.com"]

experiments/reasoning-trace/.gitignore

@@ -0,0 +1 @@
+output/

experiments/reasoning-trace/README.md

@@ -0,0 +1,139 @@
+# Reasoning Trace — R&D sandbox
+
+Goal: run `docling-agent`'s RAG loop against a document already ingested in
+Docling-Studio, capture the `RAGResult` (per-iteration reasoning trace), and
+inspect what the agent does.
+
+Fully **isolated** from the Studio backend: no deps added to
+`document-parser/`, no services modified. Just a script + uv inline deps.
+
+---
+
+## What it does
+
+1. Reads the pre-parsed `DoclingDocument` directly from Studio's SQLite
+   (`analysis_jobs.document_json`) — no PDF re-conversion.
+2. Instantiates `DoclingRAGAgent` against a local Ollama model.
+3. Calls `agent._rag_loop()` directly (the public `.run()` method discards the
+   `RAGResult`; we need the iterations to see the reasoning trace).
+4. Dumps the full `RAGResult` as JSON to `output/`.
+
+---
+
+## Prerequisites
+
+### 1. Ollama running
+```sh
+# If not already running as a service:
+ollama serve          # in another terminal
+```
+
+### 2. A model pulled
+Recommended (Peter Staar's default, ~3B params, good JSON adherence):
+```sh
+ollama pull granite4:micro-h
+```
+
+Alternative already on your machine (2 GB, may struggle with strict JSON
+rejection sampling):
+```
+llama3.2:3b
+```
+
+Bigger/more reliable but slower (20B):
+```sh
+ollama pull gpt-oss:20b
+```
+
+### 3. Pick an analysis job id
+Any `COMPLETED` row from `analysis_jobs` with a non-null `document_json`:
+
+```sh
+sqlite3 document-parser/data/docling_studio.db \
+  "SELECT aj.id, d.filename, length(aj.document_json)
+   FROM analysis_jobs aj JOIN documents d ON d.id=aj.document_id
+   WHERE aj.document_json IS NOT NULL AND aj.status='COMPLETED'
+   ORDER BY length(aj.document_json) DESC LIMIT 5;"
+```
+
+On this machine, the biggest one right now is:
+```
+722d5631-0089-44a3-a64a-7ce5b99579d3  — CCI - Conférence IA - Offre Commerciale v1.0
+```
+
+---
+
+## Run
+
+```sh
+uv run experiments/reasoning-trace/inspect_doc.py \
+    --job-id 722d5631-0089-44a3-a64a-7ce5b99579d3 \
+    --query "Quels sont les livrables principaux proposés ?" \
+    --model granite4:micro-h
+```
+
+Flags:
+- `--job-id` — required, analysis_jobs.id
+- `--query` — required, the question
+- `--model` — either a mellea catalog constant (`IBM_GRANITE_4_HYBRID_MICRO`)
+  or a raw Ollama tag (`granite4:micro-h`, `llama3.2:3b`). Default:
+  `granite4:micro-h`.
+- `--max-iters` — default 5 (agent's own default)
+- `--quiet` — disable the rich panels during the loop
+
+First run will take ~1–2 min: `uv` solves the `docling-agent` env (pulls
+docling-core, mellea, pydantic, rich, …) into a cached virtualenv. Subsequent
+runs are instant.
+
+---
+
+## Output
+
+`experiments/reasoning-trace/output/<job-id-prefix>_<utc>.json`
+
+Schema:
+```json
+{
+  "job_id": "…",
+  "filename": "…",
+  "query": "…",
+  "model": { "ollama_name": "…", "hf_model_name": "…" },
+  "max_iterations": 5,
+  "result": {
+    "answer": "…",
+    "converged": true,
+    "iterations": [
+      { "iteration": 1, "section_ref": "#/texts/3",
+        "reason": "…", "section_text_length": 412,
+        "can_answer": false, "response": "…" },
+      …
+    ]
+  }
+}
+```
+
+This is the artifact the v1 Studio endpoint (`POST /api/rag/inspect`) will
+import — so anything that works here should work there.
+
+---
+
+## Things to check on first run
+
+- **Do we actually get a trace?** `iterations` list should have ≥ 1 entries
+  (empty means "no section headers found" fallback — bad sign for the viz idea).
+- **Are `section_ref` values `#/texts/N` paths or `#/groups/N`?** Determines
+  how the resolver walks the tree.
+- **Reasoning quality**: does `reason` actually explain the pick, or is it
+  LLM filler? That affects whether the trace is worth surfacing visually.
+- **Convergence rate**: with `max_iters=5`, does a small model converge at all,
+  or hit the cap and return a partial answer?
+- **Latency**: per-iteration wall-clock on your M-series machine with granite4.
+
+---
+
+## Next step (if the above looks promising)
+
+Resolve each `iteration.section_ref` → `(page_no, bbox)` using the same
+`DoclingDocument` that was loaded here. That's the `reasoning_service.py`
+resolver described in `docs/design/reasoning-trace.md` §3.2 — implement it in
+a second script here (`resolve_trace.py`) before touching Studio.

experiments/reasoning-trace/inspect_doc.py

@@ -0,0 +1,154 @@
+#!/usr/bin/env python3
+# /// script
+# requires-python = ">=3.11"
+# dependencies = [
+#   "docling-agent",
+#   "rich",
+# ]
+# ///
+"""
+Run a docling-agent RAG inspection on a Docling-Studio analysis job and dump the RAGResult.
+
+Bypasses `DoclingRAGAgent.run()` (which discards the RAGResult) and calls the private
+`_rag_loop()` directly so we can capture the per-iteration trace.
+
+Loads the DoclingDocument from Studio's SQLite (`analysis_jobs.document_json`), so no
+re-parsing of the PDF is needed — same doc the UI is showing.
+
+Usage:
+    uv run experiments/reasoning-trace/inspect_doc.py \\
+        --job-id 722d5631-0089-44a3-a64a-7ce5b99579d3 \\
+        --query "Quels sont les points clés de l'offre ?" \\
+        --model granite4:micro-h
+
+Output:
+    experiments/reasoning-trace/output/<job-id-prefix>_<utc-timestamp>.json
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import sqlite3
+import sys
+from datetime import datetime, timezone
+from pathlib import Path
+
+from docling_agent.agent.rag import DoclingRAGAgent
+from docling_core.types.doc.document import DoclingDocument
+from mellea.backends import model_ids as M
+from mellea.backends.model_ids import ModelIdentifier
+
+HERE = Path(__file__).resolve().parent
+REPO = HERE.parents[1]
+DB_PATH = REPO / "document-parser" / "data" / "docling_studio.db"
+OUT_DIR = HERE / "output"
+
+
+def load_doc(job_id: str) -> tuple[DoclingDocument, str]:
+    if not DB_PATH.exists():
+        sys.exit(f"SQLite DB not found at {DB_PATH}")
+    con = sqlite3.connect(DB_PATH)
+    con.row_factory = sqlite3.Row
+    row = con.execute(
+        """
+        SELECT aj.document_json, d.filename
+        FROM analysis_jobs aj
+        JOIN documents d ON d.id = aj.document_id
+        WHERE aj.id = ?
+        """,
+        (job_id,),
+    ).fetchone()
+    con.close()
+    if row is None:
+        sys.exit(f"No analysis job with id {job_id}")
+    if not row["document_json"]:
+        sys.exit(f"Analysis job {job_id} has no document_json (not completed?)")
+    return DoclingDocument.model_validate_json(row["document_json"]), row["filename"]
+
+
+def resolve_model(name: str) -> ModelIdentifier:
+    """Accept either a mellea catalog constant name (e.g. 'IBM_GRANITE_4_HYBRID_MICRO')
+    or a raw Ollama tag (e.g. 'granite4:micro-h', 'llama3.2:3b')."""
+    const = getattr(M, name.upper(), None)
+    if isinstance(const, ModelIdentifier):
+        return const
+    return ModelIdentifier(ollama_name=name)
+
+
+def summarize_structure(doc: DoclingDocument) -> str:
+    from docling_core.types.doc.document import SectionHeaderItem, TitleItem
+
+    headers = [
+        item for item, _ in doc.iterate_items()
+        if isinstance(item, (TitleItem, SectionHeaderItem))
+    ]
+    return (
+        f"texts={len(doc.texts)} "
+        f"tables={len(doc.tables)} "
+        f"pictures={len(doc.pictures)} "
+        f"groups={len(doc.groups)} "
+        f"section_headers={len(headers)}"
+    )
+
+
+def main() -> None:
+    p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    p.add_argument("--job-id", required=True, help="analysis_jobs.id from Studio SQLite")
+    p.add_argument("--query", required=True, help="Question to ask the document")
+    p.add_argument(
+        "--model",
+        default="granite4:micro-h",
+        help="Ollama tag or mellea catalog constant (default: granite4:micro-h)",
+    )
+    p.add_argument("--max-iters", type=int, default=5)
+    p.add_argument("--quiet", action="store_true", help="disable rich progress panels")
+    args = p.parse_args()
+
+    print(f"→ Loading DoclingDocument from analysis {args.job_id[:8]}…")
+    doc, filename = load_doc(args.job_id)
+    print(f"  {filename}")
+    print(f"  {summarize_structure(doc)}")
+
+    model_id = resolve_model(args.model)
+    print(f"→ Model: ollama={model_id.ollama_name!r} hf={model_id.hf_model_name!r}")
+
+    agent = DoclingRAGAgent(
+        model_id=model_id,
+        tools=[],
+        max_iterations=args.max_iters,
+        verbose=not args.quiet,
+    )
+
+    print(f"→ Running RAG loop (query: {args.query!r})\n")
+    # Intentional: agent.run() discards the RAGResult. _rag_loop gives us the trace.
+    result = agent._rag_loop(query=args.query, doc=doc)
+
+    OUT_DIR.mkdir(exist_ok=True)
+    ts = datetime.now(timezone.utc).strftime("%Y%m%dT%H%M%SZ")
+    out_path = OUT_DIR / f"{args.job_id[:8]}_{ts}.json"
+    payload = {
+        "job_id": args.job_id,
+        "filename": filename,
+        "query": args.query,
+        "model": {
+            "ollama_name": model_id.ollama_name,
+            "hf_model_name": model_id.hf_model_name,
+        },
+        "max_iterations": args.max_iters,
+        "result": json.loads(result.model_dump_json()),
+    }
+    out_path.write_text(json.dumps(payload, indent=2, ensure_ascii=False))
+
+    print()
+    print(f"✓ Wrote {out_path.relative_to(REPO)}")
+    print(
+        f"  converged={result.converged}  "
+        f"iterations={len(result.iterations)}  "
+        f"answer_chars={len(result.answer)}"
+    )
+    if result.iterations:
+        print("  section_refs visited:", [it.section_ref for it in result.iterations])
+
+
+if __name__ == "__main__":
+    main()

experiments/reasoning-trace/inspect_graph.py

@@ -0,0 +1,367 @@
+#!/usr/bin/env python3
+# /// script
+# requires-python = ">=3.11"
+# dependencies = [
+#   "neo4j>=5.20,<6",
+# ]
+# ///
+"""
+Compare the Neo4j graph for a given doc_id against the source DoclingDocument
+stored in SQLite. Reports whether the graph is "well-formed" — all elements
+present, hierarchy intact, no orphans, reading order faithful.
+
+Use when you want to sanity-check the Neo4j writer before building anything
+on top of the graph (e.g. the live RAG overlay).
+
+Usage:
+    uv run experiments/reasoning-trace/inspect_graph.py \\
+        --doc-id 307ad2ba-93d8-4dfd-8e38-c1ea06d23f0d
+
+Env:
+    NEO4J_URI        default bolt://localhost:7687
+    NEO4J_USER       default neo4j
+    NEO4J_PASSWORD   default changeme
+"""
+from __future__ import annotations
+
+import argparse
+import asyncio
+import json
+import os
+import sqlite3
+import sys
+from collections import Counter
+from pathlib import Path
+
+HERE = Path(__file__).resolve().parent
+REPO = HERE.parents[1]
+DB_PATH = REPO / "document-parser" / "data" / "docling_studio.db"
+
+
+def _load_doc_json(doc_id: str) -> tuple[str, dict]:
+    """Return (filename, parsed DoclingDocument dict) for the latest completed
+    analysis of this doc."""
+    if not DB_PATH.exists():
+        sys.exit(f"SQLite DB not found at {DB_PATH}")
+    con = sqlite3.connect(DB_PATH)
+    con.row_factory = sqlite3.Row
+    row = con.execute(
+        """
+        SELECT d.filename, aj.document_json
+        FROM analysis_jobs aj
+        JOIN documents d ON d.id = aj.document_id
+        WHERE aj.document_id = ?
+          AND aj.status = 'COMPLETED'
+          AND aj.document_json IS NOT NULL
+        ORDER BY aj.completed_at DESC LIMIT 1
+        """,
+        (doc_id,),
+    ).fetchone()
+    con.close()
+    if not row:
+        sys.exit(f"No completed analysis with document_json for doc_id={doc_id}")
+    return row["filename"], json.loads(row["document_json"])
+
+
+# ---------------------------------------------------------------------------
+# Source-side summaries (DoclingDocument JSON)
+# ---------------------------------------------------------------------------
+_ITEM_LISTS = ("texts", "tables", "pictures", "groups")
+
+
+def _iter_items(doc: dict):
+    for key in _ITEM_LISTS:
+        for it in doc.get(key) or []:
+            yield key, it
+
+
+def _parent_ref(item: dict) -> str | None:
+    parent = item.get("parent")
+    if isinstance(parent, dict):
+        return parent.get("$ref") or parent.get("cref")
+    return None
+
+
+def _dfs_reading_order(doc: dict) -> list[str]:
+    by_ref: dict[str, dict] = {}
+    for _, it in _iter_items(doc):
+        ref = it.get("self_ref")
+        if ref:
+            by_ref[ref] = it
+    body = doc.get("body") or {}
+    order: list[str] = []
+
+    def walk(children):
+        for ch in children or []:
+            ref = ch.get("$ref") or ch.get("cref")
+            if not ref:
+                continue
+            order.append(ref)
+            walk((by_ref.get(ref) or {}).get("children"))
+
+    walk(body.get("children"))
+    return order
+
+
+def _summarize_source(doc: dict) -> dict:
+    items_by_kind = Counter()
+    labels = Counter()
+    roots_from_body = 0
+    with_prov = 0
+    no_text_no_caption = 0
+    section_headers: list[str] = []
+
+    for kind, it in _iter_items(doc):
+        items_by_kind[kind] += 1
+        label = (it.get("label") or "").lower()
+        labels[label] += 1
+        if _parent_ref(it) == "#/body":
+            roots_from_body += 1
+        if it.get("prov"):
+            with_prov += 1
+        if not (it.get("text") or it.get("caption")):
+            no_text_no_caption += 1
+        if label in ("section_header", "title"):
+            section_headers.append(it.get("self_ref") or "<?>")
+
+    pages = doc.get("pages") or {}
+    reading_order = _dfs_reading_order(doc)
+
+    return {
+        "filename_in_doc": doc.get("name"),
+        "items_by_kind": dict(items_by_kind),
+        "labels": dict(labels),
+        "total_items": sum(items_by_kind.values()),
+        "section_headers_count": len(section_headers),
+        "section_headers_sample": section_headers[:5],
+        "roots_from_body": roots_from_body,
+        "items_with_prov": with_prov,
+        "items_without_prov": sum(items_by_kind.values()) - with_prov,
+        "items_without_text_nor_caption": no_text_no_caption,
+        "pages_count": len(pages),
+        "reading_order_length": len(reading_order),
+        "reading_order_sample": reading_order[:5],
+    }
+
+
+# ---------------------------------------------------------------------------
+# Graph-side summaries (Neo4j)
+# ---------------------------------------------------------------------------
+async def _summarize_graph(neo, doc_id: str) -> dict:
+    async with neo.driver.session(database=neo.database) as s:
+
+        async def q(cypher: str, **params) -> list[dict]:
+            res = await s.run(cypher, doc_id=doc_id, **params)
+            return [dict(r) async for r in res]
+
+        doc = await q("MATCH (d:Document {id: $doc_id}) RETURN d.title AS title")
+        if not doc:
+            return {"exists": False}
+
+        elements = await q(
+            "MATCH (e:Element {doc_id: $doc_id}) "
+            "RETURN labels(e) AS labels, e.docling_label AS docling_label, "
+            "       e.self_ref AS self_ref, e.prov_page AS page"
+        )
+        pages = await q(
+            "MATCH (p:Page {doc_id: $doc_id}) RETURN p.page_no AS page_no"
+        )
+
+        edges_by_type = await q(
+            """
+            MATCH (n {doc_id: $doc_id})-[r]->(m)
+            WHERE (m:Document OR m:Element OR m:Page OR m:Chunk)
+            RETURN type(r) AS type, count(r) AS n
+            """
+        )
+        # HAS_ROOT starts at :Document, so n.doc_id filter above miss the edge
+        # (doc has id= not doc_id=); fetch separately.
+        has_root = await q(
+            "MATCH (d:Document {id: $doc_id})-[:HAS_ROOT]->(c:Element) "
+            "RETURN count(c) AS n"
+        )
+
+        # Orphan detection: elements with no incoming PARENT_OF and not a root.
+        orphans = await q(
+            """
+            MATCH (e:Element {doc_id: $doc_id})
+            WHERE NOT (()-[:PARENT_OF]->(e))
+              AND NOT ((:Document {id: $doc_id})-[:HAS_ROOT]->(e))
+            RETURN e.self_ref AS self_ref, e.docling_label AS label
+            """
+        )
+        # Elements with prov_page but no ON_PAGE edge.
+        on_page_missing = await q(
+            """
+            MATCH (e:Element {doc_id: $doc_id})
+            WHERE e.prov_page IS NOT NULL AND NOT (e)-[:ON_PAGE]->(:Page)
+            RETURN count(e) AS n
+            """
+        )
+        # Pages with no element attached.
+        empty_pages = await q(
+            """
+            MATCH (p:Page {doc_id: $doc_id})
+            WHERE NOT (:Element {doc_id: $doc_id})-[:ON_PAGE]->(p)
+            RETURN p.page_no AS page_no
+            """
+        )
+
+    specific_label_counter: Counter = Counter()
+    docling_label_counter: Counter = Counter()
+    for e in elements:
+        specifics = [lbl for lbl in e["labels"] if lbl != "Element"]
+        specific_label_counter[specifics[0] if specifics else "<none>"] += 1
+        docling_label_counter[e["docling_label"] or "<none>"] += 1
+
+    edges = {row["type"]: row["n"] for row in edges_by_type}
+    edges["HAS_ROOT"] = has_root[0]["n"] if has_root else 0
+
+    return {
+        "exists": True,
+        "title": doc[0]["title"],
+        "element_count": len(elements),
+        "page_count": len(pages),
+        "specific_labels": dict(specific_label_counter),
+        "docling_labels": dict(docling_label_counter),
+        "edges": edges,
+        "orphan_elements": orphans,
+        "elements_with_prov_but_no_on_page": on_page_missing[0]["n"] if on_page_missing else 0,
+        "empty_pages": [r["page_no"] for r in empty_pages],
+    }
+
+
+# ---------------------------------------------------------------------------
+# Coherence checks
+# ---------------------------------------------------------------------------
+async def _coherence(neo, doc_id: str, doc: dict) -> dict:
+    source_refs = {it.get("self_ref") for _, it in _iter_items(doc) if it.get("self_ref")}
+    async with neo.driver.session(database=neo.database) as s:
+        res = await s.run(
+            "MATCH (e:Element {doc_id: $doc_id}) RETURN e.self_ref AS r",
+            doc_id=doc_id,
+        )
+        graph_refs = {row["r"] async for row in res}
+
+    return {
+        "source_refs_count": len(source_refs),
+        "graph_refs_count": len(graph_refs),
+        "missing_in_graph": sorted(source_refs - graph_refs)[:20],
+        "missing_in_graph_count": len(source_refs - graph_refs),
+        "extra_in_graph": sorted(graph_refs - source_refs)[:20],
+        "extra_in_graph_count": len(graph_refs - source_refs),
+        "reading_order_source_len": len(_dfs_reading_order(doc)),
+    }
+
+
+# ---------------------------------------------------------------------------
+# Pretty report
+# ---------------------------------------------------------------------------
+def _section(title: str, payload: dict) -> str:
+    lines = [f"\n── {title} " + "─" * max(2, 70 - len(title))]
+    for k, v in payload.items():
+        if isinstance(v, (dict, list)):
+            lines.append(f"  {k}: {json.dumps(v, ensure_ascii=False, default=str)[:220]}")
+        else:
+            lines.append(f"  {k}: {v}")
+    return "\n".join(lines)
+
+
+def _verdict(source: dict, graph: dict, coherence: dict) -> str:
+    issues: list[str] = []
+
+    if not graph.get("exists"):
+        return "❌  No graph in Neo4j for this doc."
+
+    if coherence["missing_in_graph_count"]:
+        issues.append(
+            f"MISSING — {coherence['missing_in_graph_count']} self_ref(s) in "
+            "the source DoclingDocument are not in Neo4j"
+        )
+    if coherence["extra_in_graph_count"]:
+        issues.append(
+            f"EXTRA — {coherence['extra_in_graph_count']} self_ref(s) in the "
+            "graph are not in the source doc"
+        )
+
+    # Edge integrity:
+    edges = graph.get("edges", {})
+    expected_has_root = source["roots_from_body"]
+    actual_has_root = edges.get("HAS_ROOT", 0)
+    if actual_has_root != expected_has_root:
+        issues.append(
+            f"HAS_ROOT mismatch — source has {expected_has_root} top-level "
+            f"items, graph has {actual_has_root}"
+        )
+
+    # Reading order: NEXT chain should be reading_order_length - 1.
+    expected_next = max(source["reading_order_length"] - 1, 0)
+    actual_next = edges.get("NEXT", 0)
+    if actual_next != expected_next:
+        issues.append(
+            f"NEXT chain mismatch — expected {expected_next}, got {actual_next}"
+        )
+
+    # ON_PAGE: every item with prov should have one ON_PAGE edge.
+    expected_on_page = source["items_with_prov"]
+    actual_on_page = edges.get("ON_PAGE", 0)
+    if actual_on_page != expected_on_page:
+        issues.append(
+            f"ON_PAGE mismatch — {expected_on_page} items have prov, "
+            f"{actual_on_page} ON_PAGE edges exist"
+        )
+
+    if graph.get("elements_with_prov_but_no_on_page"):
+        issues.append(
+            f"ON_PAGE broken — {graph['elements_with_prov_but_no_on_page']} "
+            "elements have prov_page but no ON_PAGE edge"
+        )
+    if graph.get("orphan_elements"):
+        orphans = graph["orphan_elements"]
+        issues.append(f"ORPHANS — {len(orphans)} disconnected Element node(s)")
+    if graph.get("empty_pages"):
+        issues.append(
+            f"EMPTY PAGES — pages with no element attached: {graph['empty_pages']}"
+        )
+
+    if not issues:
+        return "✅  Graph is well-formed — matches source doc 1:1."
+    return "⚠️   Findings:\n  " + "\n  ".join(f"• {x}" for x in issues)
+
+
+async def main_async() -> None:
+    p = argparse.ArgumentParser(
+        description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter
+    )
+    p.add_argument("--doc-id", required=True, help="documents.id (not analysis_jobs.id)")
+    args = p.parse_args()
+
+    filename, doc = _load_doc_json(args.doc_id)
+
+    # Imports deferred so we can patch sys.path from within the script runtime.
+    sys.path.insert(0, str(REPO / "document-parser"))
+    from infra.neo4j import close_driver, get_driver
+
+    neo = await get_driver(
+        os.environ.get("NEO4J_URI", "bolt://localhost:7687"),
+        os.environ.get("NEO4J_USER", "neo4j"),
+        os.environ.get("NEO4J_PASSWORD", "changeme"),
+    )
+    try:
+        source = _summarize_source(doc)
+        graph = await _summarize_graph(neo, args.doc_id)
+        coh = await _coherence(neo, args.doc_id, doc)
+
+        print(f"Document: {filename}")
+        print(f"doc_id: {args.doc_id}")
+        print(_section("SOURCE (DoclingDocument from SQLite)", source))
+        print(_section("GRAPH (Neo4j)", graph))
+        print(_section("COHERENCE", coh))
+        print()
+        print(_verdict(source, graph, coh))
+    finally:
+        await close_driver()
+
+
+if __name__ == "__main__":
+    asyncio.run(main_async())

experiments/reasoning-trace/inspect_semantics.py

@@ -0,0 +1,175 @@
+#!/usr/bin/env python3
+# /// script
+# requires-python = ">=3.11"
+# dependencies = ["neo4j>=5.20,<6"]
+# ///
+"""
+Second-pass inspector: how well does our Neo4j graph model the *semantics*
+of a DoclingDocument — the parts Peter Staar would care about?
+
+Goes beyond structural 1:1 coverage (inspect_graph.py) to answer:
+ - Are SectionHeader levels preserved (outline hierarchy intact)?
+ - Can we reconstruct "section scope" (section → its content) from the graph?
+ - Are lists vs list-items distinguishable?
+ - Are figure captions linked to their figures?
+ - What's the depth of the actual hierarchy vs the flat body?
+"""
+from __future__ import annotations
+
+import argparse
+import asyncio
+import os
+import sys
+from pathlib import Path
+
+HERE = Path(__file__).resolve().parent
+REPO = HERE.parents[1]
+
+
+async def main_async(doc_id: str) -> None:
+    sys.path.insert(0, str(REPO / "document-parser"))
+    from infra.neo4j import close_driver, get_driver
+
+    neo = await get_driver(
+        os.environ.get("NEO4J_URI", "bolt://localhost:7687"),
+        os.environ.get("NEO4J_USER", "neo4j"),
+        os.environ.get("NEO4J_PASSWORD", "changeme"),
+    )
+    try:
+        async with neo.driver.session(database=neo.database) as s:
+
+            async def q(cypher: str, **params):
+                res = await s.run(cypher, doc_id=doc_id, **params)
+                return [dict(r) async for r in res]
+
+            print("=== 1. OUTLINE HIERARCHY (section header levels) ===")
+            rows = await q(
+                """
+                MATCH (e:SectionHeader {doc_id: $doc_id})
+                RETURN e.self_ref AS ref, e.level AS level, substring(e.text, 0, 60) AS text
+                ORDER BY e.self_ref
+                """
+            )
+            for r in rows:
+                print(f"  level={r['level']}  {r['ref']:<14}  {r['text']}")
+            levels = [r["level"] for r in rows]
+            print(f"  → distinct levels: {sorted(set(levels))}")
+
+            print("\n=== 2. DIRECT TREE DEPTH via PARENT_OF ===")
+            rows = await q(
+                """
+                MATCH (root:Element {doc_id: $doc_id})
+                WHERE NOT (()-[:PARENT_OF]->(root))
+                OPTIONAL MATCH path = (root)-[:PARENT_OF*]->(leaf)
+                WITH root, max(length(path)) AS depth
+                RETURN
+                  labels(root) AS labels,
+                  root.docling_label AS docling_label,
+                  coalesce(depth, 0) AS depth
+                ORDER BY depth DESC
+                LIMIT 10
+                """
+            )
+            for r in rows:
+                specific = [l for l in r["labels"] if l != "Element"][0]
+                print(f"  depth={r['depth']}  {specific:<15}  ({r['docling_label']})")
+
+            print("\n=== 3. SECTION SCOPE (can we infer section content from NEXT?) ===")
+            # For each section header, walk NEXT until the next section header —
+            # that's the section's content span as per docling-agent's logic.
+            rows = await q(
+                """
+                MATCH (sh:SectionHeader {doc_id: $doc_id})
+                OPTIONAL MATCH p = (sh)-[:NEXT*]->(next:SectionHeader {doc_id: $doc_id})
+                WITH sh, min(length(p)) AS span
+                RETURN
+                  sh.self_ref AS ref,
+                  sh.level AS level,
+                  coalesce(span - 1, -1) AS items_in_scope_if_span_works,
+                  substring(sh.text, 0, 50) AS title
+                ORDER BY sh.self_ref
+                """
+            )
+            for r in rows:
+                span = r["items_in_scope_if_span_works"]
+                label = f"~{span} items" if span >= 0 else "last (unknown span)"
+                print(f"  level={r['level']}  {r['ref']:<14}  {label:<18}  {r['title']}")
+
+            print("\n=== 4. LIST CONTAINER vs LIST ITEM distinction ===")
+            rows = await q(
+                """
+                MATCH (e:ListItem {doc_id: $doc_id})
+                RETURN e.docling_label AS docling_label, count(*) AS n
+                ORDER BY docling_label
+                """
+            )
+            for r in rows:
+                print(f"  docling_label={r['docling_label']:<12}  neo4j_label=:ListItem  count={r['n']}")
+            print("  ⚠️  Both 'list' (container) and 'list_item' get :ListItem in Neo4j.")
+
+            print("\n=== 5. FIGURE ↔ CAPTION linkage ===")
+            captions = await q(
+                "MATCH (c:Caption {doc_id: $doc_id}) RETURN count(c) AS n"
+            )
+            linked = await q(
+                """
+                MATCH (fig:Figure {doc_id: $doc_id})-[:PARENT_OF]-(c:Caption {doc_id: $doc_id})
+                RETURN count(DISTINCT fig) AS figs_with_caption, count(DISTINCT c) AS captions_linked
+                """
+            )
+            print(
+                f"  captions={captions[0]['n']}  "
+                f"figures_with_caption={linked[0]['figs_with_caption']}  "
+                f"captions_linked={linked[0]['captions_linked']}"
+            )
+
+            print("\n=== 6. TABLE CELL CONTENT — graph-addressable or opaque? ===")
+            rows = await q(
+                """
+                MATCH (t:Table {doc_id: $doc_id})
+                RETURN t.self_ref AS ref,
+                       CASE WHEN t.cells_json IS NOT NULL THEN 'JSON-blob' ELSE 'missing' END
+                        AS cells_mode,
+                       size(coalesce(t.cells_json, '')) AS cells_bytes
+                """
+            )
+            for r in rows:
+                print(f"  {r['ref']:<14}  cells={r['cells_mode']}  ({r['cells_bytes']} bytes)")
+            if rows:
+                print("  ⚠️  Cells are a JSON string on the Table node — not queryable as graph nodes.")
+
+            print("\n=== 7. WHAT AN AGENT VISIT LOOKS LIKE (section subgraph) ===")
+            # Pick the first section header and show what would be highlighted
+            # if we traversed its scope (via NEXT until next same-or-higher-level section).
+            rows = await q(
+                """
+                MATCH (sh:SectionHeader {doc_id: $doc_id})
+                WITH sh ORDER BY sh.self_ref LIMIT 1
+                OPTIONAL MATCH chain = (sh)-[:NEXT*0..50]->(e:Element {doc_id: $doc_id})
+                WITH sh, e, length(chain) AS pos
+                OPTIONAL MATCH (e)<-[:NEXT*0..]-(_stop:SectionHeader)
+                WHERE _stop <> sh AND _stop.level <= sh.level
+                WITH sh, e, pos
+                ORDER BY pos
+                LIMIT 8
+                RETURN pos, labels(e) AS labels, e.docling_label AS kind,
+                       substring(e.text, 0, 60) AS text
+                """
+            )
+            for r in rows:
+                specific = [l for l in r["labels"] if l != "Element"][0]
+                print(f"  pos={r['pos']:<3}  {specific:<15}  {r['text']}")
+            print(
+                "  → Visiting a section on the graph shows ONE node. Its 'scope' "
+                "(the content) must be inferred via NEXT-walk — not materialized as edges."
+            )
+
+    finally:
+        await close_driver()
+
+
+if __name__ == "__main__":
+    p = argparse.ArgumentParser(description=__doc__)
+    p.add_argument("--doc-id", required=True)
+    args = p.parse_args()
+    asyncio.run(main_async(args.doc_id))

experiments/reasoning-trace/prime_neo4j.py

@@ -0,0 +1,148 @@
+#!/usr/bin/env python3
+# /// script
+# requires-python = ">=3.11"
+# dependencies = [
+#   "neo4j>=5.20,<6",
+#   "python-dotenv>=1.0",
+# ]
+# ///
+"""
+Populate Neo4j with the DoclingDocument tree + pages for one or more analysis
+jobs, **without re-parsing the PDFs**. Reuses Studio's own TreeWriter so the
+graph is byte-identical to what an in-UI analysis would produce.
+
+Use when:
+  - Neo4j was brought up after some docs were already analyzed (orphan graphs).
+  - You want to prime a demo environment from existing SQLite state.
+
+Usage:
+    # Single job
+    uv run experiments/reasoning-trace/prime_neo4j.py \\
+        --job-id 722d5631-0089-44a3-a64a-7ce5b99579d3
+
+    # All completed analyses that have a document_json but no graph yet
+    uv run experiments/reasoning-trace/prime_neo4j.py --all-missing
+
+Env (defaults match docker-compose.dev.yml):
+    NEO4J_URI       default bolt://localhost:7687
+    NEO4J_USER      default neo4j
+    NEO4J_PASSWORD  default changeme
+"""
+from __future__ import annotations
+
+import argparse
+import asyncio
+import os
+import sqlite3
+import sys
+from datetime import UTC, datetime
+from pathlib import Path
+
+HERE = Path(__file__).resolve().parent
+REPO = HERE.parents[1]
+DB_PATH = REPO / "document-parser" / "data" / "docling_studio.db"
+
+# Studio's own TreeWriter lives in document-parser/infra/neo4j. Import it by
+# adding document-parser to sys.path — this keeps us byte-identical with what
+# the live backend writes, instead of re-implementing the walk.
+sys.path.insert(0, str(REPO / "document-parser"))
+
+
+def _fetch_row(job_id: str) -> tuple[str, str, str] | None:
+    con = sqlite3.connect(DB_PATH)
+    con.row_factory = sqlite3.Row
+    row = con.execute(
+        """
+        SELECT aj.document_id, d.filename, aj.document_json
+        FROM analysis_jobs aj
+        JOIN documents d ON d.id = aj.document_id
+        WHERE aj.id = ? AND aj.document_json IS NOT NULL
+        """,
+        (job_id,),
+    ).fetchone()
+    con.close()
+    return (row["document_id"], row["filename"], row["document_json"]) if row else None
+
+
+def _fetch_all_completed() -> list[tuple[str, str, str, str]]:
+    """Latest completed analysis per document that has a document_json."""
+    con = sqlite3.connect(DB_PATH)
+    con.row_factory = sqlite3.Row
+    rows = con.execute(
+        """
+        SELECT aj.id, aj.document_id, d.filename, aj.document_json
+        FROM analysis_jobs aj
+        JOIN documents d ON d.id = aj.document_id
+        WHERE aj.document_json IS NOT NULL
+          AND aj.status = 'COMPLETED'
+        GROUP BY aj.document_id
+        HAVING MAX(aj.completed_at)
+        """,
+    ).fetchall()
+    con.close()
+    return [(r["id"], r["document_id"], r["filename"], r["document_json"]) for r in rows]
+
+
+async def prime(job_id: str, doc_id: str, filename: str, document_json: str) -> None:
+    # Imports deferred until after sys.path is patched.
+    from infra.neo4j import bootstrap_schema, close_driver, get_driver, write_document
+
+    uri = os.environ.get("NEO4J_URI", "bolt://localhost:7687")
+    user = os.environ.get("NEO4J_USER", "neo4j")
+    pwd = os.environ.get("NEO4J_PASSWORD", "changeme")
+
+    neo = await get_driver(uri, user, pwd)
+    try:
+        # Schema is idempotent; safe to run every time.
+        await bootstrap_schema(neo)
+        result = await write_document(
+            neo,
+            doc_id=doc_id,
+            filename=filename,
+            document_json=document_json,
+        )
+        print(
+            f"  ✓ {doc_id[:8]}  {filename[:40]:<40}  "
+            f"elements={result.element_count} pages={result.page_count}"
+        )
+    finally:
+        await close_driver()
+
+
+async def main_async() -> None:
+    p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    g = p.add_mutually_exclusive_group(required=True)
+    g.add_argument("--job-id", help="analysis_jobs.id to prime")
+    g.add_argument(
+        "--all-missing",
+        action="store_true",
+        help="prime every completed analysis with a document_json (latest per doc)",
+    )
+    args = p.parse_args()
+
+    if not DB_PATH.exists():
+        sys.exit(f"SQLite DB not found at {DB_PATH}")
+
+    started = datetime.now(tz=UTC)
+    if args.job_id:
+        row = _fetch_row(args.job_id)
+        if row is None:
+            sys.exit(f"No analysis with id {args.job_id} or no document_json")
+        doc_id, filename, document_json = row
+        print(f"→ Priming Neo4j for job {args.job_id[:8]} (doc {doc_id[:8]})")
+        await prime(args.job_id, doc_id, filename, document_json)
+    else:
+        rows = _fetch_all_completed()
+        print(f"→ Priming Neo4j for {len(rows)} document(s)")
+        for job_id, doc_id, filename, document_json in rows:
+            try:
+                await prime(job_id, doc_id, filename, document_json)
+            except Exception as e:
+                print(f"  ✗ {doc_id[:8]}  {filename[:40]:<40}  FAILED: {e}")
+
+    elapsed = (datetime.now(tz=UTC) - started).total_seconds()
+    print(f"Done in {elapsed:.1f}s")
+
+
+if __name__ == "__main__":
+    asyncio.run(main_async())

frontend/env.d.ts

@@ -7,3 +7,10 @@ declare module '*.vue' {
   const component: DefineComponent<Record<string, unknown>, Record<string, unknown>, unknown>
   export default component
 }
+
+// Cytoscape plugins we use in GraphView — they ship no types. Treated as
+// opaque plugin objects; the runtime APIs they add to `cy` are called via
+// `(cy as any)` at the call site and typed loosely.
+declare module 'cytoscape-expand-collapse'
+declare module 'cytoscape-navigator'
+declare module 'cytoscape-navigator/cytoscape.js-navigator.css'

frontend/package-lock.json

@@ -1,13 +1,16 @@
 {
   "name": "docling-studio",
-  "version": "0.3.1",
+  "version": "0.4.0",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "docling-studio",
-      "version": "0.3.1",
+      "version": "0.4.0",
       "dependencies": {
+        "cytoscape": "^3.30.0",
+        "cytoscape-dagre": "^2.5.0",
+        "cytoscape-expand-collapse": "^4.1.1",
         "dompurify": "^3.3.3",
         "marked": "^17.0.4",
         "pinia": "^2.3.0",
@@ -16,6 +19,8 @@
       },
       "devDependencies": {
         "@eslint/js": "^9.0.0",
+        "@types/cytoscape": "^3.21.4",
+        "@types/cytoscape-dagre": "^2.3.3",
         "@types/dompurify": "^3.2.0",
         "@vitejs/plugin-vue": "^6.0.5",
         "@vitest/mocker": "^4.1.2",
@@ -1021,6 +1026,23 @@
         "assertion-error": "^2.0.1"
       }
     },
+    "node_modules/@types/cytoscape": {
+      "version": "3.21.9",
+      "resolved": "https://registry.npmjs.org/@types/cytoscape/-/cytoscape-3.21.9.tgz",
+      "integrity": "sha512-JyrG4tllI6jvuISPjHK9j2Xv/LTbnLekLke5otGStjFluIyA9JjgnvgZrSBsp8cEDpiTjwgZUZwpPv8TSBcoLw==",
+      "dev": true,
+      "license": "MIT"
+    },
+    "node_modules/@types/cytoscape-dagre": {
+      "version": "2.3.4",
+      "resolved": "https://registry.npmjs.org/@types/cytoscape-dagre/-/cytoscape-dagre-2.3.4.tgz",
+      "integrity": "sha512-uOGXuPfPLFoKZaegjHl9oj4tqONNJuhUl180FiJgRZ35rVijBs6J4UP1Ah6mA6S46h+7pv4ICqpgfdC3EADZlw==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "cytoscape": "^3.31"
+      }
+    },
     "node_modules/@types/deep-eql": {
       "version": "4.0.2",
       "resolved": "https://registry.npmjs.org/@types/deep-eql/-/deep-eql-4.0.2.tgz",
@@ -1824,6 +1846,45 @@
       "resolved": "https://registry.npmjs.org/csstype/-/csstype-3.2.3.tgz",
       "integrity": "sha512-z1HGKcYy2xA8AGQfwrn0PAy+PB7X/GSj3UVJW9qKyn43xWa+gl5nXmU4qqLMRzWVLFC8KusUX8T/0kCiOYpAIQ=="
     },
+    "node_modules/cytoscape": {
+      "version": "3.33.2",
+      "resolved": "https://registry.npmjs.org/cytoscape/-/cytoscape-3.33.2.tgz",
+      "integrity": "sha512-sj4HXd3DokGhzZAdjDejGvTPLqlt84vNFN8m7bGsOzDY5DyVcxIb2ejIXat2Iy7HxWhdT/N1oKyheJ5YdpsGuw==",
+      "license": "MIT",
+      "engines": {
+        "node": ">=0.10"
+      }
+    },
+    "node_modules/cytoscape-dagre": {
+      "version": "2.5.0",
+      "resolved": "https://registry.npmjs.org/cytoscape-dagre/-/cytoscape-dagre-2.5.0.tgz",
+      "integrity": "sha512-VG2Knemmshop4kh5fpLO27rYcyUaaDkRw+6PiX4bstpB+QFt0p2oauMrsjVbUamGWQ6YNavh7x2em2uZlzV44g==",
+      "license": "MIT",
+      "dependencies": {
+        "dagre": "^0.8.5"
+      },
+      "peerDependencies": {
+        "cytoscape": "^3.2.22"
+      }
+    },
+    "node_modules/cytoscape-expand-collapse": {
+      "version": "4.1.1",
+      "resolved": "https://registry.npmjs.org/cytoscape-expand-collapse/-/cytoscape-expand-collapse-4.1.1.tgz",
+      "integrity": "sha512-MI4/GsA6Rf6RRzNR1aCitBLSnxiIKLxvZyCzF+oti/zn/ui1jmf769VcEFAEbjjsAtwteGsTmczI+niCMWJNvA==",
+      "peerDependencies": {
+        "cytoscape": "^3.3.0"
+      }
+    },
+    "node_modules/dagre": {
+      "version": "0.8.5",
+      "resolved": "https://registry.npmjs.org/dagre/-/dagre-0.8.5.tgz",
+      "integrity": "sha512-/aTqmnRta7x7MCCpExk7HQL2O4owCT2h8NT//9I1OQ9vt29Pa0BzSAkR5lwFUcQ7491yVi/3CXU9jQ5o0Mn2Sw==",
+      "license": "MIT",
+      "dependencies": {
+        "graphlib": "^2.1.8",
+        "lodash": "^4.17.15"
+      }
+    },
     "node_modules/de-indent": {
       "version": "1.0.2",
       "resolved": "https://registry.npmjs.org/de-indent/-/de-indent-1.0.2.tgz",
@@ -2252,6 +2313,15 @@
         "url": "https://github.com/sponsors/sindresorhus"
       }
     },
+    "node_modules/graphlib": {
+      "version": "2.1.8",
+      "resolved": "https://registry.npmjs.org/graphlib/-/graphlib-2.1.8.tgz",
+      "integrity": "sha512-jcLLfkpoVGmH7/InMC/1hIvOPSUh38oJtGhvrOFGzioE1DZ+0YW16RgmOJhHiuWTvGiJQ9Z1Ik43JvkRPRvE+A==",
+      "license": "MIT",
+      "dependencies": {
+        "lodash": "^4.17.15"
+      }
+    },
     "node_modules/has-flag": {
       "version": "4.0.0",
       "resolved": "https://registry.npmjs.org/has-flag/-/has-flag-4.0.0.tgz",
@@ -2401,8 +2471,7 @@
     "node_modules/lodash": {
       "version": "4.17.23",
       "resolved": "https://registry.npmjs.org/lodash/-/lodash-4.17.23.tgz",
-      "integrity": "sha512-LgVTMpQtIopCi79SJeDiP0TfWi5CNEc/L/aRdTh3yIvmZXTnheWpKjSZhnvMl8iXbC1tFg9gdHHDMLoV7CnG+w==",
+      "integrity": "sha512-LgVTMpQtIopCi79SJeDiP0TfWi5CNEc/L/aRdTh3yIvmZXTnheWpKjSZhnvMl8iXbC1tFg9gdHHDMLoV7CnG+w=="
-      "dev": true
     },
     "node_modules/lodash.merge": {
       "version": "4.6.2",

frontend/package.json

@@ -16,6 +16,9 @@
     "format:check": "prettier --check src/"
   },
   "dependencies": {
+    "cytoscape": "^3.30.0",
+    "cytoscape-dagre": "^2.5.0",
+    "cytoscape-expand-collapse": "^4.1.1",
     "dompurify": "^3.3.3",
     "marked": "^17.0.4",
     "pinia": "^2.3.0",
@@ -24,9 +27,11 @@
   },
   "devDependencies": {
     "@eslint/js": "^9.0.0",
-    "@vitest/mocker": "^4.1.2",
+    "@types/cytoscape": "^3.21.4",
+    "@types/cytoscape-dagre": "^2.3.3",
     "@types/dompurify": "^3.2.0",
     "@vitejs/plugin-vue": "^6.0.5",
+    "@vitest/mocker": "^4.1.2",
     "eslint": "^9.0.0",
     "eslint-plugin-vue": "^9.32.0",
     "prettier": "^3.4.0",

frontend/src/app/router/index.ts

@@ -27,6 +27,22 @@ const routes: RouteRecordRaw[] = [
     name: 'search',
     component: () => import('../../pages/SearchPage.vue'),
   },
+  {
+    // Reasoning-trace tunnel. Route is always registered; the page shows
+    // an empty state when the `reasoning` feature flag is off (same pattern
+    // as /search does for ingestion).
+    path: '/reasoning',
+    name: 'reasoning',
+    component: () => import('../../pages/ReasoningPage.vue'),
+  },
+  {
+    // Deep-link into a specific document's reasoning workspace, e.g. shared
+    // by Peter to a teammate.
+    path: '/reasoning/:docId',
+    name: 'reasoning-doc',
+    component: () => import('../../pages/ReasoningPage.vue'),
+    props: true,
+  },
   {
     path: '/settings',
     name: 'settings',

frontend/src/features/analysis/graphApi.ts

@@ -0,0 +1,59 @@
+import { apiFetch } from '../../shared/api/http'
+
+/**
+ * A single provenance entry for an element — matches Docling's
+ * `ProvenanceItem`. One element may have multiple (e.g. a paragraph that
+ * spans a page break has two entries, one per page).
+ */
+export interface GraphProvenance {
+  order: number
+  page_no: number | null
+  bbox_l: number
+  bbox_t: number
+  bbox_r: number
+  bbox_b: number
+  coord_origin: string
+  charspan_start: number | null
+  charspan_end: number | null
+}
+
+export interface GraphNode {
+  id: string
+  group: 'document' | 'element' | 'page' | 'chunk'
+  label?: string
+  docling_label?: string
+  self_ref?: string
+  text?: string
+  prov_page?: number | null
+  /** Full list of provenances (page + bbox) — used by the bbox-highlight UI. */
+  provs?: GraphProvenance[]
+  level?: number | null
+  page_no?: number
+  chunk_index?: number
+  title?: string
+  doc_id?: string
+  token_count?: number
+  [key: string]: unknown
+}
+
+export interface GraphEdge {
+  id: string
+  source: string
+  target: string
+  type: 'HAS_ROOT' | 'PARENT_OF' | 'NEXT' | 'ON_PAGE' | 'HAS_CHUNK' | 'DERIVED_FROM'
+  order?: number | null
+}
+
+export interface GraphPayload {
+  doc_id: string
+  nodes: GraphNode[]
+  edges: GraphEdge[]
+  node_count: number
+  edge_count: number
+  truncated: boolean
+  page_count: number
+}
+
+export function fetchDocumentGraph(docId: string): Promise<GraphPayload> {
+  return apiFetch<GraphPayload>(`/api/documents/${encodeURIComponent(docId)}/graph`)
+}

frontend/src/features/analysis/legendFilters.test.ts

@@ -0,0 +1,95 @@
+import { describe, expect, it } from 'vitest'
+
+import type { GraphNode } from './graphApi'
+import { LEGEND_CHIPS, findChip, partitionByHiddenChips } from './legendFilters'
+
+function node(overrides: Partial<GraphNode>): GraphNode {
+  return {
+    id: overrides.id ?? 'n1',
+    group: overrides.group ?? 'element',
+    ...overrides,
+  }
+}
+
+describe('LEGEND_CHIPS matching', () => {
+  it('section matches only SectionHeader elements', () => {
+    const chip = findChip('section')!
+    expect(chip.match(node({ group: 'element', label: 'SectionHeader' }))).toBe(true)
+    expect(chip.match(node({ group: 'element', label: 'Paragraph' }))).toBe(false)
+    // An element with no label should NOT match section.
+    expect(chip.match(node({ group: 'element' }))).toBe(false)
+  })
+
+  it('paragraph matches Paragraph and TextElement (fallback)', () => {
+    const chip = findChip('paragraph')!
+    expect(chip.match(node({ group: 'element', label: 'Paragraph' }))).toBe(true)
+    // TextElement is the generic fallback Docling-labels that don't have a
+    // specific mapping end up with. It's visually rendered in the paragraph
+    // color, so the chip covers both.
+    expect(chip.match(node({ group: 'element', label: 'TextElement' }))).toBe(true)
+    expect(chip.match(node({ group: 'element', label: 'Table' }))).toBe(false)
+  })
+
+  it('group-based chips use `group` not `label`', () => {
+    expect(findChip('document')!.match(node({ group: 'document' }))).toBe(true)
+    expect(findChip('page')!.match(node({ group: 'page', page_no: 1 }))).toBe(true)
+    expect(findChip('chunk')!.match(node({ group: 'chunk' }))).toBe(true)
+    // document chip must NOT match elements that happen to have label===undefined
+    expect(findChip('document')!.match(node({ group: 'element' }))).toBe(false)
+  })
+
+  it('exposes a stable list of chips', () => {
+    // The sidebar CSS targets `.legend-${key}` — renaming a key here would
+    // silently break the styles. Pin the expected order + keys.
+    expect(LEGEND_CHIPS.map((c) => c.key)).toEqual([
+      'document',
+      'section',
+      'paragraph',
+      'table',
+      'figure',
+      'page',
+      'chunk',
+    ])
+  })
+})
+
+describe('partitionByHiddenChips', () => {
+  const nodes: GraphNode[] = [
+    node({ id: 'd', group: 'document' }),
+    node({ id: 's1', group: 'element', label: 'SectionHeader' }),
+    node({ id: 'p1', group: 'element', label: 'Paragraph' }),
+    node({ id: 'f1', group: 'element', label: 'Figure' }),
+    node({ id: 'pg1', group: 'page' }),
+  ]
+
+  it('returns everything in `show` when the hidden set is empty', () => {
+    const { hide, show } = partitionByHiddenChips(nodes, new Set())
+    expect(hide).toHaveLength(0)
+    expect(show).toHaveLength(nodes.length)
+  })
+
+  it('hides the nodes matching the selected chips', () => {
+    const { hide, show } = partitionByHiddenChips(nodes, new Set(['figure', 'document']))
+    expect(hide.map((n) => n.id).sort()).toEqual(['d', 'f1'])
+    expect(show.map((n) => n.id).sort()).toEqual(['p1', 'pg1', 's1'])
+  })
+
+  it('ignores unknown chip keys (no silent breakage)', () => {
+    const { hide, show } = partitionByHiddenChips(nodes, new Set(['some-new-kind']))
+    expect(hide).toHaveLength(0)
+    expect(show).toHaveLength(nodes.length)
+  })
+
+  it('keeps un-classifiable nodes visible even if all chips are off', () => {
+    // A hypothetical element with an unmapped label (e.g. :Formula is in the
+    // writer but has no chip yet) must not be hidden when toggling unrelated
+    // chips — only explicit chip matches cause hiding.
+    const weird = node({ id: 'w', group: 'element', label: 'Formula' })
+    const { hide, show } = partitionByHiddenChips(
+      [weird, ...nodes],
+      new Set(['section', 'paragraph']),
+    )
+    expect(hide.map((n) => n.id)).toEqual(['s1', 'p1'])
+    expect(show.some((n) => n.id === 'w')).toBe(true)
+  })
+})

frontend/src/features/analysis/legendFilters.ts

@@ -0,0 +1,108 @@
+/**
+ * Declarative definition of GraphView's legend chips.
+ *
+ * Each chip maps to (a) a swatch color shown in the toolbar, and (b) a
+ * predicate that decides whether a given node is "of that kind". Clicking a
+ * chip toggles the matching nodes' visibility via the `.hidden` Cytoscape
+ * class — pure function kept here so it's unit-testable without mounting
+ * the component or spinning up Cytoscape.
+ */
+
+import type { GraphNode } from './graphApi'
+
+export interface LegendChip {
+  /** Stable ID used as the key in the `hiddenChips` Set + as a CSS slug. */
+  key: string
+  /** Display label shown in the chip. Kept English for now — matches
+   * existing `legend-*` class names that the design stylesheet targets. */
+  label: string
+  /** Swatch color (also used for the CSS class `legend-${key}`). */
+  color: string
+  /** Returns true when a node belongs to this chip's category. */
+  match: (node: GraphNode) => boolean
+}
+
+/**
+ * Legend order matches the chips currently rendered in GraphView's toolbar,
+ * so swapping to this source of truth is a drop-in replacement.
+ *
+ * `kindLabel` is the specific Neo4j label returned by `fetch_graph`
+ * (e.g. `SectionHeader`, `Paragraph`, `Table`, …). Group-based chips target
+ * the `group` attribute (document / page / chunk) — only `element` nodes
+ * have a useful `kindLabel`.
+ */
+export const LEGEND_CHIPS: LegendChip[] = [
+  {
+    key: 'document',
+    label: 'Document',
+    color: '#1E293B',
+    match: (n) => n.group === 'document',
+  },
+  {
+    key: 'section',
+    label: 'Section',
+    color: '#F97316',
+    match: (n) => n.group === 'element' && n.label === 'SectionHeader',
+  },
+  {
+    key: 'paragraph',
+    label: 'Paragraph',
+    color: '#3B82F6',
+    match: (n) => n.group === 'element' && (n.label === 'Paragraph' || n.label === 'TextElement'),
+  },
+  {
+    key: 'table',
+    label: 'Table',
+    color: '#8B5CF6',
+    match: (n) => n.group === 'element' && n.label === 'Table',
+  },
+  {
+    key: 'figure',
+    label: 'Figure',
+    color: '#22C55E',
+    match: (n) => n.group === 'element' && n.label === 'Figure',
+  },
+  {
+    key: 'page',
+    label: 'Page',
+    color: '#94A3B8',
+    match: (n) => n.group === 'page',
+  },
+  {
+    key: 'chunk',
+    label: 'Chunk',
+    color: '#DC2626',
+    match: (n) => n.group === 'chunk',
+  },
+]
+
+/** Lookup by key — tiny helper to keep GraphView concise. */
+export function findChip(key: string): LegendChip | undefined {
+  return LEGEND_CHIPS.find((c) => c.key === key)
+}
+
+/**
+ * Partition an iterable of nodes into `{ hide, show }` based on a set of
+ * chip keys to hide. A node is hidden if it matches any chip in the set.
+ * Nodes matching no chip at all (e.g. a future new kind we haven't added
+ * to the legend) are always shown — no silent disappearance.
+ */
+export function partitionByHiddenChips(
+  nodes: Iterable<GraphNode>,
+  hiddenChipKeys: ReadonlySet<string>,
+): { hide: GraphNode[]; show: GraphNode[] } {
+  const hide: GraphNode[] = []
+  const show: GraphNode[] = []
+  for (const n of nodes) {
+    let matchedHidden = false
+    for (const key of hiddenChipKeys) {
+      const chip = findChip(key)
+      if (chip && chip.match(n)) {
+        matchedHidden = true
+        break
+      }
+    }
+    ;(matchedHidden ? hide : show).push(n)
+  }
+  return { hide, show }
+}

frontend/src/features/analysis/sectionParenting.test.ts

@@ -0,0 +1,136 @@
+import { describe, expect, it } from 'vitest'
+
+import type { GraphEdge, GraphNode } from './graphApi'
+import { computeSectionParents, explicitParentMap, mergeParentMaps } from './sectionParenting'
+
+function el(selfRef: string, label: string): GraphNode {
+  return {
+    id: `elem::${selfRef}`,
+    group: 'element',
+    label,
+    self_ref: selfRef,
+  }
+}
+
+function nextEdge(from: string, to: string): GraphEdge {
+  return {
+    id: `NEXT::elem::${from}::elem::${to}`,
+    source: `elem::${from}`,
+    target: `elem::${to}`,
+    type: 'NEXT',
+  }
+}
+
+function parentEdge(parent: string, child: string): GraphEdge {
+  return {
+    id: `PARENT_OF::elem::${parent}::elem::${child}`,
+    source: `elem::${parent}`,
+    target: `elem::${child}`,
+    type: 'PARENT_OF',
+  }
+}
+
+describe('computeSectionParents', () => {
+  it('returns empty when no SectionHeader exists', () => {
+    const nodes = [el('#/texts/0', 'Paragraph'), el('#/texts/1', 'Paragraph')]
+    const edges = [nextEdge('#/texts/0', '#/texts/1')]
+    expect(computeSectionParents(nodes, edges)).toEqual(new Map())
+  })
+
+  it('parents every element to the preceding SectionHeader', () => {
+    // SH ─▶ P ─▶ P ─▶ SH ─▶ P
+    const nodes = [
+      el('#/texts/0', 'SectionHeader'),
+      el('#/texts/1', 'Paragraph'),
+      el('#/texts/2', 'Paragraph'),
+      el('#/texts/3', 'SectionHeader'),
+      el('#/texts/4', 'Paragraph'),
+    ]
+    const edges = [
+      nextEdge('#/texts/0', '#/texts/1'),
+      nextEdge('#/texts/1', '#/texts/2'),
+      nextEdge('#/texts/2', '#/texts/3'),
+      nextEdge('#/texts/3', '#/texts/4'),
+    ]
+    const parents = computeSectionParents(nodes, edges)
+    expect(parents.get('elem::#/texts/1')).toBe('elem::#/texts/0')
+    expect(parents.get('elem::#/texts/2')).toBe('elem::#/texts/0')
+    expect(parents.get('elem::#/texts/4')).toBe('elem::#/texts/3')
+    // Headers are never themselves synthetically parented.
+    expect(parents.has('elem::#/texts/0')).toBe(false)
+    expect(parents.has('elem::#/texts/3')).toBe(false)
+  })
+
+  it('does not override an explicit PARENT_OF', () => {
+    // SH ─▶ List ─▶ ListItem, with PARENT_OF(List, ListItem)
+    const nodes = [
+      el('#/texts/0', 'SectionHeader'),
+      el('#/groups/0', 'List'),
+      el('#/texts/1', 'ListItem'),
+    ]
+    const edges: GraphEdge[] = [
+      nextEdge('#/texts/0', '#/groups/0'),
+      nextEdge('#/groups/0', '#/texts/1'),
+      parentEdge('#/groups/0', '#/texts/1'),
+    ]
+    const parents = computeSectionParents(nodes, edges)
+    // The List falls under the section…
+    expect(parents.get('elem::#/groups/0')).toBe('elem::#/texts/0')
+    // …but the ListItem keeps its explicit List parent, not the section.
+    expect(parents.has('elem::#/texts/1')).toBe(false)
+  })
+
+  it('leaves elements preceding the first SectionHeader unparented', () => {
+    // Page-header stuff can come before any SectionHeader — don't force-parent it.
+    const nodes = [
+      el('#/texts/0', 'PageHeader'),
+      el('#/texts/1', 'SectionHeader'),
+      el('#/texts/2', 'Paragraph'),
+    ]
+    const edges = [nextEdge('#/texts/0', '#/texts/1'), nextEdge('#/texts/1', '#/texts/2')]
+    const parents = computeSectionParents(nodes, edges)
+    expect(parents.has('elem::#/texts/0')).toBe(false)
+    expect(parents.get('elem::#/texts/2')).toBe('elem::#/texts/1')
+  })
+
+  it('handles multiple disconnected NEXT chains deterministically', () => {
+    // Two chains, each with its own section. Heads are sorted by id so the
+    // walk order is stable run-to-run.
+    const nodes = [
+      el('#/texts/10', 'SectionHeader'),
+      el('#/texts/11', 'Paragraph'),
+      el('#/texts/20', 'SectionHeader'),
+      el('#/texts/21', 'Paragraph'),
+    ]
+    const edges = [nextEdge('#/texts/10', '#/texts/11'), nextEdge('#/texts/20', '#/texts/21')]
+    const parents = computeSectionParents(nodes, edges)
+    expect(parents.get('elem::#/texts/11')).toBe('elem::#/texts/10')
+    expect(parents.get('elem::#/texts/21')).toBe('elem::#/texts/20')
+  })
+})
+
+describe('explicitParentMap + mergeParentMaps', () => {
+  it('extracts explicit PARENT_OF into a child→parent map', () => {
+    const edges = [parentEdge('#/groups/0', '#/texts/1')]
+    const m = explicitParentMap(edges)
+    expect(m.get('elem::#/texts/1')).toBe('elem::#/groups/0')
+  })
+
+  it('explicit wins over synthetic when both map the same child', () => {
+    const synthetic = new Map([['c', 'synthetic-parent']])
+    const explicit = new Map([['c', 'real-parent']])
+    const merged = mergeParentMaps(explicit, synthetic)
+    expect(merged.get('c')).toBe('real-parent')
+  })
+
+  it('preserves synthetic entries that have no explicit counterpart', () => {
+    const synthetic = new Map([
+      ['c1', 'section-1'],
+      ['c2', 'section-1'],
+    ])
+    const explicit = new Map([['c2', 'list-a']])
+    const merged = mergeParentMaps(explicit, synthetic)
+    expect(merged.get('c1')).toBe('section-1')
+    expect(merged.get('c2')).toBe('list-a')
+  })
+})

frontend/src/features/analysis/sectionParenting.ts

@@ -0,0 +1,111 @@
+/**
+ * Synthesize "section scope" compound-node parents for GraphView.
+ *
+ * The Docling graph is structurally flat — most elements are direct children
+ * of `#/body` (HAS_ROOT) with no explicit parent relationship to the section
+ * header that "owns" them in the reader's mental model. To enable the
+ * expand-collapse UX per section, we derive that ownership here, at render
+ * time, from the NEXT reading-order chain:
+ *
+ *   - Walk NEXT from the head of the chain.
+ *   - Every time we hit a SectionHeader, it becomes the "current section".
+ *   - Every non-SectionHeader node thereafter — unless it has an explicit
+ *     PARENT_OF edge (e.g. list_item → list) — gets that section as its
+ *     compound parent.
+ *
+ * The function returns a `Map<childId, parentId>` where both ids are full
+ * Cytoscape node ids (`elem::<self_ref>`). The caller merges this with any
+ * existing PARENT_OF to produce the final `data.parent` on each node.
+ *
+ * Edge cases handled:
+ *   - No SectionHeader in the doc → returns an empty map (nothing to collapse).
+ *   - Multiple disconnected NEXT chains → each is walked independently.
+ *   - Explicit PARENT_OF on a child → never overridden.
+ *   - SectionHeader that is itself a list_item or nested → its descendants
+ *     still use it as their section anchor; the SectionHeader keeps its own
+ *     explicit parent.
+ */
+
+import type { GraphEdge, GraphNode } from './graphApi'
+
+export function computeSectionParents(
+  nodes: readonly GraphNode[],
+  edges: readonly GraphEdge[],
+): Map<string, string> {
+  const elementIds = new Set<string>()
+  const kindById = new Map<string, string | undefined>()
+  for (const n of nodes) {
+    if (n.group !== 'element') continue
+    elementIds.add(n.id)
+    kindById.set(n.id, n.label)
+  }
+  if (elementIds.size === 0) return new Map()
+
+  const explicitParentOf = new Set<string>()
+  const nextMap = new Map<string, string>()
+  const hasIncomingNext = new Set<string>()
+
+  for (const e of edges) {
+    if (!elementIds.has(e.source) || !elementIds.has(e.target)) continue
+    if (e.type === 'PARENT_OF') {
+      explicitParentOf.add(e.target)
+    } else if (e.type === 'NEXT') {
+      // Preserve first-seen NEXT edge per source (should be unique anyway).
+      if (!nextMap.has(e.source)) nextMap.set(e.source, e.target)
+      hasIncomingNext.add(e.target)
+    }
+  }
+
+  // Walk heads deterministically (sorted) so the same graph always produces
+  // the same parenting map — useful for tests and reproducible debugging.
+  const heads = [...elementIds].filter((id) => !hasIncomingNext.has(id)).sort()
+
+  const parents = new Map<string, string>()
+  const visited = new Set<string>()
+
+  for (const head of heads) {
+    let currentSection: string | null = null
+    let node: string | undefined = head
+
+    while (node && !visited.has(node)) {
+      visited.add(node)
+      const kind = kindById.get(node)
+
+      if (kind === 'SectionHeader') {
+        currentSection = node
+        // The SectionHeader itself does NOT get a synthetic parent — it is
+        // the anchor. If it has an explicit PARENT_OF, that one remains
+        // authoritative downstream (the merger respects it).
+      } else if (currentSection && !explicitParentOf.has(node)) {
+        parents.set(node, currentSection)
+      }
+
+      node = nextMap.get(node)
+    }
+  }
+
+  return parents
+}
+
+/**
+ * Merge an explicit PARENT_OF map with synthetic section parents. Explicit
+ * wins. Produces the final `childId → parentId` map that callers set as
+ * `data.parent` on Cytoscape nodes.
+ */
+export function mergeParentMaps(
+  explicit: ReadonlyMap<string, string>,
+  synthetic: ReadonlyMap<string, string>,
+): Map<string, string> {
+  const out = new Map(synthetic)
+  for (const [child, parent] of explicit) out.set(child, parent)
+  return out
+}
+
+/** Extract the explicit PARENT_OF map from the edge list — convenience. */
+export function explicitParentMap(edges: readonly GraphEdge[]): Map<string, string> {
+  const out = new Map<string, string>()
+  for (const e of edges) {
+    if (e.type === 'PARENT_OF') out.set(e.target, e.source)
+  }
+  return out
+}

frontend/src/features/analysis/ui/GraphView.vue

@@ -0,0 +1,695 @@
+<template>
+  <div class="graph-view" data-e2e="graph-view">
+    <div v-if="loading" class="graph-placeholder">
+      <div class="spinner-large" />
+      <span>{{ t('results.graphLoading') }}</span>
+    </div>
+    <div v-else-if="error" class="graph-placeholder error" data-e2e="graph-error">
+      <span>{{ error }}</span>
+      <button class="retry-btn" @click="load">{{ t('results.retry') }}</button>
+    </div>
+    <div v-else-if="empty" class="graph-placeholder">
+      <span>{{ t('results.graphEmpty') }}</span>
+    </div>
+    <template v-else>
+      <div class="graph-toolbar">
+        <span class="graph-stats">
+          {{ payload?.node_count }} nodes · {{ payload?.edge_count }} edges ·
+          {{ payload?.page_count }} pages
+        </span>
+        <span class="graph-legend">
+          <button
+            v-for="chip in visibleChips"
+            :key="chip.key"
+            type="button"
+            class="legend-chip"
+            :class="[`legend-${chip.key}`, { 'legend-off': hiddenChips.has(chip.key) }]"
+            :data-e2e="`legend-${chip.key}`"
+            :title="
+              hiddenChips.has(chip.key) ? `Show ${chip.label} nodes` : `Hide ${chip.label} nodes`
+            "
+            :aria-pressed="!hiddenChips.has(chip.key)"
+            @click="toggleChip(chip.key)"
+          >
+            {{ chip.label }}
+          </button>
+        </span>
+      </div>
+      <div class="graph-body">
+        <div class="graph-canvas-wrap">
+          <div ref="containerRef" class="graph-canvas" data-e2e="graph-canvas" />
+          <!-- Hover tooltip, positioned inside the canvas wrap so its
+               coords match cy.renderedPosition() exactly. -->
+          <div
+            v-if="tooltip"
+            class="graph-tooltip"
+            data-e2e="graph-tooltip"
+            :style="{ left: tooltip.x + 'px', top: tooltip.y + 'px' }"
+          >
+            {{ tooltip.text }}
+          </div>
+        </div>
+        <NodeDetailsPanel
+          :node="selectedNode"
+          :contents="selectedNodeContents"
+          @close="closeDetails"
+          @navigate="navigateToNode"
+        />
+      </div>
+    </template>
+  </div>
+</template>
+
+<script setup lang="ts">
+import type { Core } from 'cytoscape'
+import { computed, onMounted, onBeforeUnmount, ref, watch, nextTick } from 'vue'
+import { useI18n } from '../../../shared/i18n'
+import { reasoningOverlayStyles } from '../../reasoning/graphReasoningOverlay'
+import { fetchDocumentGraph, type GraphNode, type GraphPayload } from '../graphApi'
+import { LEGEND_CHIPS, findChip } from '../legendFilters'
+import { computeSectionParents, explicitParentMap, mergeParentMaps } from '../sectionParenting'
+import NodeDetailsPanel from './NodeDetailsPanel.vue'
+
+// `fetcher` is optional so Maintain can keep using the Neo4j-backed endpoint
+// (`fetchDocumentGraph`, the default) while the reasoning-trace page can
+// inject a SQLite-backed fetcher that returns the same `GraphPayload` shape
+// without requiring Neo4j. Keeping this at the component boundary means the
+// rendering pipeline below doesn't care where the graph came from.
+const props = withDefaults(
+  defineProps<{
+    docId: string | null
+    fetcher?: (docId: string) => Promise<GraphPayload>
+  }>(),
+  { fetcher: fetchDocumentGraph },
+)
+const emit = defineEmits<{
+  /** Emitted on node tap with the element's `self_ref` (null when the tap
+   * cleared the selection, or when the tapped node has no self_ref —
+   * Document / Page / Chunk). Consumers can mirror the selection elsewhere
+   * (e.g. the ReasoningWorkspace syncs it to the PDF viewer). */
+  nodeFocus: [selfRef: string | null]
+}>()
+const { t } = useI18n()
+
+const containerRef = ref<HTMLDivElement | null>(null)
+const payload = ref<GraphPayload | null>(null)
+const loading = ref(false)
+const error = ref<string | null>(null)
+const empty = ref(false)
+// Exposed via defineExpose so parent components (e.g. the reasoning trace
+// overlay) can read the live Cytoscape instance reactively. Null while the
+// graph is loading / empty / unmounted.
+const cy = ref<Core | null>(null)
+
+// Legend-driven visibility: user clicks a chip → its key lands here and the
+// matching nodes get the `.hidden` Cytoscape class. Reset to empty every time
+// the doc changes (see the `watch(() => props.docId)` below) — kept as a Set
+// because order doesn't matter and membership checks are the hot path.
+const hiddenChips = ref<Set<string>>(new Set())
+
+// Click → details panel. Null = panel hidden.
+const selectedNode = ref<GraphNode | null>(null)
+
+// Compound parenting map (childId → parentId), kept in sync with the
+// Cytoscape render so the details panel can show "this section contains …".
+// Updated at the end of `renderGraph` — before that it's empty.
+const parentMap = ref<Map<string, string>>(new Map())
+
+// Inverse index of parentMap: parentId → childId[]. Enables the
+// NodeDetailsPanel "contents" section (click a section → see what's in it).
+const childrenByParent = computed<Map<string, GraphNode[]>>(() => {
+  const out = new Map<string, GraphNode[]>()
+  const byId = new Map<string, GraphNode>()
+  for (const n of payload.value?.nodes ?? []) byId.set(n.id, n)
+  for (const [childId, parentId] of parentMap.value) {
+    const child = byId.get(childId)
+    if (!child) continue
+    if (!out.has(parentId)) out.set(parentId, [])
+    out.get(parentId)!.push(child)
+  }
+  return out
+})
+
+const selectedNodeContents = computed<GraphNode[]>(() => {
+  const id = selectedNode.value?.id
+  if (!id) return []
+  return childrenByParent.value.get(id) ?? []
+})
+
+// Only surface chips that actually have matching nodes in the current
+// payload. Keeps the legend in sync with the source (e.g. Reasoning view
+// never emits Chunk nodes, so the Chunk chip would dangle) without
+// hardcoding per-view chip lists.
+const visibleChips = computed(() => {
+  const nodes = payload.value?.nodes ?? []
+  if (nodes.length === 0) return []
+  return LEGEND_CHIPS.filter((chip) => nodes.some((n) => chip.match(n)))
+})
+
+// Hover tooltip: position (px within .graph-canvas) + text. Null hides it.
+const tooltip = ref<{ x: number; y: number; text: string } | null>(null)
+
+const NODE_COLORS: Record<string, string> = {
+  document: '#1E293B',
+  SectionHeader: '#F97316',
+  Paragraph: '#3B82F6',
+  TextElement: '#3B82F6',
+  Table: '#8B5CF6',
+  Figure: '#22C55E',
+  ListItem: '#06B6D4',
+  Formula: '#EC4899',
+  Code: '#14B8A6',
+  Caption: '#EAB308',
+  Page: '#94A3B8',
+  Chunk: '#DC2626',
+}
+
+function nodeColor(n: GraphPayload['nodes'][number]): string {
+  if (n.group === 'document') return NODE_COLORS.document
+  if (n.group === 'page') return NODE_COLORS.Page
+  if (n.group === 'chunk') return NODE_COLORS.Chunk
+  return NODE_COLORS[n.label || 'TextElement'] || NODE_COLORS.TextElement
+}
+
+function nodeLabel(n: GraphPayload['nodes'][number]): string {
+  if (n.group === 'document') return n.title || n.id
+  if (n.group === 'page') return `p.${n.page_no}`
+  if (n.group === 'chunk') return `chunk #${n.chunk_index}`
+  const txt = (n.text || '').slice(0, 40)
+  return txt || n.label || n.docling_label || n.self_ref || n.id
+}
+
+async function load(): Promise<void> {
+  if (!props.docId) {
+    empty.value = true
+    return
+  }
+  loading.value = true
+  error.value = null
+  empty.value = false
+  try {
+    payload.value = await props.fetcher(props.docId)
+    if (!payload.value.nodes.length) {
+      empty.value = true
+      return
+    }
+    // Flip loading off so the canvas <div> mounts, then wait a tick before init.
+    loading.value = false
+    await nextTick()
+    await renderGraph()
+    // Re-apply any chips the user had toggled before this load (e.g. they
+    // hid Pages on doc A, then navigated to doc B — keep Pages hidden).
+    applyHiddenChips()
+  } catch (e) {
+    error.value = (e as Error).message || 'Failed to load graph'
+    console.error('Failed to load graph', e)
+  } finally {
+    loading.value = false
+  }
+}
+
+async function renderGraph(): Promise<void> {
+  if (!containerRef.value || !payload.value) return
+  // Dynamic imports keep the heavy Cytoscape bundle out of the main chunk.
+  const [{ default: cytoscape }, { default: dagre }, ecMod] = await Promise.all([
+    import('cytoscape'),
+    import('cytoscape-dagre'),
+    import('cytoscape-expand-collapse'),
+  ])
+
+  const C = cytoscape as any
+  C.use(dagre)
+  // Plugin registration is idempotent — calling use() twice is a no-op.
+  C.use((ecMod as any).default ?? ecMod)
+
+  if (cy.value) {
+    cy.value.destroy()
+    cy.value = null
+  }
+
+  // Compute compound parenting: merge docling-native PARENT_OF with the
+  // synthetic section-scope parents so every non-root element sits inside
+  // its section visually (enables per-section collapse via the legend chips
+  // and the expand-collapse plugin). Also persisted on `parentMap` so the
+  // NodeDetailsPanel can list what a given section contains.
+  const computedParentMap = mergeParentMaps(
+    explicitParentMap(payload.value.edges),
+    computeSectionParents(payload.value.nodes, payload.value.edges),
+  )
+  parentMap.value = computedParentMap
+
+  const elements = [
+    ...payload.value.nodes.map((n) => ({
+      data: {
+        id: n.id,
+        label: nodeLabel(n),
+        // `kindLabel` is the specific Neo4j label (SectionHeader, Paragraph,
+        // Figure, …) — kept as a data attribute so legend filters can match
+        // on it. `label` above is the human display string for Cytoscape.
+        kindLabel: n.label ?? null,
+        bg: nodeColor(n),
+        group: n.group,
+        // Compound-node parent: used by the expand-collapse plugin to
+        // fold/unfold a section's scope. `undefined` = this node is a root
+        // of the compound hierarchy (Documents, unparented sections, etc.).
+        parent: computedParentMap.get(n.id),
+        raw: n,
+      },
+    })),
+    ...payload.value.edges.map((e) => ({
+      data: {
+        id: e.id,
+        source: e.source,
+        target: e.target,
+        type: e.type,
+      },
+    })),
+  ]
+
+  cy.value = cytoscape({
+    container: containerRef.value,
+    elements,
+    style: [
+      {
+        selector: 'node',
+        style: {
+          'background-color': 'data(bg)',
+          label: 'data(label)',
+          color: '#0F172A',
+          'font-size': 10,
+          'text-wrap': 'ellipsis',
+          'text-max-width': '140px',
+          'text-valign': 'center',
+          'text-halign': 'center',
+          width: 28,
+          height: 28,
+          'border-width': 1,
+          'border-color': '#0F172A',
+        },
+      },
+      {
+        selector: 'node[group = "document"]',
+        style: { shape: 'round-rectangle', width: 60, height: 36, color: '#F8FAFC' },
+      },
+      {
+        selector: 'node[group = "page"]',
+        style: { shape: 'round-rectangle', width: 40, height: 24 },
+      },
+      {
+        selector: 'node[group = "chunk"]',
+        style: { shape: 'diamond', color: '#F8FAFC' },
+      },
+      {
+        selector: 'edge',
+        style: {
+          width: 1,
+          'line-color': '#94A3B8',
+          'target-arrow-color': '#94A3B8',
+          'target-arrow-shape': 'triangle',
+          'curve-style': 'bezier',
+          'font-size': 8,
+          color: '#64748B',
+        },
+      },
+      {
+        selector: 'edge[type = "PARENT_OF"]',
+        style: { 'line-color': '#1E293B', 'target-arrow-color': '#1E293B', width: 1.5 },
+      },
+      {
+        selector: 'edge[type = "NEXT"]',
+        style: { 'line-style': 'dashed', 'line-color': '#64748B' },
+      },
+      {
+        selector: 'edge[type = "ON_PAGE"]',
+        style: { 'line-color': '#CBD5E1', width: 1 },
+      },
+      {
+        selector: 'edge[type = "DERIVED_FROM"]',
+        style: { 'line-color': '#DC2626', 'target-arrow-color': '#DC2626' },
+      },
+      // Reasoning-trace overlay: visited-node class + synthetic REASONING_NEXT edges.
+      ...reasoningOverlayStyles,
+      // Legend-driven visibility: chips toggle the `.hidden` class on
+      // matching nodes. `display: none` cascades to connected edges for free.
+      { selector: 'node.hidden', style: { display: 'none' } },
+      // Compound nodes (section scope + explicit PARENT_OF containers). A
+      // node with :parent children is rendered as a bounding box here —
+      // visually groups the section's content together. Minimal padding so
+      // the layout stays compact.
+      {
+        selector: ':parent',
+        style: {
+          'background-opacity': 0.08,
+          'background-color': '#F97316',
+          'border-color': '#FDBA74',
+          'border-width': 1,
+          'padding-top': '12px',
+          'padding-bottom': '12px',
+          'padding-left': '12px',
+          'padding-right': '12px',
+          'text-valign': 'top',
+          'text-halign': 'center',
+          'text-margin-y': -4,
+          shape: 'round-rectangle',
+        },
+      },
+      // Click-selected node: stronger border so the user sees which one
+      // populated the details panel.
+      {
+        selector: 'node.nd-selected',
+        style: {
+          'border-width': 4,
+          'border-color': '#0EA5E9',
+          'overlay-color': '#0EA5E9',
+          'overlay-opacity': 0.12,
+          'overlay-padding': 4,
+          'z-index': 60,
+        },
+      },
+    ],
+
+    layout: {
+      name: 'dagre',
+      rankDir: 'TB',
+      nodeSep: 30,
+      edgeSep: 10,
+      rankSep: 40,
+    } as any,
+    wheelSensitivity: 0.15,
+  })
+
+  // --- Plugin activation ---------------------------------------------------
+  // Expand/collapse on compound nodes. `layoutBy` re-runs dagre after each
+  // toggle so the graph stays tidy. Don't animate — on big docs the per-node
+  // tween is choppy and the user just wants the end state.
+  ;(cy.value as any).expandCollapse({
+    layoutBy: { name: 'dagre', rankDir: 'TB', nodeSep: 30, rankSep: 40 },
+    fisheye: false,
+    animate: false,
+    undoable: false,
+    cueEnabled: true, // shows the +/- cue on compound nodes
+    expandCollapseCuePosition: 'top-left',
+    expandCollapseCueSize: 12,
+  })
+
+  // --- Interactions --------------------------------------------------------
+  cy.value.on('tap', 'node', (evt) => {
+    const raw = evt.target.data('raw') as GraphNode | undefined
+    if (!raw) return
+    selectedNode.value = raw
+    // Visual feedback — clear previous selection class first.
+    cy.value?.nodes('.nd-selected').removeClass('nd-selected')
+    evt.target.addClass('nd-selected')
+    // Let the outer workspace mirror the selection (e.g. into the PDF view).
+    // Nodes without a `self_ref` (Document / Page / Chunk) emit `null` so
+    // the consumer can reset its focus.
+    emit('nodeFocus', raw.self_ref ?? null)
+  })
+  // Click on background → close details panel + clear cross-view focus.
+  cy.value.on('tap', (evt) => {
+    if (evt.target === cy.value) {
+      selectedNode.value = null
+      cy.value?.nodes('.nd-selected').removeClass('nd-selected')
+      emit('nodeFocus', null)
+    }
+  })
+
+  // Hover tooltip — shows the full node text (backend truncates to 200 chars).
+  cy.value.on('mouseover', 'node', (evt) => {
+    const raw = evt.target.data('raw') as GraphNode | undefined
+    if (!raw) return
+    const text = tooltipTextFor(raw)
+    if (!text) return
+    const pos = evt.target.renderedPosition()
+    tooltip.value = { x: pos.x, y: pos.y, text }
+  })
+  cy.value.on('mouseout', 'node', () => {
+    tooltip.value = null
+  })
+  cy.value.on('pan zoom', () => {
+    // Tooltip coordinates are in rendered space; on pan/zoom they're stale.
+    tooltip.value = null
+  })
+}
+
+/**
+ * What to show on hover: prefer the node's full text, then title, then ref.
+ * Keeps the tooltip useful across node kinds (Document/Page/Chunk too).
+ */
+function tooltipTextFor(n: GraphNode): string {
+  if (n.group === 'document') return (n.title as string | undefined) ?? n.id
+  if (n.group === 'page') return `Page ${n.page_no ?? '?'}`
+  if (n.group === 'chunk') {
+    const head = ((n.text as string | undefined) ?? '').slice(0, 160)
+    return head ? `chunk #${n.chunk_index ?? '?'}\n${head}` : `chunk #${n.chunk_index ?? '?'}`
+  }
+  const text = (n.text as string | undefined) ?? ''
+  const ref = n.self_ref ?? ''
+  if (text) return text
+  return ref || n.label || ''
+}
+
+function closeDetails(): void {
+  selectedNode.value = null
+  cy.value?.nodes('.nd-selected').removeClass('nd-selected')
+}
+
+/**
+ * Triggered when the user clicks a child row inside the NodeDetailsPanel
+ * (e.g. the "Contents" list of a section). Switch the selection, center the
+ * viewport on the target, and flash the node briefly so the eye can catch it.
+ */
+function navigateToNode(target: GraphNode): void {
+  selectedNode.value = target
+  if (!cy.value) return
+  cy.value.nodes('.nd-selected').removeClass('nd-selected')
+  const el = cy.value.getElementById(target.id)
+  if (el && el.length > 0) {
+    el.addClass('nd-selected')
+    cy.value.animate({ center: { eles: el }, duration: 250 })
+  }
+}
+
+/**
+ * Mirror an external selection (e.g. user clicked a bbox in the PDF view)
+ * onto the graph: select the matching node, scroll it into view, update
+ * the details panel. No-op if the element isn't in the current graph
+ * (common for a PDF-only element that the reasoning graph didn't emit).
+ */
+function selectBySelfRef(selfRef: string): void {
+  const node = payload.value?.nodes.find((n) => n.self_ref === selfRef) ?? null
+  if (!node) return
+  navigateToNode(node)
+}
+
+function disposeGraph(): void {
+  if (cy.value) {
+    cy.value.destroy()
+    cy.value = null
+  }
+  selectedNode.value = null
+  tooltip.value = null
+}
+
+/**
+ * Apply the current `hiddenChips` set to the Cytoscape instance — marks
+ * every node whose chip is in the set with the `.hidden` class, and clears
+ * the class from nodes whose chip is no longer hidden.
+ *
+ * Called after every re-render (so chip state survives a doc reload) and
+ * whenever the user toggles a chip.
+ */
+function applyHiddenChips(): void {
+  const c = cy.value
+  if (!c) return
+
+  c.nodes().forEach((n: any) => {
+    const raw = n.data('raw')
+    if (!raw) return
+    const hiddenByChip = [...hiddenChips.value].some((key) => {
+      const chip = findChip(key)
+      return chip?.match(raw) ?? false
+    })
+    if (hiddenByChip) n.addClass('hidden')
+    else n.removeClass('hidden')
+  })
+}
+
+function toggleChip(key: string): void {
+  const next = new Set(hiddenChips.value)
+  if (next.has(key)) next.delete(key)
+  else next.add(key)
+  hiddenChips.value = next
+  applyHiddenChips()
+}
+
+onMounted(load)
+onBeforeUnmount(disposeGraph)
+watch(
+  () => props.docId,
+  () => {
+    disposeGraph()
+    load()
+  },
+)
+
+// Let parent components observe the live Cytoscape instance (e.g. the
+// reasoning-trace overlay reads it via `graphViewRef.value?.cy`).
+defineExpose({ cy, load, selectBySelfRef })
+</script>
+
+<style scoped>
+.graph-view {
+  display: flex;
+  flex-direction: column;
+  height: 100%;
+  overflow: hidden;
+}
+
+.graph-toolbar {
+  display: flex;
+  justify-content: space-between;
+  align-items: center;
+  padding: 8px 12px;
+  border-bottom: 1px solid var(--border);
+  gap: 12px;
+  flex-wrap: wrap;
+}
+
+.graph-stats {
+  font-family: 'IBM Plex Mono', monospace;
+  font-size: 11px;
+  color: var(--text-muted);
+}
+
+.graph-legend {
+  display: flex;
+  gap: 6px;
+  flex-wrap: wrap;
+}
+
+.legend-chip {
+  font-size: 10px;
+  font-weight: 600;
+  padding: 2px 8px;
+  border-radius: 10px;
+  color: #f8fafc;
+  border: 0;
+  cursor: pointer;
+  font-family: inherit;
+  transition: opacity var(--transition);
+}
+
+.legend-chip:hover {
+  filter: brightness(1.12);
+}
+
+/* Inactive (user clicked the chip to hide that node kind). */
+.legend-chip.legend-off {
+  opacity: 0.32;
+  text-decoration: line-through;
+  filter: saturate(0.4);
+}
+
+.legend-document {
+  background: #1e293b;
+}
+.legend-section {
+  background: #f97316;
+}
+.legend-paragraph {
+  background: #3b82f6;
+}
+.legend-table {
+  background: #8b5cf6;
+}
+.legend-figure {
+  background: #22c55e;
+}
+.legend-page {
+  background: #94a3b8;
+  color: #0f172a;
+}
+.legend-chunk {
+  background: #dc2626;
+}
+
+.graph-body {
+  flex: 1;
+  min-height: 0;
+  display: flex;
+  flex-direction: row;
+  overflow: hidden;
+}
+
+.graph-canvas-wrap {
+  flex: 1 1 auto;
+  min-width: 0;
+  position: relative;
+  overflow: hidden;
+}
+
+.graph-canvas {
+  position: absolute;
+  inset: 0;
+  background: var(--bg);
+}
+
+.graph-tooltip {
+  position: absolute;
+  transform: translate(-50%, calc(-100% - 14px));
+  max-width: 280px;
+  padding: 6px 10px;
+  background: rgba(15, 23, 42, 0.94);
+  color: #f8fafc;
+  font-size: 11px;
+  line-height: 1.4;
+  border-radius: var(--radius-sm);
+  pointer-events: none;
+  white-space: pre-wrap;
+  box-shadow: 0 4px 12px rgba(0, 0, 0, 0.2);
+  z-index: 20;
+}
+
+.graph-placeholder {
+  display: flex;
+  flex-direction: column;
+  align-items: center;
+  justify-content: center;
+  height: 100%;
+  gap: 12px;
+  color: var(--text-muted);
+  font-size: 14px;
+  padding: 32px;
+  text-align: center;
+}
+
+.graph-placeholder.error {
+  color: var(--error);
+}
+
+.retry-btn {
+  background: var(--accent);
+  color: white;
+  border: none;
+  padding: 6px 16px;
+  border-radius: var(--radius-sm);
+  cursor: pointer;
+  font-size: 12px;
+}
+
+.spinner-large {
+  width: 32px;
+  height: 32px;
+  border: 3px solid var(--border-light);
+  border-top-color: var(--accent);
+  border-radius: 50%;
+  animation: spin 0.8s linear infinite;
+}
+
+@keyframes spin {
+  to {
+    transform: rotate(360deg);
+  }
+}
+</style>

frontend/src/features/analysis/ui/NodeDetailsPanel.vue

@@ -0,0 +1,409 @@
+<template>
+  <aside
+    v-if="node"
+    class="nd-panel"
+    data-e2e="node-details-panel"
+    role="complementary"
+    :aria-label="t('graph.nodeDetails')"
+  >
+    <header class="nd-header">
+      <span class="nd-kind-chip" :style="{ background: kindColor }">{{ kindLabel }}</span>
+      <button class="nd-close" :aria-label="t('graph.close')" @click="$emit('close')">✕</button>
+    </header>
+
+    <dl class="nd-fields">
+      <template v-if="selfRef">
+        <dt>self_ref</dt>
+        <dd class="nd-mono">{{ selfRef }}</dd>
+      </template>
+
+      <template v-if="doclingLabel">
+        <dt>docling_label</dt>
+        <dd class="nd-mono">{{ doclingLabel }}</dd>
+      </template>
+
+      <template v-if="level != null">
+        <dt>level</dt>
+        <dd class="nd-mono">{{ level }}</dd>
+      </template>
+
+      <template v-if="pageNo != null">
+        <dt>{{ t('graph.page') }}</dt>
+        <dd class="nd-mono">p.{{ pageNo }}</dd>
+      </template>
+
+      <template v-if="chunkIndex != null">
+        <dt>chunk index</dt>
+        <dd class="nd-mono">#{{ chunkIndex }}</dd>
+      </template>
+    </dl>
+
+    <section v-if="text" class="nd-text-block">
+      <h4 class="nd-section-title">{{ t('graph.text') }}</h4>
+      <p class="nd-text">{{ text }}</p>
+    </section>
+
+    <section v-if="provs.length > 0" class="nd-provs-block">
+      <h4 class="nd-section-title">
+        {{ t('graph.provenances').replace('{n}', String(provs.length)) }}
+      </h4>
+      <ul class="nd-provs">
+        <li v-for="(p, i) in provs" :key="i" class="nd-prov">
+          <span class="nd-prov-page">p.{{ p.page_no ?? '?' }}</span>
+          <span class="nd-prov-bbox">
+            [{{ fmt(p.bbox_l) }}, {{ fmt(p.bbox_t) }}, {{ fmt(p.bbox_r) }}, {{ fmt(p.bbox_b) }}]
+          </span>
+          <span v-if="p.coord_origin" class="nd-prov-origin">{{ p.coord_origin }}</span>
+        </li>
+      </ul>
+    </section>
+
+    <section v-if="contents && contents.length > 0" class="nd-contents-block">
+      <h4 class="nd-section-title">
+        {{ t('graph.contains').replace('{n}', String(contents.length)) }}
+      </h4>
+      <ul class="nd-contents">
+        <li v-for="child in contents" :key="child.id">
+          <button
+            type="button"
+            class="nd-child"
+            :data-e2e="`node-details-child-${child.id}`"
+            @click="$emit('navigate', child)"
+          >
+            <span
+              class="nd-child-chip"
+              :style="{ background: kindColorFor(child) }"
+              :title="child.label ?? child.group"
+            >
+              {{ kindLabelFor(child) }}
+            </span>
+            <span class="nd-child-text">{{ previewText(child) }}</span>
+          </button>
+        </li>
+      </ul>
+    </section>
+  </aside>
+</template>
+
+<script setup lang="ts">
+import { computed } from 'vue'
+
+import { useI18n } from '../../../shared/i18n'
+import type { GraphNode, GraphProvenance } from '../graphApi'
+
+const props = defineProps<{
+  node: GraphNode | null
+  /**
+   * Nodes whose compound parent (PARENT_OF or synthetic section scope) is the
+   * currently-selected node. Computed upstream in GraphView so we don't have
+   * to re-walk the whole edge list here. Empty or null for leaf nodes.
+   */
+  contents?: readonly GraphNode[] | null
+}>()
+defineEmits<{
+  close: []
+  /** User clicked a child row — GraphView pans + swaps selection. */
+  navigate: [node: GraphNode]
+}>()
+
+const { t } = useI18n()
+
+// Centralised colour map, mirrors NODE_COLORS in GraphView — keeping a copy
+// here avoids coupling the detail panel to GraphView's internal state.
+const KIND_COLORS: Record<string, string> = {
+  document: '#1E293B',
+  SectionHeader: '#F97316',
+  Paragraph: '#3B82F6',
+  TextElement: '#3B82F6',
+  Table: '#8B5CF6',
+  Figure: '#22C55E',
+  List: '#06B6D4',
+  ListItem: '#06B6D4',
+  Formula: '#EC4899',
+  Code: '#14B8A6',
+  Caption: '#EAB308',
+  PageHeader: '#64748B',
+  PageFooter: '#64748B',
+  KeyValueArea: '#D946EF',
+  FormArea: '#D946EF',
+  DocumentIndex: '#0EA5E9',
+  Page: '#94A3B8',
+  Chunk: '#DC2626',
+}
+
+const kindLabel = computed<string>(() => {
+  const n = props.node
+  if (!n) return ''
+  if (n.group === 'document') return 'Document'
+  if (n.group === 'page') return 'Page'
+  if (n.group === 'chunk') return 'Chunk'
+  return n.label ?? 'Element'
+})
+
+const kindColor = computed<string>(() => {
+  const n = props.node
+  if (!n) return '#64748B'
+  if (n.group === 'document') return KIND_COLORS.document
+  if (n.group === 'page') return KIND_COLORS.Page
+  if (n.group === 'chunk') return KIND_COLORS.Chunk
+  return KIND_COLORS[n.label ?? ''] || KIND_COLORS.TextElement
+})
+
+const selfRef = computed(() => props.node?.self_ref ?? null)
+const doclingLabel = computed(() => (props.node?.docling_label as string | undefined) ?? null)
+const level = computed<number | null>(() => {
+  const v = props.node?.level
+  return typeof v === 'number' ? v : null
+})
+const pageNo = computed<number | null>(() => {
+  const v = props.node?.page_no ?? props.node?.prov_page
+  return typeof v === 'number' ? v : null
+})
+const chunkIndex = computed<number | null>(() => {
+  const v = props.node?.chunk_index
+  return typeof v === 'number' ? v : null
+})
+const text = computed<string>(() => (props.node?.text as string | undefined) ?? '')
+const provs = computed<GraphProvenance[]>(() => (props.node?.provs as GraphProvenance[]) ?? [])
+
+function fmt(n: number | null | undefined): string {
+  if (n == null) return '—'
+  return n.toFixed(1)
+}
+
+// Label + color helpers factored so they work for children too, not just the
+// currently-selected node. Keep them consistent with the chips above.
+function kindLabelFor(n: GraphNode): string {
+  if (n.group === 'document') return 'Document'
+  if (n.group === 'page') return 'Page'
+  if (n.group === 'chunk') return 'Chunk'
+  return n.label ?? 'Element'
+}
+
+function kindColorFor(n: GraphNode): string {
+  if (n.group === 'document') return KIND_COLORS.document
+  if (n.group === 'page') return KIND_COLORS.Page
+  if (n.group === 'chunk') return KIND_COLORS.Chunk
+  return KIND_COLORS[n.label ?? ''] || KIND_COLORS.TextElement
+}
+
+/**
+ * Short label for a child row. Prefer the node's own text (truncated), fall
+ * back to its self_ref so users can still recognise / debug missing text.
+ */
+function previewText(n: GraphNode): string {
+  const raw = (n.text as string | undefined) ?? ''
+  const clean = raw.replace(/\s+/g, ' ').trim()
+  if (clean) return clean.length > 80 ? clean.slice(0, 80) + '…' : clean
+  if (n.group === 'page') return `p.${n.page_no ?? '?'}`
+  if (n.group === 'chunk') return `chunk #${n.chunk_index ?? '?'}`
+  return n.self_ref ?? n.id
+}
+</script>
+
+<style scoped>
+.nd-panel {
+  display: flex;
+  flex-direction: column;
+  gap: 14px;
+  width: 320px;
+  flex: 0 0 320px;
+  padding: 14px 16px;
+  background: var(--bg);
+  border-left: 1px solid var(--border);
+  overflow-y: auto;
+  height: 100%;
+}
+
+.nd-header {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+}
+
+.nd-kind-chip {
+  display: inline-block;
+  padding: 3px 10px;
+  border-radius: 10px;
+  color: #f8fafc;
+  font-size: 11px;
+  font-weight: 600;
+  letter-spacing: 0.3px;
+}
+
+.nd-close {
+  background: transparent;
+  border: 0;
+  color: var(--text-muted);
+  font-size: 16px;
+  cursor: pointer;
+  padding: 2px 8px;
+  border-radius: var(--radius-sm);
+}
+
+.nd-close:hover {
+  background: var(--border-light);
+  color: var(--text);
+}
+
+.nd-fields {
+  display: grid;
+  grid-template-columns: 100px 1fr;
+  gap: 6px 10px;
+  margin: 0;
+  font-size: 12px;
+}
+
+.nd-fields dt {
+  color: var(--text-muted);
+  font-weight: 500;
+  text-transform: uppercase;
+  font-size: 10px;
+  letter-spacing: 0.4px;
+  align-self: center;
+}
+
+.nd-fields dd {
+  margin: 0;
+  color: var(--text);
+}
+
+.nd-mono {
+  font-family: 'IBM Plex Mono', monospace;
+  font-size: 11px;
+}
+
+.nd-section-title {
+  margin: 0 0 6px;
+  font-size: 10px;
+  font-weight: 600;
+  text-transform: uppercase;
+  letter-spacing: 0.5px;
+  color: var(--text-muted);
+}
+
+.nd-text-block {
+  display: flex;
+  flex-direction: column;
+  gap: 4px;
+  padding-top: 10px;
+  border-top: 1px solid var(--border-light);
+}
+
+.nd-text {
+  margin: 0;
+  font-size: 12px;
+  line-height: 1.5;
+  color: var(--text);
+  white-space: pre-wrap;
+}
+
+.nd-provs-block {
+  display: flex;
+  flex-direction: column;
+  gap: 4px;
+  padding-top: 10px;
+  border-top: 1px solid var(--border-light);
+}
+
+.nd-provs {
+  list-style: none;
+  padding: 0;
+  margin: 0;
+  display: flex;
+  flex-direction: column;
+  gap: 4px;
+}
+
+.nd-prov {
+  display: grid;
+  grid-template-columns: 40px 1fr auto;
+  gap: 6px;
+  align-items: baseline;
+  font-family: 'IBM Plex Mono', monospace;
+  font-size: 10px;
+  padding: 4px 6px;
+  background: var(--border-light);
+  border-radius: var(--radius-sm);
+}
+
+.nd-prov-page {
+  font-weight: 700;
+  color: var(--text);
+}
+
+.nd-prov-bbox {
+  color: var(--text-secondary);
+}
+
+.nd-prov-origin {
+  color: var(--text-muted);
+  font-size: 9px;
+  text-transform: lowercase;
+}
+
+.nd-contents-block {
+  display: flex;
+  flex-direction: column;
+  gap: 4px;
+  padding-top: 10px;
+  border-top: 1px solid var(--border-light);
+}
+
+.nd-contents {
+  list-style: none;
+  padding: 0;
+  margin: 0;
+  display: flex;
+  flex-direction: column;
+  gap: 4px;
+  /* Cap the list height so a section with hundreds of paragraphs doesn't
+   * blow the panel out. Scroll internally above that. */
+  max-height: 340px;
+  overflow-y: auto;
+}
+
+.nd-child {
+  display: flex;
+  align-items: baseline;
+  gap: 8px;
+  width: 100%;
+  text-align: left;
+  background: transparent;
+  border: 1px solid transparent;
+  border-radius: var(--radius-sm);
+  padding: 5px 8px;
+  cursor: pointer;
+  font: inherit;
+  color: inherit;
+  transition: all var(--transition);
+}
+
+.nd-child:hover {
+  background: var(--border-light);
+  border-color: var(--border);
+}
+
+.nd-child-chip {
+  flex: 0 0 auto;
+  display: inline-block;
+  padding: 1px 7px;
+  border-radius: 8px;
+  color: #f8fafc;
+  font-size: 9px;
+  font-weight: 600;
+  letter-spacing: 0.3px;
+}
+
+.nd-child-text {
+  flex: 1 1 auto;
+  font-size: 12px;
+  line-height: 1.4;
+  color: var(--text);
+  overflow: hidden;
+  text-overflow: ellipsis;
+  display: -webkit-box;
+  -webkit-line-clamp: 2;
+  -webkit-box-orient: vertical;
+}
+</style>

frontend/src/features/analysis/ui/StructureViewer.vue

@@ -40,8 +40,10 @@
       <canvas
         ref="canvasRef"
         class="overlay-canvas"
+        :class="{ selectable }"
         @mousemove="onMouseMove"
         @mouseleave="hoveredElement = null"
+        @click="onCanvasClick"
       />
       <!-- Tooltip -->
       <div v-if="hoveredElement" class="tooltip" :style="tooltipStyle">
@@ -76,8 +78,39 @@ const ELEMENT_COLORS: Record<string, string> = {
 const props = defineProps({
   pages: { type: Array as () => Page[], default: () => [] },
   documentId: String,
+  /**
+   * Reasoning-trace integration hooks. Optional — when unset, StructureViewer
+   * renders like before (Studio "Structure" tab). When set, enables overlays
+   * for the reasoning viewer without forking the component:
+   *
+   * - `visitedBySelfRef`: elements whose `self_ref` is in this map render in
+   *   the reasoning accent color with a numbered badge (the visit order).
+   * - `focusedSelfRef`: when it changes, auto-scroll to the page of that
+   *   element and pulse its bbox briefly.
+   * - `selectable`: when true, clicking a bbox emits `elementFocus` so a
+   *   parent can sync the selection with the graph view.
+   */
+  visitedBySelfRef: {
+    type: Object as () => Map<string, number> | null,
+    default: null,
+  },
+  focusedSelfRef: { type: String as () => string | null, default: null },
+  selectable: { type: Boolean, default: false },
+  /**
+   * When true AND `visitedBySelfRef` is set, non-visited elements are drawn
+   * with reduced alpha so the visited ones pop. Matches the reasoning
+   * panel's "Focus" toggle behavior on the graph.
+   */
+  dimNonVisited: { type: Boolean, default: false },
 })
 
+const emit = defineEmits<{
+  /** Fired when the user clicks a bbox — only if `selectable` is true. */
+  elementFocus: [selfRef: string]
+}>()
+
+const REASONING_COLOR = '#EA580C'
+
 const selectedPage = ref(1)
 const hiddenTypes = reactive(new Set<string>())
 const containerRef = ref<HTMLDivElement | null>(null)
@@ -136,39 +169,94 @@ function drawOverlay() {
 
   const scale = computeScale(img.clientWidth, img.clientHeight, page.width, page.height)
 
+  // Two-pass draw so reasoning overlays (highlight + pulse) sit on top of
+  // the base element strokes without being painted over by subsequent
+  // elements. First pass = base, second pass = accents.
   for (const el of visibleElements.value) {
     const rect = bboxToRect(el.bbox, scale)
-    const color = ELEMENT_COLORS[el.type] || ELEMENT_COLORS.text
+    const baseColor = ELEMENT_COLORS[el.type] || ELEMENT_COLORS.text
+    const isVisited =
+      props.visitedBySelfRef !== null && !!el.self_ref && props.visitedBySelfRef.has(el.self_ref)
 
-    ctx.strokeStyle = color
+    if (isVisited) {
-    ctx.lineWidth = 2
+      // Reasoning-visited element — reasoning accent color, bolder stroke,
-    ctx.strokeRect(rect.x, rect.y, rect.w, rect.h)
+      // more saturated fill than the base element. The visit-order badge
-
+      // is drawn in the second pass below.
-    ctx.fillStyle = color + '20'
+      ctx.strokeStyle = REASONING_COLOR
-    ctx.fillRect(rect.x, rect.y, rect.w, rect.h)
+      ctx.lineWidth = 3
+      ctx.strokeRect(rect.x, rect.y, rect.w, rect.h)
+      ctx.fillStyle = REASONING_COLOR + '33'
+      ctx.fillRect(rect.x, rect.y, rect.w, rect.h)
+    } else {
+      // Dim non-visited when focus mode is on and a visited set is present,
+      // so visited bboxes pop. Otherwise keep the regular styling.
+      const dim = props.dimNonVisited && props.visitedBySelfRef !== null
+      ctx.strokeStyle = baseColor + (dim ? '22' : '')
+      ctx.lineWidth = dim ? 1 : 2
+      ctx.strokeRect(rect.x, rect.y, rect.w, rect.h)
+      ctx.fillStyle = baseColor + (dim ? '08' : '20')
+      ctx.fillRect(rect.x, rect.y, rect.w, rect.h)
+    }
   }
+
+  // Second pass — numbered badges on visited elements + focus pulse ring.
+  for (const el of visibleElements.value) {
+    const rect = bboxToRect(el.bbox, scale)
+    const order =
+      props.visitedBySelfRef !== null && el.self_ref
+        ? props.visitedBySelfRef.get(el.self_ref)
+        : undefined
+    if (order !== undefined) {
+      drawVisitBadge(ctx, rect.x, rect.y, order)
+    }
+    if (props.focusedSelfRef && el.self_ref === props.focusedSelfRef) {
+      ctx.strokeStyle = REASONING_COLOR
+      ctx.lineWidth = 2
+      ctx.setLineDash([6, 4])
+      ctx.strokeRect(rect.x - 4, rect.y - 4, rect.w + 8, rect.h + 8)
+      ctx.setLineDash([])
+    }
+  }
+}
+
+function drawVisitBadge(ctx: CanvasRenderingContext2D, x: number, y: number, order: number): void {
+  const radius = 10
+  const cx = x
+  const cy = y
+  ctx.fillStyle = REASONING_COLOR
+  ctx.beginPath()
+  ctx.arc(cx, cy, radius, 0, Math.PI * 2)
+  ctx.fill()
+  ctx.fillStyle = '#ffffff'
+  ctx.font = 'bold 11px -apple-system, sans-serif'
+  ctx.textAlign = 'center'
+  ctx.textBaseline = 'middle'
+  ctx.fillText(String(order), cx, cy + 0.5)
+}
+
+function elementAt(e: MouseEvent): PageElement | null {
+  const canvas = canvasRef.value
+  const page = currentPageData.value
+  const img = imageRef.value
+  if (!canvas || !page || !img) return null
+  const canvasRect = canvas.getBoundingClientRect()
+  const mx = e.clientX - canvasRect.left
+  const my = e.clientY - canvasRect.top
+  const scale = computeScale(img.clientWidth, img.clientHeight, page.width, page.height)
+  for (const el of visibleElements.value) {
+    if (pointInRect(mx, my, bboxToRect(el.bbox, scale))) return el
+  }
+  return null
 }
 
 function onMouseMove(e: MouseEvent) {
   const canvas = canvasRef.value
-  const page = currentPageData.value
+  if (!canvas) return
-  const img = imageRef.value
-  if (!canvas || !page || !img) return
-
   const canvasRect = canvas.getBoundingClientRect()
   const mx = e.clientX - canvasRect.left
   const my = e.clientY - canvasRect.top
 
-  const scale = computeScale(img.clientWidth, img.clientHeight, page.width, page.height)
+  const found = elementAt(e)
-
-  let found: PageElement | null = null
-  for (const el of visibleElements.value) {
-    if (pointInRect(mx, my, bboxToRect(el.bbox, scale))) {
-      found = el
-      break
-    }
-  }
-
   hoveredElement.value = found
   if (found) {
     tooltipStyle.value = {
@@ -178,9 +266,51 @@ function onMouseMove(e: MouseEvent) {
   }
 }
 
+function onCanvasClick(e: MouseEvent): void {
+  if (!props.selectable) return
+  const el = elementAt(e)
+  if (el?.self_ref) emit('elementFocus', el.self_ref)
+}
+
 watch([() => props.pages, selectedPage, hiddenTypes], () => {
   nextTick(drawOverlay)
 })
+
+watch(
+  () => [props.visitedBySelfRef, props.dimNonVisited],
+  () => nextTick(drawOverlay),
+)
+
+// When the caller sets a focused self_ref (e.g. the user clicked a node in
+// the graph), find which page that element lives on and jump to it. The
+// overlay redraw will then show the dashed focus ring around its bbox.
+function scrollToFocused(ref: string | null): void {
+  if (!ref) {
+    nextTick(drawOverlay)
+    return
+  }
+  for (const page of props.pages) {
+    if (page.elements.some((e) => e.self_ref === ref)) {
+      if (selectedPage.value !== page.page_number) {
+        selectedPage.value = page.page_number
+        // Let <img> reload before drawing — drawOverlay runs on @load.
+      } else {
+        nextTick(drawOverlay)
+      }
+      return
+    }
+  }
+  // Ref not on any page (e.g. a #/body node) — just redraw to clear the
+  // previous focus ring.
+  nextTick(drawOverlay)
+}
+
+watch(() => props.focusedSelfRef, scrollToFocused)
+
+// Imperative entry point so callers can re-trigger a scroll on the same
+// self_ref (the watch above only fires on value change). Used by the
+// reasoning workspace when the user re-clicks the active iteration card.
+defineExpose({ scrollToFocused })
 </script>
 
 <style scoped>
@@ -282,6 +412,10 @@ watch([() => props.pages, selectedPage, hiddenTypes], () => {
   pointer-events: auto;
 }
 
+.overlay-canvas.selectable {
+  cursor: pointer;
+}
+
 .tooltip {
   position: absolute;
   background: var(--bg-surface);

frontend/src/features/feature-flags/store.test.ts

@@ -29,12 +29,12 @@ describe('useFeatureFlagStore', () => {
     expect(store.isEnabled('chunking')).toBe(true)
   })
 
-  it('disables chunking when engine is remote', async () => {
+  it('enables chunking when engine is remote', async () => {
     mockApiFetch.mockResolvedValue({ status: 'ok', engine: 'remote' })
     const store = useFeatureFlagStore()
     await store.load()
     expect(store.engine).toBe('remote')
-    expect(store.isEnabled('chunking')).toBe(false)
+    expect(store.isEnabled('chunking')).toBe(true)
   })
 
   it('enables disclaimer when deploymentMode is huggingface', async () => {

frontend/src/features/feature-flags/store.ts

@@ -16,7 +16,7 @@ interface HealthResponse {
   ingestionAvailable?: boolean
 }
 
-export type FeatureFlag = 'chunking' | 'disclaimer' | 'ingestion'
+export type FeatureFlag = 'chunking' | 'disclaimer' | 'ingestion' | 'reasoning'
 
 interface FeatureFlagDef {
   description: string
@@ -32,7 +32,7 @@ interface FeatureFlagContext {
 const featureRegistry: Record<FeatureFlag, FeatureFlagDef> = {
   chunking: {
     description: 'Document chunking for RAG preparation',
-    isEnabled: (ctx) => ctx.engine === 'local',
+    isEnabled: (ctx) => ctx.engine !== null,
   },
   disclaimer: {
     description: 'Show shared-instance disclaimer banner',
@@ -42,6 +42,13 @@ const featureRegistry: Record<FeatureFlag, FeatureFlagDef> = {
     description: 'OpenSearch ingestion pipeline (embedding + vector indexing)',
     isEnabled: (ctx) => ctx.ingestionAvailable,
   },
+  reasoning: {
+    // SQLite-backed (builds the graph from `document_json` on the fly), so no
+    // server-side gating needed. Kept as a flag so a future deployment can
+    // still kill-switch the UI if it wants to.
+    description: 'Reasoning trace tunnel (docling-agent RAGResult viewer)',
+    isEnabled: () => true,
+  },
 }
 
 export const useFeatureFlagStore = defineStore('feature-flags', () => {

frontend/src/features/feature-flags/useFeatureFlag.test.ts

@@ -23,6 +23,6 @@ describe('useFeatureFlag', () => {
     expect(flag.value).toBe(true)
 
     store.$patch({ engine: 'remote' })
-    expect(flag.value).toBe(false)
+    expect(flag.value).toBe(true)
   })
 })

frontend/src/features/reasoning/api.ts

@@ -0,0 +1,41 @@
+import { apiFetch } from '../../shared/api/http'
+import type { GraphPayload } from '../analysis/graphApi'
+import type { RAGResult } from './types'
+
+/**
+ * Fetch the reasoning-trace graph for a document — built on the backend from
+ * the SQLite `document_json` blob, not Neo4j. This is intentionally decoupled
+ * from Maintain's richer Neo4j graph: reasoning only needs the structural
+ * view (sections, parent/child, reading order, pages) to overlay iterations
+ * onto, and should work even if Neo4j isn't configured.
+ *
+ * 404 if no completed analysis with `document_json` exists for the doc.
+ */
+export function fetchReasoningGraph(docId: string): Promise<GraphPayload> {
+  return apiFetch<GraphPayload>(`/api/documents/${encodeURIComponent(docId)}/reasoning-graph`)
+}
+
+/**
+ * Kick off a `docling-agent` RAG run against a document and wait for the
+ * `RAGResult` (no streaming yet — the backend blocks on `_rag_loop` and
+ * returns once the loop converges or hits `max_iterations`).
+ *
+ * Runs typically take 20–40s depending on the model + Ollama latency. The
+ * caller should show a loading state.
+ *
+ * Errors:
+ *  - 503 if `RAG_ENABLED=false` server-side or docling-agent isn't installed
+ *  - 404 if no completed analysis exists for the doc
+ *  - 500 if the loop itself raises (Ollama unreachable, model missing, …)
+ */
+export function runReasoning(docId: string, query: string, modelId?: string): Promise<RAGResult> {
+  return apiFetch<RAGResult>(`/api/documents/${encodeURIComponent(docId)}/rag`, {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json' },
+    body: JSON.stringify({
+      query,
+      // Backend accepts snake_case; don't camelCase here.
+      model_id: modelId || undefined,
+    }),
+  })
+}

frontend/src/features/reasoning/graphReasoningOverlay.test.ts

@@ -0,0 +1,245 @@
+import cytoscape from 'cytoscape'
+import type { Core } from 'cytoscape'
+import { beforeEach, describe, expect, it } from 'vitest'
+
+import {
+  REASONING_EDGE_TYPE,
+  applyReasoningOverlay,
+  buildDegradedOverlay,
+  clearReasoningOverlay,
+  focusIteration,
+  nodeIdForSectionRef,
+} from './graphReasoningOverlay'
+import type { RAGResult } from './types'
+
+function seed(): Core {
+  // Headless mode — no DOM container needed.
+  return cytoscape({
+    headless: true,
+    elements: [
+      { data: { id: 'elem::#/texts/0', group: 'element' } },
+      { data: { id: 'elem::#/texts/3', group: 'element' } },
+      { data: { id: 'elem::#/texts/7', group: 'element' } },
+      { data: { id: 'elem::#/groups/1', group: 'element' } },
+    ],
+  })
+}
+
+function result(iterations: Array<Partial<RAGResult['iterations'][number]>>): RAGResult {
+  return {
+    answer: 'x',
+    converged: true,
+    iterations: iterations.map((it, i) => ({
+      iteration: i + 1,
+      section_ref: '',
+      reason: '',
+      section_text_length: 0,
+      can_answer: false,
+      response: '',
+      ...it,
+    })),
+  }
+}
+
+describe('nodeIdForSectionRef', () => {
+  it('matches the backend _element_node id format', () => {
+    // Kept in sync with document-parser/infra/neo4j/queries.py::_element_node.
+    expect(nodeIdForSectionRef('#/texts/3')).toBe('elem::#/texts/3')
+  })
+})
+
+describe('applyReasoningOverlay', () => {
+  let cy: Core
+  beforeEach(() => {
+    cy = seed()
+  })
+
+  it('marks every resolvable section as visited with its iteration order', () => {
+    const res = applyReasoningOverlay(
+      cy,
+      result([{ section_ref: '#/texts/0' }, { section_ref: '#/texts/3' }]),
+    )
+
+    expect(res.presentCount).toBe(2)
+    expect(res.missingCount).toBe(0)
+    expect(cy.getElementById('elem::#/texts/0').hasClass('visited')).toBe(true)
+    expect(cy.getElementById('elem::#/texts/0').data('visitOrder')).toBe(1)
+    expect(cy.getElementById('elem::#/texts/3').data('visitOrder')).toBe(2)
+  })
+
+  it('adds a REASONING_NEXT edge between each consecutive visited pair', () => {
+    applyReasoningOverlay(
+      cy,
+      result([
+        { section_ref: '#/texts/0' },
+        { section_ref: '#/texts/3' },
+        { section_ref: '#/texts/7' },
+      ]),
+    )
+
+    const edges = cy.$(`edge[type = "${REASONING_EDGE_TYPE}"]`)
+    expect(edges.length).toBe(2)
+    const ids = edges.map((e) => e.id()).sort()
+    expect(ids).toEqual([
+      'REASONING_NEXT::elem::#/texts/0::elem::#/texts/3',
+      'REASONING_NEXT::elem::#/texts/3::elem::#/texts/7',
+    ])
+  })
+
+  it('reports missing section refs and does not crash on them', () => {
+    const res = applyReasoningOverlay(
+      cy,
+      result([
+        { section_ref: '#/texts/0' },
+        { section_ref: '#/texts/999' }, // not in graph
+        { section_ref: '#/texts/3' },
+      ]),
+    )
+
+    expect(res.presentCount).toBe(2)
+    expect(res.missingCount).toBe(1)
+    expect(res.resolved[1].present).toBe(false)
+    expect(cy.getElementById('elem::#/texts/999').nonempty()).toBe(false)
+  })
+
+  it('breaks the arrow chain across a missing iteration (no ghost edges)', () => {
+    applyReasoningOverlay(
+      cy,
+      result([
+        { section_ref: '#/texts/0' },
+        { section_ref: '#/texts/999' }, // missing
+        { section_ref: '#/texts/3' },
+      ]),
+    )
+
+    // Only the chain between present-to-present pairs gets edges. With one
+    // missing in the middle, we still draw 0→3 because the filter keeps the
+    // present ones adjacent in the sequence used for edge drawing. Assert it:
+    // one edge between 0 and 3.
+    const edges = cy.$(`edge[type = "${REASONING_EDGE_TYPE}"]`)
+    expect(edges.length).toBe(1)
+    expect(edges[0].data('source')).toBe('elem::#/texts/0')
+    expect(edges[0].data('target')).toBe('elem::#/texts/3')
+  })
+
+  it('is idempotent — re-applying replaces the previous overlay', () => {
+    applyReasoningOverlay(cy, result([{ section_ref: '#/texts/0' }]))
+    applyReasoningOverlay(cy, result([{ section_ref: '#/texts/3' }, { section_ref: '#/texts/7' }]))
+
+    expect(cy.getElementById('elem::#/texts/0').hasClass('visited')).toBe(false)
+    expect(cy.getElementById('elem::#/texts/3').hasClass('visited')).toBe(true)
+    expect(cy.$(`edge[type = "${REASONING_EDGE_TYPE}"]`).length).toBe(1)
+  })
+
+  it('dims every non-visited node so the trace pops visually', () => {
+    applyReasoningOverlay(cy, result([{ section_ref: '#/texts/0' }, { section_ref: '#/texts/3' }]))
+
+    // Visited ones keep their full opacity (no dimmed class).
+    expect(cy.getElementById('elem::#/texts/0').hasClass('dimmed')).toBe(false)
+    expect(cy.getElementById('elem::#/texts/3').hasClass('dimmed')).toBe(false)
+    // Everything else gets dimmed.
+    expect(cy.getElementById('elem::#/texts/7').hasClass('dimmed')).toBe(true)
+    expect(cy.getElementById('elem::#/groups/1').hasClass('dimmed')).toBe(true)
+  })
+
+  it('does not dim anything when the trace has no resolvable iterations', () => {
+    // All missing → we don't want to wash out the whole graph for nothing.
+    applyReasoningOverlay(cy, result([{ section_ref: '#/texts/999' }]))
+    expect(cy.$('.dimmed').length).toBe(0)
+  })
+
+  it('skips dimming entirely when focusMode is off', () => {
+    applyReasoningOverlay(
+      cy,
+      result([{ section_ref: '#/texts/0' }, { section_ref: '#/texts/3' }]),
+      { focusMode: false },
+    )
+    // Visited still highlighted...
+    expect(cy.getElementById('elem::#/texts/0').hasClass('visited')).toBe(true)
+    // ...but nothing else gets dimmed.
+    expect(cy.$('.dimmed').length).toBe(0)
+  })
+
+  it('does not dim the synthetic REASONING_NEXT edges', () => {
+    applyReasoningOverlay(
+      cy,
+      result([
+        { section_ref: '#/texts/0' },
+        { section_ref: '#/texts/3' },
+        { section_ref: '#/texts/7' },
+      ]),
+    )
+    const reasoningEdges = cy.$(`edge[type = "${REASONING_EDGE_TYPE}"]`)
+    expect(reasoningEdges.length).toBe(2)
+    reasoningEdges.forEach((e) => {
+      expect(e.hasClass('dimmed')).toBe(false)
+    })
+  })
+
+  it('preserves the original graph elements (no destructive mutations)', () => {
+    const beforeIds = cy
+      .elements()
+      .map((e) => e.id())
+      .sort()
+    applyReasoningOverlay(cy, result([{ section_ref: '#/texts/0' }]))
+    clearReasoningOverlay(cy)
+    const afterIds = cy
+      .elements()
+      .map((e) => e.id())
+      .sort()
+    expect(afterIds).toEqual(beforeIds)
+  })
+})
+
+describe('clearReasoningOverlay', () => {
+  it('removes class, data, dimming, and synthetic edges', () => {
+    const cy = seed()
+    applyReasoningOverlay(cy, result([{ section_ref: '#/texts/0' }, { section_ref: '#/texts/3' }]))
+    clearReasoningOverlay(cy)
+
+    expect(cy.$('.visited').length).toBe(0)
+    expect(cy.$('.dimmed').length).toBe(0)
+    expect(cy.getElementById('elem::#/texts/0').data('visitOrder')).toBeUndefined()
+    expect(cy.$(`edge[type = "${REASONING_EDGE_TYPE}"]`).length).toBe(0)
+  })
+
+  it('is a no-op when nothing is overlaid', () => {
+    const cy = seed()
+    expect(() => clearReasoningOverlay(cy)).not.toThrow()
+  })
+})
+
+describe('buildDegradedOverlay', () => {
+  // Used when the Neo4j graph isn't loaded — we still want to show the
+  // reasoning trace cards, just without graph positioning.
+  it('returns every iteration with present=false', () => {
+    const out = buildDegradedOverlay(
+      result([
+        { section_ref: '#/texts/0', reason: 'r0', can_answer: false },
+        { section_ref: '#/texts/3', reason: 'r1', can_answer: true, response: 'done' },
+      ]),
+    )
+    expect(out.resolved).toHaveLength(2)
+    expect(out.presentCount).toBe(0)
+    expect(out.missingCount).toBe(2)
+    expect(out.resolved.every((r) => !r.present)).toBe(true)
+    expect(out.resolved[0].reason).toBe('r0')
+    expect(out.resolved[1].canAnswer).toBe(true)
+    expect(out.resolved[1].response).toBe('done')
+    expect(out.resolved[0].nodeId).toBe('elem::#/texts/0')
+  })
+
+  it('handles an empty iterations list', () => {
+    const out = buildDegradedOverlay(result([]))
+    expect(out.resolved).toEqual([])
+    expect(out.presentCount).toBe(0)
+    expect(out.missingCount).toBe(0)
+  })
+})
+
+describe('focusIteration', () => {
+  it('is a no-op for a missing node', () => {
+    const cy = seed()
+    expect(() => focusIteration(cy, 'elem::#/texts/does-not-exist')).not.toThrow()
+  })
+})

frontend/src/features/reasoning/graphReasoningOverlay.ts

@@ -0,0 +1,278 @@
+/**
+ * Pure Cytoscape-manipulation helpers for the reasoning-trace overlay.
+ *
+ * Keeping this as a plain TS module (not a component) makes it trivially
+ * testable against `cytoscape({ headless: true })` and lets the same code
+ * drive both the v1 static-import UX and the v2 streaming-runner UX.
+ *
+ * The overlay is purely visual: it adds a class + a transient `visitOrder`
+ * data attribute to existing element nodes and injects synthetic
+ * `REASONING_NEXT` edges between successive visited nodes. None of this is
+ * persisted to Neo4j.
+ */
+import type { Core } from 'cytoscape'
+
+import type { OverlayResult, RAGResult, ResolvedIteration } from './types'
+
+export const REASONING_EDGE_TYPE = 'REASONING_NEXT'
+const VISITED_CLASS = 'visited'
+const DIMMED_CLASS = 'dimmed'
+
+/**
+ * Build the Cytoscape node id for a given `section_ref`.
+ *
+ * Must stay in sync with `document-parser/infra/neo4j/queries.py::_element_node`
+ * which emits `f"elem::{self_ref}"`. The `#` that lives inside a Docling
+ * self_ref (e.g. `#/texts/3`) is fine inside an id string — we just always
+ * look up via `cy.getElementById()` rather than the `#id` selector syntax,
+ * which would conflict with the leading `#`.
+ */
+export function nodeIdForSectionRef(sectionRef: string): string {
+  return `elem::${sectionRef}`
+}
+
+export interface OverlayOptions {
+  /**
+   * When true (default), dim every non-visited element and hide the Document
+   * root — the user sees the trace pop against a muted background.
+   * When false, the graph keeps its full colors; only visited nodes get the
+   * orange ring + numbered badge and the REASONING_NEXT arrows are drawn.
+   * Toggled from the ReasoningPanel header.
+   */
+  focusMode?: boolean
+}
+
+/**
+ * Apply the overlay for a freshly imported `RAGResult`.
+ *
+ * Idempotent: any previous overlay is cleared first. Returns a summary that
+ * the caller (store / panel) uses to drive the UI (list of iterations,
+ * missing count, etc).
+ */
+export function applyReasoningOverlay(
+  cy: Core,
+  result: RAGResult,
+  options: OverlayOptions = {},
+): OverlayResult {
+  const focusMode = options.focusMode ?? true
+  clearReasoningOverlay(cy)
+
+  const resolved: ResolvedIteration[] = result.iterations.map((it) => {
+    const nodeId = nodeIdForSectionRef(it.section_ref)
+    const node = cy.getElementById(nodeId)
+    const present = node.nonempty()
+    if (present) {
+      node.addClass(VISITED_CLASS)
+      node.data('visitOrder', it.iteration)
+    }
+    return {
+      iteration: it.iteration,
+      sectionRef: it.section_ref,
+      nodeId,
+      present,
+      reason: it.reason,
+      canAnswer: it.can_answer,
+      response: it.response,
+      sectionTextLength: it.section_text_length,
+    }
+  })
+
+  // Dim everything that isn't part of the trace. We do this BEFORE injecting
+  // the synthetic REASONING_NEXT edges so those edges never receive `.dimmed`
+  // — they're the foreground of the viz. Takes the inspiration from the
+  // BboxOverlay approach (dim non-highlighted bboxes, keep the active ones
+  // fully opaque) and applies it to the graph.
+  //
+  // The Document root node is hidden outright (via a `display: none` rule on
+  // `node.dimmed[group = "document"]`) — it sits at the center of the layout
+  // and adds zero signal to a reasoning trace, only visual noise. Its
+  // connected `HAS_ROOT` edge is hidden automatically by Cytoscape.
+  const present = resolved.filter((r) => r.present)
+  if (focusMode && present.length > 0) {
+    cy.elements().not(`.${VISITED_CLASS}`).addClass(DIMMED_CLASS)
+  }
+
+  // Draw trace arrows only between successively-present iterations. Gaps
+  // (missing nodes) break the chain — we don't draw edges to ghosts.
+  for (let i = 0; i < present.length - 1; i++) {
+    const src = present[i]
+    const tgt = present[i + 1]
+    cy.add({
+      group: 'edges',
+      data: {
+        id: `${REASONING_EDGE_TYPE}::${src.nodeId}::${tgt.nodeId}`,
+        source: src.nodeId,
+        target: tgt.nodeId,
+        type: REASONING_EDGE_TYPE,
+        order: i,
+      },
+    })
+  }
+
+  const visitedEles = cy.$(`.${VISITED_CLASS}`)
+  if (visitedEles.nonempty()) {
+    // Padding keeps the arrows readable when the trace is one or two nodes.
+    cy.fit(visitedEles, 80)
+  }
+
+  return {
+    resolved,
+    presentCount: present.length,
+    missingCount: resolved.length - present.length,
+  }
+}
+
+/**
+ * Build a "degraded" overlay result when no Cytoscape instance is available
+ * (graph failed to load, or graph is empty for this document). Every iteration
+ * is marked `present: false`; the panel can still render the iteration cards
+ * so the user sees the reasoning — they just don't get highlights on the graph.
+ */
+export function buildDegradedOverlay(result: RAGResult): OverlayResult {
+  const resolved: ResolvedIteration[] = result.iterations.map((it) => ({
+    iteration: it.iteration,
+    sectionRef: it.section_ref,
+    nodeId: nodeIdForSectionRef(it.section_ref),
+    present: false,
+    reason: it.reason,
+    canAnswer: it.can_answer,
+    response: it.response,
+    sectionTextLength: it.section_text_length,
+  }))
+  return {
+    resolved,
+    presentCount: 0,
+    missingCount: resolved.length,
+  }
+}
+
+/**
+ * Remove every overlay artifact from the graph. Safe to call when nothing
+ * is overlaid — it becomes a no-op.
+ */
+export function clearReasoningOverlay(cy: Core): void {
+  cy.$(`.${VISITED_CLASS}`).forEach((n) => {
+    n.removeClass(VISITED_CLASS)
+    n.removeData('visitOrder')
+  })
+  cy.$(`.${DIMMED_CLASS}`).removeClass(DIMMED_CLASS)
+  cy.$(`edge[type = "${REASONING_EDGE_TYPE}"]`).remove()
+}
+
+/**
+ * Center the viewport on a single iteration's node and give it a quick
+ * visual pulse. No-op if the node isn't present (missing iteration).
+ */
+export function focusIteration(cy: Core, nodeId: string): void {
+  const node = cy.getElementById(nodeId)
+  if (!node.nonempty()) return
+  cy.animate(
+    {
+      center: { eles: node },
+      zoom: Math.max(cy.zoom(), 1.0),
+    },
+    { duration: 300 },
+  )
+  // flashClass is a cytoscape-builtin: add the class then remove it after N ms.
+  node.flashClass('pulse', 800)
+}
+
+/**
+ * Shape of a single Cytoscape stylesheet block used for the overlay. We keep
+ * it intentionally loose (`Record<string, unknown>` for `style`) because the
+ * upstream cytoscape types fork selectors by node-vs-edge prefix and flag
+ * cross-type properties, while the runtime itself accepts everything we emit
+ * here.
+ */
+interface StyleBlock {
+  selector: string
+  style: Record<string, unknown>
+}
+
+/**
+ * Style rules appended to the existing GraphView stylesheet. Exported so the
+ * component can spread them into its Cytoscape config.
+ */
+export const reasoningOverlayStyles: StyleBlock[] = [
+  // Non-trace elements get dimmed — BboxOverlay-inspired: keep the colors,
+  // drop the opacity hard so the trace pops. Node opacity cascades to label
+  // + border; edges get a stronger fade because they add visual noise.
+  {
+    selector: `node.${DIMMED_CLASS}`,
+    style: {
+      opacity: 0.18,
+      'text-opacity': 0.25,
+    },
+  },
+  {
+    selector: `edge.${DIMMED_CLASS}`,
+    style: {
+      opacity: 0.08,
+    },
+  },
+  // Hide the Document root node entirely when a trace is active — it sits
+  // at the center of the dagre layout and is pure noise for reasoning-path
+  // inspection. Its `HAS_ROOT` edge is hidden automatically by Cytoscape
+  // because `display: none` cascades to connected edges.
+  {
+    selector: `node.${DIMMED_CLASS}[group = "document"]`,
+    style: {
+      display: 'none',
+    },
+  },
+  // Visited nodes: orange ring + label-less numbered badge above them.
+  // Explicit opacity: 1 prevents inheritance quirks when the user re-applies
+  // an overlay on top of an existing one.
+  {
+    selector: `node.${VISITED_CLASS}`,
+    style: {
+      'border-color': '#EA580C',
+      'border-width': 4,
+      'overlay-color': '#EA580C',
+      'overlay-opacity': 0.08,
+      'overlay-padding': 4,
+      opacity: 1,
+      'z-index': 50,
+    },
+  },
+  {
+    selector: `node.${VISITED_CLASS}[visitOrder]`,
+    style: {
+      label: 'data(visitOrder)',
+      'text-valign': 'top',
+      'text-margin-y': -6,
+      'text-background-color': '#EA580C',
+      'text-background-opacity': 1,
+      'text-background-padding': '3px',
+      'text-background-shape': 'roundrectangle',
+      'text-border-color': '#FFFFFF',
+      'text-border-width': 1.5,
+      'text-border-opacity': 1,
+      color: '#FFFFFF',
+      'font-weight': 700,
+      'font-size': 12,
+      'text-opacity': 1,
+    },
+  },
+  {
+    selector: `edge[type = "${REASONING_EDGE_TYPE}"]`,
+    style: {
+      'line-color': '#EA580C',
+      'target-arrow-color': '#EA580C',
+      'target-arrow-shape': 'triangle',
+      'arrow-scale': 1.4,
+      'curve-style': 'bezier',
+      'control-point-step-size': 40,
+      width: 3,
+      opacity: 0.95,
+      'z-index': 99,
+    },
+  },
+  {
+    selector: 'node.pulse',
+    style: {
+      'border-width': 7,
+      'border-color': '#F59E0B',
+    },
+  },
+]

frontend/src/features/reasoning/store.test.ts

@@ -0,0 +1,84 @@
+import { createPinia, setActivePinia } from 'pinia'
+import { beforeEach, describe, expect, it } from 'vitest'
+
+import { parseImportedTrace, useReasoningStore } from './store'
+
+describe('parseImportedTrace', () => {
+  const bare = {
+    answer: 'ok',
+    converged: true,
+    iterations: [],
+  }
+
+  it('accepts a bare RAGResult', () => {
+    const parsed = parseImportedTrace(bare)
+    expect(parsed?.result.answer).toBe('ok')
+    expect(parsed?.envelope).toBeNull()
+  })
+
+  it('accepts a sidecar envelope and extracts the result', () => {
+    const parsed = parseImportedTrace({
+      job_id: 'abc',
+      filename: 'x.pdf',
+      result: bare,
+    })
+    expect(parsed?.result.answer).toBe('ok')
+    expect(parsed?.envelope?.job_id).toBe('abc')
+  })
+
+  it('rejects shapes that are neither envelope nor bare RAGResult', () => {
+    expect(parseImportedTrace(null)).toBeNull()
+    expect(parseImportedTrace('string')).toBeNull()
+    expect(parseImportedTrace({ foo: 'bar' })).toBeNull()
+    expect(parseImportedTrace({ answer: 'x' })).toBeNull() // missing converged + iterations
+  })
+})
+
+describe('useReasoningStore', () => {
+  beforeEach(() => {
+    setActivePinia(createPinia())
+  })
+
+  it('starts empty', () => {
+    const s = useReasoningStore()
+    expect(s.hasTrace).toBe(false)
+    expect(s.iterations).toEqual([])
+    expect(s.presentCount).toBe(0)
+  })
+
+  it('setResult populates and resets active iteration', () => {
+    const s = useReasoningStore()
+    s.setActiveIteration(3)
+    s.setResult({ answer: 'a', converged: true, iterations: [] }, null)
+    expect(s.hasTrace).toBe(true)
+    expect(s.activeIteration).toBeNull()
+  })
+
+  it('toggleFocusMode flips the flag', () => {
+    const s = useReasoningStore()
+    expect(s.focusMode).toBe(true)
+    s.toggleFocusMode()
+    expect(s.focusMode).toBe(false)
+    s.toggleFocusMode()
+    expect(s.focusMode).toBe(true)
+  })
+
+  it('reset restores focusMode to the default on', () => {
+    const s = useReasoningStore()
+    s.toggleFocusMode()
+    expect(s.focusMode).toBe(false)
+    s.reset()
+    expect(s.focusMode).toBe(true)
+  })
+
+  it('reset clears everything including the dialog flag', () => {
+    const s = useReasoningStore()
+    s.openImportDialog()
+    s.setResult({ answer: 'a', converged: true, iterations: [] }, null)
+    s.setError('boom')
+    s.reset()
+    expect(s.hasTrace).toBe(false)
+    expect(s.error).toBeNull()
+    expect(s.importDialogOpen).toBe(false)
+  })
+})

frontend/src/features/reasoning/store.ts

@@ -0,0 +1,156 @@
+import { defineStore } from 'pinia'
+import { computed, ref } from 'vue'
+
+import type { OverlayResult, RAGResult, ResolvedIteration, SidecarEnvelope } from './types'
+
+/**
+ * Parse an arbitrary JSON payload as either:
+ *   - a bare `RAGResult` (what `docling-agent` emits directly), or
+ *   - the sidecar envelope (`{ job_id, filename, query, model, result }`)
+ *
+ * Returns `null` if the shape doesn't match either.
+ */
+export function parseImportedTrace(
+  raw: unknown,
+): { result: RAGResult; envelope: SidecarEnvelope | null } | null {
+  if (!raw || typeof raw !== 'object') return null
+  const obj = raw as Record<string, unknown>
+
+  // Envelope shape
+  if (obj.result && typeof obj.result === 'object') {
+    const result = obj.result as RAGResult
+    if (isRAGResult(result)) {
+      return { result, envelope: obj as unknown as SidecarEnvelope }
+    }
+  }
+  // Bare RAGResult
+  if (isRAGResult(obj as unknown as RAGResult)) {
+    return { result: obj as unknown as RAGResult, envelope: null }
+  }
+  return null
+}
+
+function isRAGResult(x: RAGResult | undefined): boolean {
+  if (!x || typeof x !== 'object') return false
+  return (
+    typeof x.answer === 'string' && typeof x.converged === 'boolean' && Array.isArray(x.iterations)
+  )
+}
+
+export const useReasoningStore = defineStore('reasoning', () => {
+  const importDialogOpen = ref(false)
+  // Separate modal for the live runner (POST /api/documents/:id/rag), so it
+  // can coexist with the import dialog conceptually even if only one is ever
+  // open at a time.
+  const runDialogOpen = ref(false)
+  const running = ref(false)
+  const rawResult = ref<RAGResult | null>(null)
+  const envelope = ref<SidecarEnvelope | null>(null)
+  const overlay = ref<OverlayResult | null>(null)
+  const activeIteration = ref<number | null>(null)
+  const error = ref<string | null>(null)
+  // Focus mode: when true, non-visited elements are dimmed so the trace pops.
+  // Default ON because that's the primary value of the feature; user can
+  // switch it off from the panel to see the trace in the full graph context.
+  const focusMode = ref(true)
+
+  const hasTrace = computed(() => rawResult.value !== null)
+  const iterations = computed<ResolvedIteration[]>(() => overlay.value?.resolved ?? [])
+  const presentCount = computed(() => overlay.value?.presentCount ?? 0)
+  const missingCount = computed(() => overlay.value?.missingCount ?? 0)
+
+  function openImportDialog(): void {
+    importDialogOpen.value = true
+  }
+
+  function closeImportDialog(): void {
+    importDialogOpen.value = false
+  }
+
+  function openRunDialog(): void {
+    runDialogOpen.value = true
+  }
+
+  function closeRunDialog(): void {
+    runDialogOpen.value = false
+  }
+
+  function setRunning(v: boolean): void {
+    running.value = v
+  }
+
+  /**
+   * Called by `ImportTraceDialog` once the user has supplied a JSON file.
+   * Does NOT touch Cytoscape — the `ReasoningPanel` watches `rawResult` and
+   * reapplies the overlay via `graphReasoningOverlay.applyReasoningOverlay`.
+   */
+  function setResult(result: RAGResult, env: SidecarEnvelope | null): void {
+    rawResult.value = result
+    envelope.value = env
+    error.value = null
+    activeIteration.value = null
+  }
+
+  function setOverlay(o: OverlayResult | null): void {
+    overlay.value = o
+  }
+
+  function setActiveIteration(n: number | null): void {
+    // Pure state — drives the active-card highlight in the iteration list.
+    // Side effects (graph pan, PDF scroll) are dispatched imperatively from
+    // ReasoningWorkspace.onIterationFocus so re-clicking the same iteration
+    // still re-focuses both views (a watch here would no-op on same value).
+    activeIteration.value = n
+  }
+
+  function setError(msg: string | null): void {
+    error.value = msg
+  }
+
+  function toggleFocusMode(): void {
+    focusMode.value = !focusMode.value
+  }
+
+  /** Full reset — e.g. when the user switches document. */
+  function reset(): void {
+    rawResult.value = null
+    envelope.value = null
+    overlay.value = null
+    activeIteration.value = null
+    error.value = null
+    importDialogOpen.value = false
+    runDialogOpen.value = false
+    running.value = false
+    focusMode.value = true
+  }
+
+  return {
+    // state
+    importDialogOpen,
+    runDialogOpen,
+    running,
+    rawResult,
+    envelope,
+    overlay,
+    activeIteration,
+    error,
+    focusMode,
+    // computed
+    hasTrace,
+    iterations,
+    presentCount,
+    missingCount,
+    // actions
+    openImportDialog,
+    closeImportDialog,
+    openRunDialog,
+    closeRunDialog,
+    setRunning,
+    setResult,
+    setOverlay,
+    setActiveIteration,
+    setError,
+    toggleFocusMode,
+    reset,
+  }
+})

frontend/src/features/reasoning/types.ts

@@ -0,0 +1,65 @@
+/**
+ * Types mirroring the `docling-agent` RAG output.
+ *
+ * The JSON imported by the user is produced either by:
+ *   - the R&D sidecar (`experiments/reasoning-trace/inspect_doc.py`), or
+ *   - any external `docling-agent` run that was serialized to JSON.
+ *
+ * Since `docling-agent` uses plain pydantic (no alias generator), field names
+ * are **snake_case** here. This is one of the rare spots in the frontend where
+ * we don't normalize to camelCase — keeping the shape 1:1 with upstream means
+ * a schema drift upstream gives us a clean type error rather than silent
+ * re-mapping.
+ *
+ * Source of truth: docling-project/docling-agent @ docling_agent/agent/rag_models.py
+ */
+
+export interface RAGIteration {
+  iteration: number
+  section_ref: string
+  reason: string
+  section_text_length: number
+  can_answer: boolean
+  response: string
+}
+
+export interface RAGResult {
+  answer: string
+  iterations: RAGIteration[]
+  converged: boolean
+}
+
+/**
+ * Envelope written by the R&D sidecar. The viewer also accepts a bare
+ * `RAGResult` (see `parseImportedTrace` in the store).
+ */
+export interface SidecarEnvelope {
+  job_id?: string
+  filename?: string
+  query?: string
+  model?: { ollama_name?: string | null; hf_model_name?: string | null }
+  max_iterations?: number
+  result: RAGResult
+}
+
+/**
+ * One iteration after matching its `section_ref` against the currently-loaded
+ * Cytoscape graph. `present=false` means the section ref has no corresponding
+ * node (doc not through Maintain, or a different version of the doc).
+ */
+export interface ResolvedIteration {
+  iteration: number
+  sectionRef: string
+  nodeId: string
+  present: boolean
+  reason: string
+  canAnswer: boolean
+  response: string
+  sectionTextLength: number
+}
+
+export interface OverlayResult {
+  resolved: ResolvedIteration[]
+  presentCount: number
+  missingCount: number
+}

frontend/src/features/reasoning/ui/DocumentView.vue

@@ -0,0 +1,107 @@
+<template>
+  <div class="rdv-root" data-e2e="reasoning-document-view">
+    <div v-if="!pages || pages.length === 0" class="rdv-empty">
+      {{ t('reasoning.docNoContent') }}
+    </div>
+    <StructureViewer
+      v-else
+      ref="structureViewerRef"
+      :pages="pages"
+      :document-id="docId"
+      :visited-by-self-ref="visitedBySelfRef"
+      :focused-self-ref="focusedSelfRef"
+      :dim-non-visited="reasoningStore.focusMode"
+      selectable
+      class="rdv-viewer"
+      @element-focus="(ref) => emit('elementFocus', ref)"
+    />
+  </div>
+</template>
+
+<script setup lang="ts">
+/**
+ * PDF rendering of the document (per-page PNG via /api/documents/:id/preview),
+ * augmented with reasoning overlays:
+ *   - elements visited by the RAG loop get a bold orange stroke + numbered
+ *     badge showing the visit order
+ *   - the current `focusedSelfRef` (usually driven by a graph-node click
+ *     upstream) auto-jumps to the right page and pulses its bbox
+ *   - clicking a bbox emits `elementFocus` so the graph can mirror the
+ *     selection (bidirectional sync handled in ReasoningWorkspace).
+ */
+import { computed, ref } from 'vue'
+
+import StructureViewer from '../../analysis/ui/StructureViewer.vue'
+import { useAnalysisStore } from '../../analysis/store'
+import { useI18n } from '../../../shared/i18n'
+import type { Page } from '../../../shared/types'
+import { useReasoningStore } from '../store'
+
+const props = defineProps<{
+  docId: string
+  focusedSelfRef: string | null
+}>()
+const emit = defineEmits<{ elementFocus: [selfRef: string] }>()
+
+const structureViewerRef = ref<InstanceType<typeof StructureViewer> | null>(null)
+
+// Imperative passthrough so the reasoning workspace can re-trigger a scroll
+// on the same self_ref (e.g. user re-clicks the active iteration card —
+// the prop hasn't changed so the watch wouldn't fire).
+function scrollToFocused(selfRef: string | null): void {
+  structureViewerRef.value?.scrollToFocused(selfRef)
+}
+
+defineExpose({ scrollToFocused })
+
+const analysisStore = useAnalysisStore()
+const reasoningStore = useReasoningStore()
+const { t } = useI18n()
+
+const pages = computed<Page[]>(() => {
+  const hit = analysisStore.analyses.find(
+    (a) => a.documentId === props.docId && a.status === 'COMPLETED' && a.pagesJson,
+  )
+  if (!hit?.pagesJson) return []
+  try {
+    return JSON.parse(hit.pagesJson) as Page[]
+  } catch {
+    return []
+  }
+})
+
+const visitedBySelfRef = computed<Map<string, number>>(() => {
+  const out = new Map<string, number>()
+  for (const it of reasoningStore.iterations) {
+    if (!it.present || !it.sectionRef) continue
+    if (!out.has(it.sectionRef)) out.set(it.sectionRef, it.iteration)
+  }
+  return out
+})
+</script>
+
+<style scoped>
+.rdv-root {
+  flex: 1 1 auto;
+  min-width: 0;
+  min-height: 0;
+  overflow-y: auto;
+  background: var(--bg);
+  padding: 16px 20px;
+}
+
+.rdv-viewer {
+  max-width: 960px;
+  margin: 0 auto;
+}
+
+.rdv-empty {
+  display: flex;
+  align-items: center;
+  justify-content: center;
+  height: 100%;
+  color: var(--text-muted);
+  font-size: 13px;
+  font-style: italic;
+}
+</style>

frontend/src/features/reasoning/ui/ImportTraceDialog.vue

@@ -0,0 +1,340 @@
+<template>
+  <div
+    v-if="store.importDialogOpen"
+    class="trace-modal-backdrop"
+    data-e2e="reasoning-import-modal"
+    @click.self="close"
+  >
+    <div
+      class="trace-modal"
+      role="dialog"
+      aria-modal="true"
+      :aria-label="t('reasoning.importTitle')"
+    >
+      <div class="trace-modal-header">
+        <h3>{{ t('reasoning.importTitle') }}</h3>
+        <button class="trace-modal-close" :aria-label="t('reasoning.close')" @click="close">
+          ✕
+        </button>
+      </div>
+
+      <p class="trace-modal-hint">{{ t('reasoning.importHint') }}</p>
+
+      <div
+        class="trace-drop"
+        :class="{ dragging, parsing }"
+        data-e2e="reasoning-drop-zone"
+        @dragover.prevent="dragging = true"
+        @dragleave.prevent="dragging = false"
+        @drop.prevent="onDrop"
+        @click="openFilePicker"
+      >
+        <input
+          ref="fileInput"
+          type="file"
+          accept=".json,application/json"
+          hidden
+          @change="onFileSelect"
+        />
+        <div v-if="parsing" class="trace-drop-state">
+          <div class="spinner" />
+          <span>{{ t('reasoning.parsing') }}</span>
+        </div>
+        <div v-else class="trace-drop-state">
+          <svg
+            viewBox="0 0 24 24"
+            fill="none"
+            stroke="currentColor"
+            stroke-width="1.5"
+            class="trace-icon"
+          >
+            <path d="M12 16V4m0 0L8 8m4-4l4 4M4 17v2a1 1 0 001 1h14a1 1 0 001-1v-2" />
+          </svg>
+          <span class="trace-drop-title">{{ t('reasoning.drop') }}</span>
+          <span class="trace-drop-sub">{{ t('reasoning.dropSub') }}</span>
+        </div>
+      </div>
+
+      <details class="trace-paste">
+        <summary>{{ t('reasoning.pasteToggle') }}</summary>
+        <textarea
+          v-model="pastedJson"
+          class="trace-paste-area"
+          :placeholder="t('reasoning.pastePlaceholder')"
+          rows="6"
+          data-e2e="reasoning-paste-area"
+        />
+        <button
+          class="trace-paste-btn"
+          :disabled="!pastedJson.trim() || parsing"
+          @click="submitPasted"
+        >
+          {{ t('reasoning.pasteSubmit') }}
+        </button>
+      </details>
+
+      <div v-if="errorMsg" class="trace-modal-error" data-e2e="reasoning-import-error">
+        {{ errorMsg }}
+      </div>
+    </div>
+  </div>
+</template>
+
+<script setup lang="ts">
+import { ref } from 'vue'
+
+import { useI18n } from '../../../shared/i18n'
+import { useReasoningStore } from '../store'
+import { parseImportedTrace } from '../store'
+
+const store = useReasoningStore()
+const { t } = useI18n()
+
+const fileInput = ref<HTMLInputElement | null>(null)
+const dragging = ref(false)
+const parsing = ref(false)
+const pastedJson = ref('')
+const errorMsg = ref<string | null>(null)
+
+function close(): void {
+  store.closeImportDialog()
+  errorMsg.value = null
+  pastedJson.value = ''
+}
+
+function openFilePicker(): void {
+  fileInput.value?.click()
+}
+
+async function readFileAsText(file: File): Promise<string> {
+  return new Promise((resolve, reject) => {
+    const reader = new FileReader()
+    reader.onload = () => resolve(String(reader.result ?? ''))
+    reader.onerror = () => reject(reader.error ?? new Error('read failed'))
+    reader.readAsText(file)
+  })
+}
+
+function ingest(rawText: string): boolean {
+  let raw: unknown
+  try {
+    raw = JSON.parse(rawText)
+  } catch (e) {
+    errorMsg.value = t('reasoning.errJson').replace('{msg}', (e as Error).message)
+    return false
+  }
+  const parsed = parseImportedTrace(raw)
+  if (!parsed) {
+    errorMsg.value = t('reasoning.errShape')
+    return false
+  }
+  errorMsg.value = null
+  store.setResult(parsed.result, parsed.envelope)
+  close()
+  return true
+}
+
+async function handleFile(file: File): Promise<void> {
+  errorMsg.value = null
+  parsing.value = true
+  try {
+    const text = await readFileAsText(file)
+    ingest(text)
+  } catch (e) {
+    errorMsg.value = (e as Error).message
+  } finally {
+    parsing.value = false
+  }
+}
+
+async function onFileSelect(e: Event): Promise<void> {
+  const target = e.target as HTMLInputElement
+  const file = target.files?.[0]
+  if (file) await handleFile(file)
+  target.value = ''
+}
+
+async function onDrop(e: DragEvent): Promise<void> {
+  dragging.value = false
+  const file = e.dataTransfer?.files?.[0]
+  if (file) await handleFile(file)
+}
+
+function submitPasted(): void {
+  if (!pastedJson.value.trim()) return
+  parsing.value = true
+  try {
+    ingest(pastedJson.value)
+  } finally {
+    parsing.value = false
+  }
+}
+</script>
+
+<style scoped>
+.trace-modal-backdrop {
+  position: fixed;
+  inset: 0;
+  background: rgba(15, 23, 42, 0.55);
+  display: flex;
+  align-items: center;
+  justify-content: center;
+  z-index: 1000;
+  padding: 16px;
+}
+
+.trace-modal {
+  background: var(--bg);
+  border: 1px solid var(--border);
+  border-radius: var(--radius);
+  padding: 20px;
+  width: min(560px, 100%);
+  max-height: 90vh;
+  overflow-y: auto;
+  box-shadow: 0 12px 48px rgba(15, 23, 42, 0.25);
+}
+
+.trace-modal-header {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  margin-bottom: 4px;
+}
+
+.trace-modal-header h3 {
+  margin: 0;
+  font-size: 16px;
+  font-weight: 600;
+  color: var(--text);
+}
+
+.trace-modal-close {
+  background: transparent;
+  border: 0;
+  color: var(--text-muted);
+  font-size: 16px;
+  cursor: pointer;
+  padding: 4px 8px;
+  border-radius: var(--radius-sm);
+}
+
+.trace-modal-close:hover {
+  background: var(--border-light);
+  color: var(--text);
+}
+
+.trace-modal-hint {
+  font-size: 13px;
+  color: var(--text-muted);
+  margin: 4px 0 16px;
+}
+
+.trace-drop {
+  border: 2px dashed var(--border-light);
+  border-radius: var(--radius);
+  padding: 28px 16px;
+  text-align: center;
+  cursor: pointer;
+  transition: all var(--transition);
+}
+
+.trace-drop:hover,
+.trace-drop.dragging {
+  border-color: var(--accent);
+  background: var(--accent-muted);
+}
+
+.trace-drop.parsing {
+  pointer-events: none;
+  opacity: 0.7;
+}
+
+.trace-drop-state {
+  display: flex;
+  flex-direction: column;
+  align-items: center;
+  gap: 8px;
+}
+
+.trace-icon {
+  width: 32px;
+  height: 32px;
+  color: var(--text-muted);
+}
+
+.trace-drop-title {
+  font-size: 14px;
+  font-weight: 500;
+  color: var(--text-secondary);
+}
+
+.trace-drop-sub {
+  font-size: 12px;
+  color: var(--text-muted);
+}
+
+.trace-paste {
+  margin-top: 12px;
+}
+
+.trace-paste summary {
+  cursor: pointer;
+  font-size: 13px;
+  color: var(--text-secondary);
+  padding: 4px 0;
+}
+
+.trace-paste-area {
+  display: block;
+  width: 100%;
+  margin-top: 8px;
+  padding: 8px;
+  font-family: 'IBM Plex Mono', monospace;
+  font-size: 11px;
+  border: 1px solid var(--border);
+  border-radius: var(--radius-sm);
+  background: var(--bg);
+  color: var(--text);
+  resize: vertical;
+}
+
+.trace-paste-btn {
+  margin-top: 8px;
+  background: var(--accent);
+  color: white;
+  border: none;
+  padding: 6px 16px;
+  border-radius: var(--radius-sm);
+  cursor: pointer;
+  font-size: 12px;
+}
+
+.trace-paste-btn:disabled {
+  opacity: 0.5;
+  cursor: not-allowed;
+}
+
+.trace-modal-error {
+  margin-top: 12px;
+  padding: 8px 12px;
+  border-radius: var(--radius-sm);
+  background: rgba(220, 38, 38, 0.08);
+  color: var(--error, #dc2626);
+  font-size: 12px;
+}
+
+.spinner {
+  width: 24px;
+  height: 24px;
+  border: 2px solid var(--border-light);
+  border-top-color: var(--accent);
+  border-radius: 50%;
+  animation: spin 0.6s linear infinite;
+}
+
+@keyframes spin {
+  to {
+    transform: rotate(360deg);
+  }
+}
+</style>

frontend/src/features/reasoning/ui/IterationCard.vue

@@ -0,0 +1,173 @@
+<template>
+  <button
+    type="button"
+    class="it-card"
+    :class="{ active, missing: !iteration.present, converged: iteration.canAnswer }"
+    :data-e2e="`reasoning-iteration-${iteration.iteration}`"
+    @click="$emit('focus', iteration.iteration)"
+  >
+    <div class="it-row">
+      <span class="it-badge">{{ iteration.iteration }}</span>
+      <span class="it-ref" :title="iteration.sectionRef">{{ iteration.sectionRef }}</span>
+      <span
+        class="it-status"
+        :class="{
+          ok: iteration.canAnswer,
+          more: !iteration.canAnswer && iteration.present,
+          missing: !iteration.present,
+        }"
+      >
+        {{ statusLabel }}
+      </span>
+    </div>
+    <p v-if="iteration.reason" class="it-reason" :class="{ placeholder: isPlaceholderReason }">
+      {{ isPlaceholderReason ? t('reasoning.reasonPlaceholder') : iteration.reason }}
+    </p>
+    <div class="it-meta">
+      <span v-if="iteration.sectionTextLength">
+        {{ t('reasoning.charsLabel').replace('{n}', String(iteration.sectionTextLength)) }}
+      </span>
+    </div>
+  </button>
+</template>
+
+<script setup lang="ts">
+import { computed } from 'vue'
+
+import { useI18n } from '../../../shared/i18n'
+import type { ResolvedIteration } from '../types'
+
+const props = defineProps<{
+  iteration: ResolvedIteration
+  active: boolean
+}>()
+
+defineEmits<{ focus: [iteration: number] }>()
+
+const { t } = useI18n()
+
+const statusLabel = computed(() => {
+  if (!props.iteration.present) return t('reasoning.statusMissing')
+  if (props.iteration.canAnswer) return t('reasoning.statusAnswered')
+  return t('reasoning.statusMore')
+})
+
+// docling-agent emits the literal string "fallback" for `reason` when its
+// `select_from_failure` branch runs (the model's structured output didn't
+// parse N times in a row). Don't show that noise — render a dash-style
+// placeholder the user can visually skip.
+const isPlaceholderReason = computed(() => {
+  const r = (props.iteration.reason || '').trim().toLowerCase()
+  return r === '' || r === 'fallback'
+})
+</script>
+
+<style scoped>
+.it-card {
+  display: block;
+  width: 100%;
+  text-align: left;
+  background: var(--bg);
+  border: 1px solid var(--border);
+  border-radius: var(--radius-sm);
+  padding: 10px 12px;
+  cursor: pointer;
+  transition: all var(--transition);
+  font-family: inherit;
+  color: inherit;
+}
+
+.it-card:hover {
+  border-color: var(--accent);
+  background: var(--accent-muted);
+}
+
+.it-card.active {
+  border-color: #ea580c;
+  box-shadow: 0 0 0 2px rgba(234, 88, 12, 0.2);
+  background: rgba(234, 88, 12, 0.06);
+}
+
+.it-card.missing {
+  opacity: 0.6;
+}
+
+.it-row {
+  display: flex;
+  align-items: center;
+  gap: 8px;
+  font-size: 12px;
+}
+
+.it-badge {
+  flex: 0 0 auto;
+  display: inline-flex;
+  align-items: center;
+  justify-content: center;
+  width: 22px;
+  height: 22px;
+  border-radius: 50%;
+  background: #ea580c;
+  color: #ffffff;
+  font-weight: 700;
+  font-size: 11px;
+}
+
+.it-card.missing .it-badge {
+  background: var(--text-muted);
+}
+
+.it-ref {
+  font-family: 'IBM Plex Mono', monospace;
+  font-size: 11px;
+  color: var(--text-secondary);
+  overflow: hidden;
+  text-overflow: ellipsis;
+  white-space: nowrap;
+  flex: 1 1 auto;
+}
+
+.it-status {
+  flex: 0 0 auto;
+  font-size: 10px;
+  font-weight: 600;
+  padding: 2px 8px;
+  border-radius: 10px;
+  text-transform: uppercase;
+  letter-spacing: 0.5px;
+}
+
+.it-status.ok {
+  background: rgba(22, 163, 74, 0.15);
+  color: #15803d;
+}
+
+.it-status.more {
+  background: rgba(234, 179, 8, 0.15);
+  color: #a16207;
+}
+
+.it-status.missing {
+  background: var(--border-light);
+  color: var(--text-muted);
+}
+
+.it-reason {
+  margin: 8px 0 0;
+  font-size: 12px;
+  line-height: 1.45;
+  color: var(--text);
+}
+
+.it-reason.placeholder {
+  color: var(--text-muted);
+  font-style: italic;
+}
+
+.it-meta {
+  margin-top: 6px;
+  font-size: 10px;
+  color: var(--text-muted);
+  font-family: 'IBM Plex Mono', monospace;
+}
+</style>

frontend/src/features/reasoning/ui/ReasoningDocPicker.vue

@@ -0,0 +1,355 @@
+<template>
+  <div class="rp-picker">
+    <header class="rp-picker-header">
+      <h1 class="rp-picker-title">{{ t('reasoning.pageTitle') }}</h1>
+      <p class="rp-picker-subtitle">{{ t('reasoning.pageSubtitle') }}</p>
+    </header>
+
+    <section class="rp-picker-upload" data-e2e="reasoning-upload">
+      <div
+        class="rp-drop"
+        :class="{ dragging, uploading }"
+        @dragover.prevent="dragging = true"
+        @dragleave.prevent="dragging = false"
+        @drop.prevent="onDrop"
+        @click="openFilePicker"
+      >
+        <input ref="fileInput" type="file" accept=".pdf" hidden @change="onFileSelect" />
+        <div v-if="uploading" class="rp-drop-state">
+          <div class="spinner" />
+          <span>{{ t('reasoning.uploading') }}</span>
+        </div>
+        <div v-else class="rp-drop-state">
+          <svg
+            viewBox="0 0 24 24"
+            fill="none"
+            stroke="currentColor"
+            stroke-width="1.5"
+            class="rp-upload-icon"
+          >
+            <path d="M12 16V4m0 0L8 8m4-4l4 4M4 17v2a1 1 0 001 1h14a1 1 0 001-1v-2" />
+          </svg>
+          <span class="rp-drop-title">{{ t('reasoning.dropPdf') }}</span>
+          <span class="rp-drop-sub">{{ t('reasoning.dropPdfHint') }}</span>
+        </div>
+      </div>
+      <p v-if="uploadError" class="rp-upload-error" data-e2e="reasoning-upload-error">
+        {{ uploadError }}
+      </p>
+    </section>
+
+    <section v-if="docsWithAnalysis.length > 0" class="rp-picker-list">
+      <h2 class="rp-list-title">{{ t('reasoning.existingDocs') }}</h2>
+      <div class="rp-doc-grid">
+        <button
+          v-for="doc in docsWithAnalysis"
+          :key="doc.id"
+          class="rp-doc-card"
+          :data-e2e="`reasoning-doc-${doc.id}`"
+          @click="emit('select', doc.id)"
+        >
+          <div class="rp-doc-card-icon">
+            <svg viewBox="0 0 20 20" fill="currentColor">
+              <path
+                fill-rule="evenodd"
+                d="M4 4a2 2 0 012-2h4.586A2 2 0 0112 2.586L15.414 6A2 2 0 0116 7.414V16a2 2 0 01-2 2H6a2 2 0 01-2-2V4z"
+                clip-rule="evenodd"
+              />
+            </svg>
+          </div>
+          <div class="rp-doc-card-body">
+            <div class="rp-doc-name" :title="doc.filename">{{ doc.filename }}</div>
+            <div class="rp-doc-meta">
+              <span v-if="doc.pageCount">
+                {{ t('reasoning.pagesCount').replace('{n}', String(doc.pageCount)) }}
+              </span>
+              <span class="rp-doc-meta-dot">·</span>
+              <span>{{ formatDate(doc.createdAt) }}</span>
+            </div>
+          </div>
+        </button>
+      </div>
+    </section>
+    <p v-else-if="documentStore.documents.length > 0" class="rp-empty-hint">
+      {{ t('reasoning.noAnalyzedDocs') }}
+    </p>
+  </div>
+</template>
+
+<script setup lang="ts">
+import { computed, onMounted, ref } from 'vue'
+
+import { useAnalysisStore } from '../../analysis/store'
+import { useDocumentStore } from '../../document/store'
+import { useI18n } from '../../../shared/i18n'
+
+const emit = defineEmits<{
+  select: [docId: string]
+  uploaded: [docId: string]
+}>()
+
+const documentStore = useDocumentStore()
+const analysisStore = useAnalysisStore()
+const { t } = useI18n()
+
+const fileInput = ref<HTMLInputElement | null>(null)
+const dragging = ref(false)
+const uploading = ref(false)
+const uploadError = ref<string | null>(null)
+
+// Docs that have at least one analysis with document_json — the graph can
+// be primed for them without a fresh Docling run. Others need analysis
+// first (handled by the upload path or the workspace's silent analyze).
+const docsWithAnalysis = computed(() => {
+  const analyzedDocIds = new Set(
+    analysisStore.analyses
+      .filter((a) => a.hasDocumentJson && a.status === 'COMPLETED')
+      .map((a) => a.documentId),
+  )
+  return documentStore.documents
+    .filter((d) => analyzedDocIds.has(d.id))
+    .sort((a, b) => (b.createdAt ?? '').localeCompare(a.createdAt ?? ''))
+})
+
+onMounted(async () => {
+  // Fetch both lists in parallel — pickers without prior state need them.
+  await Promise.all([
+    documentStore.documents.length ? Promise.resolve() : documentStore.load(),
+    analysisStore.analyses.length ? Promise.resolve() : analysisStore.load(),
+  ])
+})
+
+function formatDate(iso: string): string {
+  try {
+    return new Date(iso).toLocaleDateString()
+  } catch {
+    return iso
+  }
+}
+
+function openFilePicker(): void {
+  fileInput.value?.click()
+}
+
+function isPdf(file: File): boolean {
+  return file.type === 'application/pdf' || file.name.toLowerCase().endsWith('.pdf')
+}
+
+async function handleFile(file: File): Promise<void> {
+  uploadError.value = null
+  if (!isPdf(file)) {
+    uploadError.value = t('upload.invalidFormat')
+    return
+  }
+  uploading.value = true
+  try {
+    const doc = await documentStore.upload(file)
+    if (doc) emit('uploaded', doc.id)
+  } catch (e) {
+    uploadError.value = (e as Error).message || t('upload.uploading')
+  } finally {
+    uploading.value = false
+  }
+}
+
+async function onFileSelect(e: Event): Promise<void> {
+  const target = e.target as HTMLInputElement
+  const file = target.files?.[0]
+  if (file) await handleFile(file)
+  target.value = ''
+}
+
+async function onDrop(e: DragEvent): Promise<void> {
+  dragging.value = false
+  const file = e.dataTransfer?.files?.[0]
+  if (file) await handleFile(file)
+}
+</script>
+
+<style scoped>
+.rp-picker {
+  display: flex;
+  flex-direction: column;
+  gap: 32px;
+  padding: 48px max(24px, 6vw);
+  max-width: 960px;
+  margin: 0 auto;
+  width: 100%;
+}
+
+.rp-picker-header {
+  text-align: center;
+}
+
+.rp-picker-title {
+  margin: 0 0 8px;
+  font-size: 28px;
+  font-weight: 700;
+  color: var(--text);
+}
+
+.rp-picker-subtitle {
+  margin: 0;
+  font-size: 14px;
+  color: var(--text-muted);
+}
+
+.rp-picker-upload {
+  display: flex;
+  flex-direction: column;
+  gap: 8px;
+}
+
+.rp-drop {
+  border: 2px dashed var(--border-light);
+  border-radius: var(--radius);
+  padding: 40px 16px;
+  text-align: center;
+  cursor: pointer;
+  transition: all var(--transition);
+}
+
+.rp-drop:hover,
+.rp-drop.dragging {
+  border-color: var(--accent);
+  background: var(--accent-muted);
+}
+
+.rp-drop.uploading {
+  pointer-events: none;
+  opacity: 0.7;
+}
+
+.rp-drop-state {
+  display: flex;
+  flex-direction: column;
+  align-items: center;
+  gap: 10px;
+}
+
+.rp-upload-icon {
+  width: 40px;
+  height: 40px;
+  color: var(--text-muted);
+}
+
+.rp-drop-title {
+  font-size: 14px;
+  font-weight: 500;
+  color: var(--text-secondary);
+}
+
+.rp-drop-sub {
+  font-size: 12px;
+  color: var(--text-muted);
+}
+
+.rp-upload-error {
+  margin: 0;
+  padding: 8px 12px;
+  font-size: 12px;
+  color: var(--error, #dc2626);
+  background: rgba(220, 38, 38, 0.08);
+  border-radius: var(--radius-sm);
+}
+
+.rp-list-title {
+  margin: 0 0 16px;
+  font-size: 13px;
+  text-transform: uppercase;
+  letter-spacing: 0.5px;
+  color: var(--text-muted);
+  font-weight: 600;
+}
+
+.rp-doc-grid {
+  display: grid;
+  grid-template-columns: repeat(auto-fill, minmax(280px, 1fr));
+  gap: 12px;
+}
+
+.rp-doc-card {
+  display: flex;
+  align-items: center;
+  gap: 12px;
+  padding: 14px;
+  background: var(--bg);
+  border: 1px solid var(--border);
+  border-radius: var(--radius-sm);
+  cursor: pointer;
+  text-align: left;
+  font-family: inherit;
+  color: inherit;
+  transition: all var(--transition);
+}
+
+.rp-doc-card:hover {
+  border-color: var(--accent);
+  background: var(--accent-muted);
+  transform: translateY(-1px);
+}
+
+.rp-doc-card-icon {
+  flex: 0 0 auto;
+  width: 36px;
+  height: 36px;
+  display: flex;
+  align-items: center;
+  justify-content: center;
+  background: var(--border-light);
+  border-radius: var(--radius-sm);
+  color: var(--text-muted);
+}
+
+.rp-doc-card-icon svg {
+  width: 20px;
+  height: 20px;
+}
+
+.rp-doc-card-body {
+  flex: 1 1 auto;
+  min-width: 0;
+}
+
+.rp-doc-name {
+  font-size: 13px;
+  font-weight: 500;
+  color: var(--text);
+  overflow: hidden;
+  text-overflow: ellipsis;
+  white-space: nowrap;
+}
+
+.rp-doc-meta {
+  margin-top: 2px;
+  font-size: 11px;
+  color: var(--text-muted);
+  font-family: 'IBM Plex Mono', monospace;
+}
+
+.rp-doc-meta-dot {
+  margin: 0 4px;
+}
+
+.rp-empty-hint {
+  font-size: 12px;
+  color: var(--text-muted);
+  font-style: italic;
+  text-align: center;
+}
+
+.spinner {
+  width: 28px;
+  height: 28px;
+  border: 2px solid var(--border-light);
+  border-top-color: var(--accent);
+  border-radius: 50%;
+  animation: spin 0.6s linear infinite;
+}
+
+@keyframes spin {
+  to {
+    transform: rotate(360deg);
+  }
+}
+</style>

frontend/src/features/reasoning/ui/ReasoningPanel.vue

@@ -0,0 +1,485 @@
+<template>
+  <aside
+    v-if="store.hasTrace || store.importDialogOpen"
+    class="reasoning-panel"
+    data-e2e="reasoning-panel"
+  >
+    <header class="rp-header">
+      <h3>{{ t('reasoning.panelTitle') }}</h3>
+      <div class="rp-header-actions">
+        <button
+          class="rp-btn-ghost rp-btn-toggle"
+          :class="{ active: store.focusMode }"
+          :aria-pressed="store.focusMode"
+          data-e2e="reasoning-focus-toggle"
+          :title="t('reasoning.focusHint')"
+          @click="store.toggleFocusMode()"
+        >
+          <span class="rp-dot" />
+          {{ t('reasoning.focus') }}
+        </button>
+        <button class="rp-btn-ghost" @click="store.openImportDialog()">
+          {{ t('reasoning.reimport') }}
+        </button>
+        <button class="rp-btn-ghost" @click="onClear">{{ t('reasoning.clear') }}</button>
+      </div>
+    </header>
+
+    <section v-if="envelope" class="rp-meta">
+      <div v-if="envelope.query" class="rp-query">
+        <span class="rp-meta-label">{{ t('reasoning.query') }}</span>
+        <span class="rp-meta-value">{{ envelope.query }}</span>
+      </div>
+      <div class="rp-meta-row">
+        <span v-if="envelope.filename" class="rp-meta-chip">{{ envelope.filename }}</span>
+        <span v-if="envelope.model?.ollama_name" class="rp-meta-chip">
+          {{ envelope.model.ollama_name }}
+        </span>
+      </div>
+    </section>
+
+    <section v-if="result" class="rp-answer">
+      <div class="rp-answer-header">
+        <span class="rp-answer-label">{{ t('reasoning.answerLabel') }}</span>
+        <span class="rp-answer-actions">
+          <span class="rp-converged" :class="{ yes: result.converged, no: !result.converged }">
+            {{ result.converged ? t('reasoning.converged') : t('reasoning.notConverged') }}
+          </span>
+          <button
+            class="rp-copy-btn"
+            :title="t('reasoning.copyAnswer')"
+            data-e2e="reasoning-copy-answer"
+            @click="copyAnswer"
+          >
+            {{ copied ? t('reasoning.copied') : t('reasoning.copy') }}
+          </button>
+        </span>
+      </div>
+      <!-- eslint-disable-next-line vue/no-v-html -- sanitized by DOMPurify -->
+      <div class="rp-answer-body markdown-body" v-html="renderedAnswer" />
+      <div class="rp-answer-footer">
+        <span class="rp-stats">
+          {{ store.presentCount }} / {{ store.iterations.length }} {{ t('reasoning.resolved') }}
+        </span>
+      </div>
+    </section>
+
+    <section v-if="store.missingCount > 0" class="rp-warn" data-e2e="reasoning-missing-warn">
+      {{ missingWarning }}
+    </section>
+
+    <section class="rp-iterations">
+      <h4 class="rp-section-title">{{ t('reasoning.iterationsTitle') }}</h4>
+      <div v-if="store.iterations.length === 0" class="rp-empty">
+        {{ t('reasoning.noIterations') }}
+      </div>
+      <div v-else class="rp-iteration-list">
+        <IterationCard
+          v-for="it in store.iterations"
+          :key="it.iteration"
+          :iteration="it"
+          :active="store.activeIteration === it.iteration"
+          @focus="(n) => emit('iterationFocus', n)"
+        />
+      </div>
+    </section>
+  </aside>
+
+  <ImportTraceDialog />
+</template>
+
+<script setup lang="ts">
+import type { Core } from 'cytoscape'
+import DOMPurify from 'dompurify'
+import { marked } from 'marked'
+import { computed, ref, watch } from 'vue'
+
+import { useI18n } from '../../../shared/i18n'
+import {
+  applyReasoningOverlay,
+  buildDegradedOverlay,
+  clearReasoningOverlay,
+} from '../graphReasoningOverlay'
+import { useReasoningStore } from '../store'
+import IterationCard from './IterationCard.vue'
+import ImportTraceDialog from './ImportTraceDialog.vue'
+
+const props = defineProps<{
+  /**
+   * The live Cytoscape instance from the GraphView. May be `null` while the
+   * graph is loading or if Maintain hasn't been run for this document.
+   * Passed down from StudioPage via `graphViewRef.cy`.
+   */
+  cy: Core | null
+}>()
+
+// Iteration clicks bubble up to the workspace, which dispatches focus to
+// both the graph and the PDF directly — keeping the panel ignorant of its
+// siblings and avoiding watch-based plumbing that misfires on repeat clicks.
+const emit = defineEmits<{ iterationFocus: [iteration: number] }>()
+
+const store = useReasoningStore()
+const { t } = useI18n()
+
+const result = computed(() => store.rawResult)
+const envelope = computed(() => store.envelope)
+
+// Render the answer as markdown so numbered lists, bold, etc. render properly.
+// Models tend to produce markdown-formatted answers (numbered lists especially),
+// and plain-text `pre-wrap` made them near-unreadable.
+const renderedAnswer = computed(() => {
+  const raw = result.value?.answer ?? ''
+  if (!raw.trim()) return ''
+  return DOMPurify.sanitize(marked.parse(raw, { async: false }) as string)
+})
+
+const copied = ref(false)
+let copyResetTimer: ReturnType<typeof setTimeout> | null = null
+
+async function copyAnswer(): Promise<void> {
+  const text = result.value?.answer
+  if (!text) return
+  try {
+    await navigator.clipboard.writeText(text)
+    copied.value = true
+    if (copyResetTimer) clearTimeout(copyResetTimer)
+    copyResetTimer = setTimeout(() => {
+      copied.value = false
+    }, 1800)
+  } catch (e) {
+    console.warn('Copy failed', e)
+  }
+}
+
+const missingWarning = computed(() => {
+  // Full miss + no cy → the graph simply isn't loaded. Different message
+  // than "N sections are actually missing from the graph".
+  if (!props.cy && store.missingCount > 0 && store.presentCount === 0) {
+    return t('reasoning.graphNotLoadedWarn')
+  }
+  return t('reasoning.missingWarn').replace('{n}', String(store.missingCount))
+})
+
+function reapplyOverlay(): void {
+  if (!store.rawResult) {
+    if (props.cy) clearReasoningOverlay(props.cy)
+    store.setOverlay(null)
+    return
+  }
+  // When the Cytoscape instance is available (graph loaded for this doc) we
+  // run the full overlay: mark visited nodes, draw REASONING_NEXT arrows.
+  // Otherwise (404 on the graph endpoint, or Maintain not run yet) we still
+  // build the iteration list in "degraded" mode so the user can read the
+  // reasoning — they just won't see nodes highlighted.
+  const out = props.cy
+    ? applyReasoningOverlay(props.cy, store.rawResult, { focusMode: store.focusMode })
+    : buildDegradedOverlay(store.rawResult)
+  store.setOverlay(out)
+}
+
+// Reapply whenever cy, rawResult, or focusMode changes. This handles:
+//  - User imports trace after graph loaded (rawResult changes).
+//  - User navigates to a different doc which swaps cy (cy changes).
+//  - Graph loads AFTER the trace was already imported (cy null → non-null).
+//  - User toggles focus mode (focusMode changes) — dim in, dim out.
+//  - User clears the trace (rawResult → null → clearReasoningOverlay).
+watch(
+  () => [props.cy, store.rawResult, store.focusMode] as const,
+  () => reapplyOverlay(),
+  { immediate: true },
+)
+
+function onClear(): void {
+  if (props.cy) clearReasoningOverlay(props.cy)
+  store.reset()
+}
+</script>
+
+<style scoped>
+.reasoning-panel {
+  display: flex;
+  flex-direction: column;
+  gap: 12px;
+  width: 340px;
+  flex: 0 0 340px;
+  padding: 16px;
+  border-left: 1px solid var(--border);
+  background: var(--bg);
+  overflow-y: auto;
+  height: 100%;
+}
+
+.rp-header {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  gap: 8px;
+}
+
+.rp-header h3 {
+  margin: 0;
+  font-size: 14px;
+  font-weight: 600;
+  color: var(--text);
+}
+
+.rp-header-actions {
+  display: flex;
+  gap: 4px;
+}
+
+.rp-btn-ghost {
+  background: transparent;
+  color: var(--text-secondary);
+  border: 1px solid var(--border);
+  padding: 4px 8px;
+  font-size: 11px;
+  border-radius: var(--radius-sm);
+  cursor: pointer;
+  transition: all var(--transition);
+}
+
+.rp-btn-ghost:hover {
+  background: var(--border-light);
+  color: var(--text);
+}
+
+.rp-btn-toggle {
+  display: inline-flex;
+  align-items: center;
+  gap: 6px;
+}
+
+.rp-btn-toggle .rp-dot {
+  width: 8px;
+  height: 8px;
+  border-radius: 50%;
+  background: var(--border);
+  transition: background var(--transition);
+}
+
+.rp-btn-toggle.active {
+  border-color: #ea580c;
+  color: #ea580c;
+  background: rgba(234, 88, 12, 0.08);
+}
+
+.rp-btn-toggle.active .rp-dot {
+  background: #ea580c;
+}
+
+.rp-meta {
+  display: flex;
+  flex-direction: column;
+  gap: 6px;
+  font-size: 12px;
+  padding-bottom: 8px;
+  border-bottom: 1px solid var(--border-light);
+}
+
+.rp-meta-label {
+  font-size: 10px;
+  font-weight: 600;
+  text-transform: uppercase;
+  letter-spacing: 0.5px;
+  color: var(--text-muted);
+  display: block;
+  margin-bottom: 2px;
+}
+
+.rp-meta-value {
+  font-size: 12px;
+  color: var(--text);
+  line-height: 1.4;
+}
+
+.rp-meta-row {
+  display: flex;
+  flex-wrap: wrap;
+  gap: 6px;
+}
+
+.rp-meta-chip {
+  display: inline-block;
+  padding: 2px 8px;
+  border-radius: 10px;
+  background: var(--border-light);
+  color: var(--text-secondary);
+  font-size: 10px;
+  font-family: 'IBM Plex Mono', monospace;
+}
+
+.rp-answer {
+  display: flex;
+  flex-direction: column;
+  gap: 10px;
+  padding: 14px 16px;
+  background: var(--bg);
+  border: 1px solid #ea580c;
+  border-radius: var(--radius);
+  box-shadow: 0 1px 3px rgba(234, 88, 12, 0.08);
+}
+
+.rp-answer-header {
+  display: flex;
+  justify-content: space-between;
+  align-items: center;
+  gap: 8px;
+}
+
+.rp-answer-label {
+  font-size: 10px;
+  font-weight: 700;
+  letter-spacing: 0.8px;
+  text-transform: uppercase;
+  color: #ea580c;
+}
+
+.rp-answer-actions {
+  display: inline-flex;
+  align-items: center;
+  gap: 8px;
+}
+
+.rp-converged {
+  font-size: 10px;
+  font-weight: 600;
+  padding: 2px 8px;
+  border-radius: 10px;
+  text-transform: uppercase;
+  letter-spacing: 0.5px;
+}
+
+.rp-converged.yes {
+  background: rgba(22, 163, 74, 0.15);
+  color: #15803d;
+}
+
+.rp-converged.no {
+  background: rgba(234, 179, 8, 0.15);
+  color: #a16207;
+}
+
+.rp-copy-btn {
+  background: transparent;
+  border: 1px solid var(--border);
+  color: var(--text-secondary);
+  padding: 2px 8px;
+  font-size: 10px;
+  border-radius: var(--radius-sm);
+  cursor: pointer;
+  transition: all var(--transition);
+}
+
+.rp-copy-btn:hover {
+  background: var(--border-light);
+  color: var(--text);
+}
+
+.rp-stats {
+  font-size: 10px;
+  color: var(--text-muted);
+  font-family: 'IBM Plex Mono', monospace;
+}
+
+.rp-answer-footer {
+  display: flex;
+  justify-content: flex-end;
+  border-top: 1px solid var(--border-light);
+  padding-top: 6px;
+}
+
+/* Markdown-rendered answer body. Mirrors a subset of MarkdownViewer styles,
+ * tuned for a narrow right-rail context (tighter sizes than the full viewer). */
+.rp-answer-body {
+  font-size: 13.5px;
+  line-height: 1.6;
+  color: var(--text);
+}
+
+.rp-answer-body :deep(p) {
+  margin: 0 0 8px;
+}
+
+.rp-answer-body :deep(p:last-child) {
+  margin-bottom: 0;
+}
+
+.rp-answer-body :deep(ol),
+.rp-answer-body :deep(ul) {
+  margin: 4px 0 8px;
+  padding-left: 22px;
+}
+
+.rp-answer-body :deep(li) {
+  margin: 2px 0;
+}
+
+.rp-answer-body :deep(strong) {
+  color: var(--text);
+  font-weight: 600;
+}
+
+.rp-answer-body :deep(code) {
+  font-family: 'IBM Plex Mono', monospace;
+  font-size: 12px;
+  background: var(--border-light);
+  padding: 1px 5px;
+  border-radius: 3px;
+}
+
+.rp-answer-body :deep(pre) {
+  font-family: 'IBM Plex Mono', monospace;
+  font-size: 12px;
+  background: var(--border-light);
+  padding: 8px 10px;
+  border-radius: var(--radius-sm);
+  overflow-x: auto;
+  margin: 6px 0;
+}
+
+.rp-answer-body :deep(h1),
+.rp-answer-body :deep(h2),
+.rp-answer-body :deep(h3),
+.rp-answer-body :deep(h4) {
+  margin: 10px 0 4px;
+  font-size: 14px;
+  font-weight: 600;
+  color: var(--text);
+}
+
+.rp-answer-body :deep(a) {
+  color: #ea580c;
+  text-decoration: underline;
+}
+
+.rp-warn {
+  padding: 8px 10px;
+  background: rgba(234, 179, 8, 0.1);
+  border: 1px solid rgba(234, 179, 8, 0.3);
+  border-radius: var(--radius-sm);
+  color: #a16207;
+  font-size: 12px;
+}
+
+.rp-section-title {
+  margin: 0 0 8px;
+  font-size: 11px;
+  text-transform: uppercase;
+  letter-spacing: 0.5px;
+  color: var(--text-muted);
+  font-weight: 600;
+}
+
+.rp-iteration-list {
+  display: flex;
+  flex-direction: column;
+  gap: 8px;
+}
+
+.rp-empty {
+  font-size: 12px;
+  color: var(--text-muted);
+  font-style: italic;
+}
+</style>

frontend/src/features/reasoning/ui/ReasoningWorkspace.vue

@@ -0,0 +1,283 @@
+<template>
+  <div class="rw-root" data-e2e="reasoning-workspace">
+    <header class="rw-topbar">
+      <button class="rw-back-btn" data-e2e="reasoning-back" @click="emit('back')">
+        ← {{ t('reasoning.changeDoc') }}
+      </button>
+      <div class="rw-doc-title" :title="docFilename ?? docId">
+        {{ docFilename ?? docId }}
+      </div>
+
+      <!-- Main-pane toggle: graph vs docling markdown. Overlay panel on the
+           right stays mounted in both modes so the iteration list keeps
+           living context — only the main pane swaps. -->
+      <div class="rw-mode-switch" role="tablist" :aria-label="t('reasoning.modeSwitchLabel')">
+        <button
+          type="button"
+          role="tab"
+          class="rw-mode-btn"
+          :class="{ active: mode === 'graph' }"
+          :aria-selected="mode === 'graph'"
+          data-e2e="reasoning-mode-graph"
+          @click="mode = 'graph'"
+        >
+          {{ t('reasoning.modeGraph') }}
+        </button>
+        <button
+          type="button"
+          role="tab"
+          class="rw-mode-btn"
+          :class="{ active: mode === 'document' }"
+          :aria-selected="mode === 'document'"
+          data-e2e="reasoning-mode-document"
+          @click="mode = 'document'"
+        >
+          {{ t('reasoning.modeDocument') }}
+        </button>
+      </div>
+
+      <button
+        class="rw-action-btn rw-action-ghost"
+        data-e2e="reasoning-workspace-import"
+        @click="reasoningStore.openImportDialog()"
+      >
+        {{ t('reasoning.importBtn') }}
+      </button>
+      <button
+        class="rw-action-btn"
+        data-e2e="reasoning-workspace-run"
+        @click="reasoningStore.openRunDialog()"
+      >
+        {{ t('reasoning.runBtn') }}
+      </button>
+    </header>
+
+    <div class="rw-body">
+      <!-- Keep GraphView mounted via v-show rather than v-if so the Cytoscape
+           instance + its layout state survive a toggle to document mode and
+           back — rebuilding is expensive and would reset pan/zoom. -->
+      <GraphView
+        v-show="mode === 'graph'"
+        ref="graphViewRef"
+        :doc-id="docId"
+        :fetcher="fetchReasoningGraph"
+        @node-focus="onGraphNodeFocus"
+      />
+      <!-- v-show (not v-if) so the StructureViewer's scroll-to-focused watch
+           sees transitions from null → sectionRef that happen while we're in
+           graph mode. Otherwise the viewer mounts with an already-set prop
+           and the initial scroll never fires. -->
+      <DocumentView
+        v-show="mode === 'document'"
+        ref="documentViewRef"
+        :doc-id="docId"
+        :focused-self-ref="focusedSelfRef"
+        @element-focus="onPdfElementFocus"
+      />
+      <ReasoningPanel :cy="graphCy" @iteration-focus="onIterationFocus" />
+    </div>
+
+    <RunReasoningDialog :doc-id="docId" :doc-filename="docFilename" />
+  </div>
+</template>
+
+<script setup lang="ts">
+import { computed, onBeforeUnmount, ref, watch } from 'vue'
+
+import GraphView from '../../analysis/ui/GraphView.vue'
+import { useI18n } from '../../../shared/i18n'
+import { fetchReasoningGraph } from '../api'
+import { focusIteration, nodeIdForSectionRef } from '../graphReasoningOverlay'
+import { useReasoningStore } from '../store'
+import DocumentView from './DocumentView.vue'
+import ReasoningPanel from './ReasoningPanel.vue'
+import RunReasoningDialog from './RunReasoningDialog.vue'
+
+type WorkspaceMode = 'graph' | 'document'
+
+const props = defineProps<{
+  docId: string
+  docFilename?: string | null
+}>()
+
+const emit = defineEmits<{ back: [] }>()
+
+const { t } = useI18n()
+const reasoningStore = useReasoningStore()
+
+const graphViewRef = ref<InstanceType<typeof GraphView> | null>(null)
+const graphCy = computed(() => graphViewRef.value?.cy ?? null)
+const documentViewRef = ref<InstanceType<typeof DocumentView> | null>(null)
+
+const mode = ref<WorkspaceMode>('graph')
+
+// Shared focused element (Docling self_ref like "#/texts/12") — the one
+// bridge between graph and PDF. Clicking a node in the graph sets this,
+// clicking a bbox in the PDF sets this. When set, both views highlight
+// the corresponding element. Persists across mode toggles so jumping from
+// Graph → Document preserves the currently-looked-at element.
+const focusedSelfRef = ref<string | null>(null)
+
+function onGraphNodeFocus(selfRef: string | null): void {
+  focusedSelfRef.value = selfRef
+}
+
+function onPdfElementFocus(selfRef: string): void {
+  focusedSelfRef.value = selfRef
+  // Mirror the selection on the graph side — if the user switches back to
+  // graph mode, they'll see the same element selected + centered.
+  graphViewRef.value?.selectBySelfRef(selfRef)
+}
+
+// Click on an iteration card in the reasoning panel. We dispatch focus to
+// both the graph and the PDF imperatively (rather than via watches on
+// `activeIteration` / `focusedSelfRef`). Watches misfire when the user
+// re-clicks the same iteration: Vue collapses synchronous "flip via null"
+// mutations and only sees the final value, equal to the previous one — so
+// nothing scrolls. Calling `focusIteration` and `scrollToFocused` directly
+// works on every click regardless of state.
+function onIterationFocus(iteration: number): void {
+  reasoningStore.setActiveIteration(iteration)
+  const hit = reasoningStore.iterations.find((i) => i.iteration === iteration)
+  if (!hit?.present || !hit.sectionRef) return
+  if (graphCy.value) focusIteration(graphCy.value, nodeIdForSectionRef(hit.sectionRef))
+  focusedSelfRef.value = hit.sectionRef
+  documentViewRef.value?.scrollToFocused(hit.sectionRef)
+}
+
+// Reset the reasoning store when switching docs — a trace imported for one
+// document is meaningless on another. The main-pane mode resets too so a
+// new doc opens on the graph (consistent default).
+watch(
+  () => props.docId,
+  () => {
+    reasoningStore.reset()
+    mode.value = 'graph'
+    focusedSelfRef.value = null
+  },
+)
+
+// Clean up so a later navigation back to the workspace starts fresh.
+onBeforeUnmount(() => reasoningStore.reset())
+</script>
+
+<style scoped>
+.rw-root {
+  display: flex;
+  flex-direction: column;
+  height: 100%;
+  overflow: hidden;
+}
+
+.rw-topbar {
+  display: flex;
+  align-items: center;
+  gap: 12px;
+  padding: 8px 12px;
+  border-bottom: 1px solid var(--border);
+  background: var(--bg);
+}
+
+.rw-back-btn {
+  background: transparent;
+  border: 1px solid var(--border);
+  color: var(--text-secondary);
+  padding: 5px 10px;
+  font-size: 12px;
+  border-radius: var(--radius-sm);
+  cursor: pointer;
+  transition: all var(--transition);
+}
+
+.rw-back-btn:hover {
+  background: var(--border-light);
+  color: var(--text);
+}
+
+.rw-doc-title {
+  flex: 1 1 auto;
+  font-size: 13px;
+  font-weight: 500;
+  color: var(--text);
+  overflow: hidden;
+  text-overflow: ellipsis;
+  white-space: nowrap;
+}
+
+.rw-action-btn {
+  background: var(--accent);
+  color: white;
+  border: 0;
+  padding: 6px 14px;
+  font-size: 12px;
+  font-weight: 500;
+  border-radius: var(--radius-sm);
+  cursor: pointer;
+  transition: all var(--transition);
+}
+
+.rw-action-btn:hover {
+  filter: brightness(0.95);
+}
+
+/* Secondary action next to the primary Run button — import is a rarer path. */
+.rw-action-ghost {
+  background: transparent;
+  color: var(--text-secondary);
+  border: 1px solid var(--border);
+}
+
+.rw-action-ghost:hover {
+  background: var(--border-light);
+  color: var(--text);
+  filter: none;
+}
+
+/* Segmented control for the main-pane mode (graph vs document). Sits
+ * between the doc title and the action buttons. */
+.rw-mode-switch {
+  display: inline-flex;
+  border: 1px solid var(--border);
+  border-radius: var(--radius-sm);
+  overflow: hidden;
+}
+
+.rw-mode-btn {
+  background: transparent;
+  border: 0;
+  padding: 5px 12px;
+  font-size: 12px;
+  color: var(--text-secondary);
+  cursor: pointer;
+  transition: all var(--transition);
+}
+
+.rw-mode-btn + .rw-mode-btn {
+  border-left: 1px solid var(--border);
+}
+
+.rw-mode-btn:hover:not(.active) {
+  background: var(--border-light);
+  color: var(--text);
+}
+
+.rw-mode-btn.active {
+  background: var(--accent);
+  color: #fff;
+  font-weight: 500;
+}
+
+.rw-body {
+  flex: 1 1 auto;
+  min-height: 0;
+  display: flex;
+  flex-direction: row;
+  overflow: hidden;
+}
+
+.rw-body > :deep(.graph-view),
+.rw-body > :deep(.rdv-root) {
+  flex: 1 1 auto;
+  min-width: 0;
+}
+</style>

frontend/src/features/reasoning/ui/RunReasoningDialog.vue

@@ -0,0 +1,296 @@
+<template>
+  <div
+    v-if="store.runDialogOpen"
+    class="run-modal-backdrop"
+    data-e2e="reasoning-run-modal"
+    @click.self="close"
+  >
+    <div class="run-modal" role="dialog" aria-modal="true" :aria-label="t('reasoning.runTitle')">
+      <div class="run-modal-header">
+        <h3>{{ t('reasoning.runTitle') }}</h3>
+        <button
+          class="run-modal-close"
+          :aria-label="t('reasoning.close')"
+          :disabled="store.running"
+          @click="close"
+        >
+          ✕
+        </button>
+      </div>
+
+      <p class="run-modal-hint">{{ t('reasoning.runHint') }}</p>
+
+      <label class="run-field">
+        <span class="run-field-label">{{ t('reasoning.runQueryLabel') }}</span>
+        <textarea
+          v-model="query"
+          class="run-field-input"
+          rows="3"
+          :placeholder="t('reasoning.runQueryPlaceholder')"
+          :disabled="store.running"
+          data-e2e="reasoning-run-query"
+        />
+      </label>
+
+      <label class="run-field">
+        <span class="run-field-label">{{ t('reasoning.runModelLabel') }}</span>
+        <input
+          v-model="modelId"
+          type="text"
+          class="run-field-input"
+          :placeholder="t('reasoning.runModelPlaceholder')"
+          :disabled="store.running"
+          data-e2e="reasoning-run-model"
+        />
+        <span class="run-field-sub">{{ t('reasoning.runModelSub') }}</span>
+      </label>
+
+      <div v-if="store.running" class="run-loading" data-e2e="reasoning-run-loading">
+        <div class="spinner" />
+        <span>{{ t('reasoning.running') }}</span>
+      </div>
+
+      <div v-if="errorMsg" class="run-modal-error" data-e2e="reasoning-run-error">
+        {{ errorMsg }}
+      </div>
+
+      <div class="run-modal-actions">
+        <button class="run-ghost" :disabled="store.running" @click="close">
+          {{ t('reasoning.cancel') }}
+        </button>
+        <button
+          class="run-primary"
+          :disabled="!query.trim() || store.running"
+          data-e2e="reasoning-run-submit"
+          @click="submit"
+        >
+          {{ t('reasoning.runSubmit') }}
+        </button>
+      </div>
+    </div>
+  </div>
+</template>
+
+<script setup lang="ts">
+import { ref } from 'vue'
+
+import { useI18n } from '../../../shared/i18n'
+import { runReasoning } from '../api'
+import { useReasoningStore } from '../store'
+import type { SidecarEnvelope } from '../types'
+
+const props = defineProps<{ docId: string; docFilename?: string | null }>()
+
+const store = useReasoningStore()
+const { t } = useI18n()
+
+const query = ref('')
+const modelId = ref('')
+const errorMsg = ref<string | null>(null)
+
+function close(): void {
+  if (store.running) return // don't let the user close mid-run
+  store.closeRunDialog()
+  errorMsg.value = null
+}
+
+async function submit(): Promise<void> {
+  const q = query.value.trim()
+  if (!q) return
+  errorMsg.value = null
+  store.setRunning(true)
+  try {
+    const result = await runReasoning(props.docId, q, modelId.value.trim() || undefined)
+    // Synthesize a sidecar-like envelope so the panel can show what was asked
+    // and which model answered, same as an imported trace.
+    const envelope: SidecarEnvelope = {
+      filename: props.docFilename ?? undefined,
+      query: q,
+      model: modelId.value.trim()
+        ? { ollama_name: modelId.value.trim(), hf_model_name: null }
+        : undefined,
+      result,
+    }
+    store.setResult(result, envelope)
+    // Keep the query for the user's reference but close the dialog.
+    store.closeRunDialog()
+  } catch (e) {
+    errorMsg.value = (e as Error).message || t('reasoning.runErrUnknown')
+  } finally {
+    store.setRunning(false)
+  }
+}
+</script>
+
+<style scoped>
+.run-modal-backdrop {
+  position: fixed;
+  inset: 0;
+  background: rgba(15, 23, 42, 0.55);
+  display: flex;
+  align-items: center;
+  justify-content: center;
+  z-index: 1000;
+  padding: 16px;
+}
+
+.run-modal {
+  background: var(--bg);
+  border: 1px solid var(--border);
+  border-radius: var(--radius);
+  padding: 20px;
+  width: min(560px, 100%);
+  max-height: 90vh;
+  overflow-y: auto;
+  box-shadow: 0 12px 48px rgba(15, 23, 42, 0.25);
+  display: flex;
+  flex-direction: column;
+  gap: 14px;
+}
+
+.run-modal-header {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+}
+
+.run-modal-header h3 {
+  margin: 0;
+  font-size: 16px;
+  font-weight: 600;
+  color: var(--text);
+}
+
+.run-modal-close {
+  background: transparent;
+  border: 0;
+  color: var(--text-muted);
+  font-size: 16px;
+  cursor: pointer;
+  padding: 4px 8px;
+  border-radius: var(--radius-sm);
+}
+
+.run-modal-close:hover:not(:disabled) {
+  background: var(--border-light);
+  color: var(--text);
+}
+
+.run-modal-close:disabled {
+  opacity: 0.4;
+  cursor: not-allowed;
+}
+
+.run-modal-hint {
+  font-size: 13px;
+  color: var(--text-muted);
+  margin: 0;
+}
+
+.run-field {
+  display: flex;
+  flex-direction: column;
+  gap: 4px;
+}
+
+.run-field-label {
+  font-size: 12px;
+  font-weight: 500;
+  color: var(--text-secondary);
+}
+
+.run-field-input {
+  width: 100%;
+  padding: 8px;
+  font-size: 13px;
+  font-family: inherit;
+  border: 1px solid var(--border);
+  border-radius: var(--radius-sm);
+  background: var(--bg);
+  color: var(--text);
+  resize: vertical;
+}
+
+.run-field-input:disabled {
+  opacity: 0.6;
+  cursor: not-allowed;
+}
+
+.run-field-sub {
+  font-size: 11px;
+  color: var(--text-muted);
+}
+
+.run-loading {
+  display: flex;
+  align-items: center;
+  gap: 10px;
+  font-size: 13px;
+  color: var(--text-secondary);
+  padding: 8px 12px;
+  background: var(--accent-muted, rgba(234, 88, 12, 0.08));
+  border-radius: var(--radius-sm);
+}
+
+.run-modal-error {
+  padding: 8px 12px;
+  border-radius: var(--radius-sm);
+  background: rgba(220, 38, 38, 0.08);
+  color: var(--error, #dc2626);
+  font-size: 12px;
+  font-family: 'IBM Plex Mono', monospace;
+  word-break: break-word;
+}
+
+.run-modal-actions {
+  display: flex;
+  justify-content: flex-end;
+  gap: 8px;
+}
+
+.run-primary {
+  background: var(--accent);
+  color: white;
+  border: 0;
+  padding: 7px 16px;
+  font-size: 13px;
+  font-weight: 500;
+  border-radius: var(--radius-sm);
+  cursor: pointer;
+}
+
+.run-primary:disabled {
+  opacity: 0.5;
+  cursor: not-allowed;
+}
+
+.run-ghost {
+  background: transparent;
+  border: 1px solid var(--border);
+  color: var(--text-secondary);
+  padding: 7px 14px;
+  font-size: 13px;
+  border-radius: var(--radius-sm);
+  cursor: pointer;
+}
+
+.run-ghost:disabled {
+  opacity: 0.5;
+  cursor: not-allowed;
+}
+
+.spinner {
+  width: 18px;
+  height: 18px;
+  border: 2px solid var(--border-light);
+  border-top-color: var(--accent);
+  border-radius: 50%;
+  animation: spin 0.6s linear infinite;
+}
+
+@keyframes spin {
+  to {
+    transform: rotate(360deg);
+  }
+}
+</style>

frontend/src/pages/ReasoningPage.vue

@@ -0,0 +1,259 @@
+<template>
+  <div v-if="phase === 'picker'" class="rp-host">
+    <ReasoningDocPicker @select="onSelect" @uploaded="onUploaded" />
+  </div>
+
+  <div v-else-if="phase === 'preparing'" class="rp-host rp-centered" data-e2e="reasoning-preparing">
+    <div class="rp-prep-card">
+      <div class="spinner-large" />
+      <div class="rp-prep-title">{{ prepTitle }}</div>
+      <div v-if="prepHint" class="rp-prep-hint">{{ prepHint }}</div>
+      <button class="rp-ghost" @click="goToPicker">{{ t('reasoning.cancel') }}</button>
+    </div>
+  </div>
+
+  <div v-else-if="phase === 'error'" class="rp-host rp-centered" data-e2e="reasoning-error">
+    <div class="rp-error-card">
+      <div class="rp-error-title">{{ t('reasoning.prepError') }}</div>
+      <p class="rp-error-msg">{{ errorMsg }}</p>
+      <div class="rp-error-actions">
+        <button class="rp-primary" @click="retry">{{ t('reasoning.retry') }}</button>
+        <button class="rp-ghost" @click="goToPicker">{{ t('reasoning.pickAnother') }}</button>
+      </div>
+    </div>
+  </div>
+
+  <div v-else-if="phase === 'ready' && currentDocId" class="rp-host">
+    <ReasoningWorkspace
+      :doc-id="currentDocId"
+      :doc-filename="currentDocFilename"
+      @back="goToPicker"
+    />
+  </div>
+</template>
+
+<script setup lang="ts">
+import { computed, onMounted, ref, watch } from 'vue'
+import { useRoute, useRouter } from 'vue-router'
+
+import { useAnalysisStore } from '../features/analysis/store'
+import { useDocumentStore } from '../features/document/store'
+import ReasoningDocPicker from '../features/reasoning/ui/ReasoningDocPicker.vue'
+import ReasoningWorkspace from '../features/reasoning/ui/ReasoningWorkspace.vue'
+import { useI18n } from '../shared/i18n'
+
+const props = defineProps<{ docId?: string }>()
+
+type Phase = 'picker' | 'preparing' | 'ready' | 'error'
+
+const { t } = useI18n()
+const route = useRoute()
+const router = useRouter()
+
+const documentStore = useDocumentStore()
+const analysisStore = useAnalysisStore()
+
+const phase = ref<Phase>('picker')
+const errorMsg = ref<string>('')
+const prepTitle = ref<string>('')
+const prepHint = ref<string | null>(null)
+
+// When navigating via `/reasoning/:docId`, the router prop populates here.
+// `route.params.docId` is the reactive source of truth.
+const currentDocId = computed<string | null>(() => {
+  const raw = props.docId ?? (route.params.docId as string | undefined)
+  return raw || null
+})
+
+const currentDocFilename = computed<string | null>(() => {
+  if (!currentDocId.value) return null
+  const doc = documentStore.documents.find((d) => d.id === currentDocId.value)
+  return doc?.filename ?? null
+})
+
+/**
+ * Drive the doc through the prepare pipeline: ensure a completed analysis
+ * exists (run one silently if not). The graph itself is built on demand from
+ * SQLite by `/api/documents/:id/reasoning-graph`, so no priming step is needed.
+ */
+async function prepareDoc(docId: string): Promise<void> {
+  phase.value = 'preparing'
+  errorMsg.value = ''
+
+  try {
+    if (documentStore.documents.length === 0) await documentStore.load()
+    if (analysisStore.analyses.length === 0) await analysisStore.load()
+
+    const completed = analysisStore.analyses.find(
+      (a) => a.documentId === docId && a.status === 'COMPLETED' && a.hasDocumentJson,
+    )
+
+    if (!completed) {
+      // Silent analysis with defaults — the tunnel doesn't expose pipeline
+      // options (see design §(b) silencieux).
+      prepTitle.value = t('reasoning.analyzing')
+      prepHint.value = t('reasoning.analyzingHint')
+      await analysisStore.run(docId)
+      await waitForAnalysisIdle()
+      const again = analysisStore.analyses.find(
+        (a) => a.documentId === docId && a.status === 'COMPLETED' && a.hasDocumentJson,
+      )
+      if (!again) {
+        throw new Error(analysisStore.error || t('reasoning.prepErrAnalysis'))
+      }
+    }
+
+    phase.value = 'ready'
+  } catch (e) {
+    errorMsg.value = (e as Error).message || t('reasoning.prepErrUnknown')
+    phase.value = 'error'
+  }
+}
+
+function waitForAnalysisIdle(timeoutMs = 10 * 60 * 1000): Promise<void> {
+  return new Promise((resolve, reject) => {
+    if (!analysisStore.running) return resolve()
+    const started = Date.now()
+    const id = window.setInterval(() => {
+      if (!analysisStore.running) {
+        window.clearInterval(id)
+        resolve()
+      } else if (Date.now() - started > timeoutMs) {
+        window.clearInterval(id)
+        reject(new Error(t('reasoning.prepErrTimeout')))
+      }
+    }, 500)
+  })
+}
+
+function onSelect(docId: string): void {
+  router.push({ name: 'reasoning-doc', params: { docId } })
+}
+
+function onUploaded(docId: string): void {
+  router.push({ name: 'reasoning-doc', params: { docId } })
+}
+
+function goToPicker(): void {
+  router.push({ name: 'reasoning' })
+}
+
+async function retry(): Promise<void> {
+  if (currentDocId.value) await prepareDoc(currentDocId.value)
+}
+
+// React to route param changes (pushes, back/forward).
+watch(
+  currentDocId,
+  (id) => {
+    if (!id) {
+      phase.value = 'picker'
+      return
+    }
+    void prepareDoc(id)
+  },
+  { immediate: false },
+)
+
+onMounted(() => {
+  if (currentDocId.value) void prepareDoc(currentDocId.value)
+  else phase.value = 'picker'
+})
+</script>
+
+<style scoped>
+.rp-host {
+  height: 100%;
+  display: flex;
+  flex-direction: column;
+  overflow: hidden;
+}
+
+.rp-centered {
+  align-items: center;
+  justify-content: center;
+  padding: 48px 24px;
+}
+
+.rp-prep-card,
+.rp-error-card {
+  display: flex;
+  flex-direction: column;
+  gap: 12px;
+  align-items: center;
+  padding: 32px 40px;
+  background: var(--bg);
+  border: 1px solid var(--border);
+  border-radius: var(--radius);
+  text-align: center;
+  max-width: 480px;
+}
+
+.rp-prep-title,
+.rp-error-title {
+  font-size: 15px;
+  font-weight: 600;
+  color: var(--text);
+}
+
+.rp-prep-hint,
+.rp-error-msg {
+  margin: 0;
+  font-size: 12px;
+  color: var(--text-muted);
+  line-height: 1.5;
+}
+
+.rp-error-msg {
+  font-family: 'IBM Plex Mono', monospace;
+  font-size: 11px;
+  padding: 8px 12px;
+  background: rgba(220, 38, 38, 0.08);
+  color: var(--error, #dc2626);
+  border-radius: var(--radius-sm);
+  width: 100%;
+  text-align: left;
+}
+
+.rp-error-actions {
+  display: flex;
+  gap: 8px;
+  margin-top: 4px;
+}
+
+.rp-primary {
+  background: var(--accent);
+  color: white;
+  border: 0;
+  padding: 6px 14px;
+  font-size: 12px;
+  font-weight: 500;
+  border-radius: var(--radius-sm);
+  cursor: pointer;
+}
+
+.rp-ghost {
+  background: transparent;
+  border: 1px solid var(--border);
+  color: var(--text-secondary);
+  padding: 6px 14px;
+  font-size: 12px;
+  border-radius: var(--radius-sm);
+  cursor: pointer;
+}
+
+.spinner-large {
+  width: 36px;
+  height: 36px;
+  border: 3px solid var(--border-light);
+  border-top-color: var(--accent);
+  border-radius: 50%;
+  animation: spin 0.8s linear infinite;
+}
+
+@keyframes spin {
+  to {
+    transform: rotate(360deg);
+  }
+}
+</style>

frontend/src/pages/StudioPage.vue

@@ -84,6 +84,27 @@
             </svg>
             {{ t('studio.ingest') }}
           </button>
+          <button
+            v-if="chunkingEnabled && ingestionEnabled && ingestionStore.available"
+            class="toggle-btn"
+            data-e2e="toggle-btn maintain-btn"
+            :class="{ active: mode === 'maintain' }"
+            @click="mode = 'maintain'"
+            :disabled="!analysisStore.currentAnalysis"
+          >
+            <svg class="toggle-icon" viewBox="0 0 20 20" fill="currentColor">
+              <path
+                d="M10 3.5a6.5 6.5 0 100 13 6.5 6.5 0 000-13zM6 10a4 4 0 118 0 4 4 0 01-8 0zm4-2a2 2 0 100 4 2 2 0 000-4z"
+              />
+              <path
+                d="M10 1v2M10 17v2M1 10h2M17 10h2M3.5 3.5l1.4 1.4M15.1 15.1l1.4 1.4M3.5 16.5l1.4-1.4M15.1 4.9l1.4-1.4"
+                stroke="currentColor"
+                stroke-width="1.5"
+                fill="none"
+              />
+            </svg>
+            {{ t('studio.maintain') }}
+          </button>
         </div>
       </div>
       <div class="topbar-actions">
@@ -475,6 +496,11 @@
             :chunk-count="analysisStore.currentChunks?.length ?? 0"
           />
         </div>
+
+        <!-- MAINTAIN MODE -->
+        <div v-if="mode === 'maintain'" class="maintain-panel">
+          <GraphView :doc-id="analysisStore.currentAnalysis?.documentId ?? null" />
+        </div>
       </div>
     </div>
   </div>
@@ -489,6 +515,7 @@ import { useIngestionStore } from '../features/ingestion/store'
 import { DocumentUpload, DocumentList } from '../features/document/index'
 import { ResultTabs } from '../features/analysis/index'
 import BboxOverlay from '../features/analysis/ui/BboxOverlay.vue'
+import GraphView from '../features/analysis/ui/GraphView.vue'
 import { ChunkPanel } from '../features/chunking'
 import { IngestPanel } from '../features/ingestion'
 import { useFeatureFlag } from '../features/feature-flags'
@@ -1404,10 +1431,11 @@ onBeforeUnmount(() => {
   padding-top: 16px;
 }
 
-/* Verify / Prepare / Ingest panels */
+/* Verify / Prepare / Ingest / Maintain panels */
 .verify-panel,
 .prepare-panel,
-.ingest-panel-wrapper {
+.ingest-panel-wrapper,
+.maintain-panel {
   height: 100%;
   overflow: hidden;
   display: flex;

frontend/src/shared/i18n.ts

@@ -11,6 +11,7 @@ const messages: Messages = {
     'nav.studio': 'Studio',
     'nav.documents': 'Documents',
     'nav.history': 'Historique',
+    'nav.reasoning': 'Raisonnement',
     'nav.settings': 'Paramètres',
     'nav.collapse': 'Réduire la barre latérale',
     'nav.expand': 'Développer la barre latérale',
@@ -81,6 +82,17 @@ const messages: Messages = {
     'results.elements': 'Éléments',
     'results.markdown': 'Markdown',
     'results.images': 'Images',
+    'results.graph': 'Graphe',
+    'results.graphLoading': 'Chargement du graphe…',
+    'results.graphEmpty': 'Pas encore de graphe pour ce document (activez Neo4j).',
+    // GraphView — node details panel & interactions
+    'graph.nodeDetails': 'Détails du nœud',
+    'graph.close': 'Fermer',
+    'graph.page': 'Page',
+    'graph.text': 'Texte',
+    'graph.provenances': 'Provenances ({n})',
+    'graph.contains': 'Contenu ({n})',
+    'results.retry': 'Réessayer',
     'results.pageOf': 'Page {current} sur {total}',
     'results.noElements': 'Aucun élément détecté sur cette page',
     'results.noImages': 'Aucune image détectée dans ce document',
@@ -110,6 +122,86 @@ const messages: Messages = {
     // Chunking
     'studio.prepare': 'Préparer',
     'studio.ingest': 'Ingérer',
+    'studio.maintain': 'Maintenir',
+    // Reasoning trace (R&D v1 — overlays a docling-agent RAGResult on the graph)
+    'reasoning.importBtn': 'Importer une trace de raisonnement',
+    'reasoning.importTitle': 'Importer une trace de raisonnement',
+    'reasoning.importHint':
+      'Dépose un JSON RAGResult produit par docling-agent (ou par le script R&D experiments/reasoning-trace).',
+    'reasoning.drop': 'Glisse un fichier .json ici',
+    'reasoning.dropSub': 'ou clique pour le choisir',
+    'reasoning.parsing': 'Analyse du fichier...',
+    'reasoning.pasteToggle': 'Coller le JSON à la place',
+    'reasoning.pastePlaceholder': "Colle ici le contenu JSON d'un RAGResult...",
+    'reasoning.pasteSubmit': 'Charger',
+    'reasoning.close': 'Fermer',
+    'reasoning.errJson': 'JSON invalide : {msg}',
+    'reasoning.errShape':
+      "Le fichier n'a pas la forme d'un RAGResult (answer, converged, iterations).",
+    'reasoning.panelTitle': 'Trace de raisonnement',
+    'reasoning.focus': 'Focus',
+    'reasoning.focusHint':
+      'Atténuer les éléments non visités pour faire ressortir le chemin de raisonnement.',
+    'reasoning.reimport': 'Réimporter',
+    'reasoning.clear': 'Effacer',
+    'reasoning.query': 'Question',
+    'reasoning.converged': 'Convergé',
+    'reasoning.notConverged': 'Itérations max atteintes',
+    'reasoning.resolved': 'sections résolues',
+    'reasoning.answerLabel': 'Réponse',
+    'reasoning.copy': 'Copier',
+    'reasoning.copied': 'Copié ✓',
+    'reasoning.copyAnswer': 'Copier la réponse dans le presse-papier',
+    'reasoning.reasonPlaceholder': '— pas de justification structurée',
+    'reasoning.missingWarn':
+      '{n} section(s) introuvable(s) dans le graphe. Le document a peut-être été re-analysé — relance « Maintenir » ou régénère la trace.',
+    'reasoning.graphNotLoadedWarn':
+      'Le graphe Neo4j de ce document n\u2019est pas chargé — les itérations sont affichées mais ne peuvent pas être positionnées sur la structure. Lance « prime_neo4j » ou re-déclenche une analyse.',
+    'reasoning.iterationsTitle': 'Itérations',
+    'reasoning.noIterations': "L'agent n'a visité aucune section (document sans en-têtes ?).",
+    'reasoning.statusAnswered': 'Répondu',
+    'reasoning.statusMore': 'Continue',
+    'reasoning.statusMissing': 'Absent',
+    'reasoning.charsLabel': '{n} caractères',
+    // Reasoning page (standalone tunnel)
+    'reasoning.pageTitle': 'Reasoning Trace',
+    'reasoning.pageSubtitle':
+      'Importe un PDF, puis dépose une trace RAGResult produite par docling-agent pour visualiser le chemin de raisonnement sur le graphe du document.',
+    'reasoning.dropPdf': 'Dépose un PDF',
+    'reasoning.dropPdfHint': 'ou clique pour en choisir un',
+    'reasoning.uploading': 'Import du document...',
+    'reasoning.existingDocs': 'Documents déjà analysés',
+    'reasoning.noAnalyzedDocs':
+      'Aucun des documents existants n\u2019a encore été analysé — lance-en un depuis Studio, ou dépose un nouveau PDF ci-dessus.',
+    'reasoning.pagesCount': '{n} pages',
+    'reasoning.changeDoc': 'Changer de document',
+    'reasoning.modeSwitchLabel': 'Mode d\u2019affichage',
+    'reasoning.modeGraph': 'Graphe',
+    'reasoning.modeDocument': 'Document',
+    'reasoning.docNoContent': 'Aucun contenu rendu disponible pour ce document.',
+    'reasoning.analyzing': 'Analyse du document...',
+    'reasoning.analyzingHint':
+      'Docling analyse le PDF avec la configuration par défaut. Cela peut prendre 1 à 3 minutes selon la taille.',
+    'reasoning.runBtn': 'Lancer le reasoning',
+    'reasoning.runTitle': 'Lancer docling-agent',
+    'reasoning.runHint':
+      'Pose une question au document. Le backend appelle docling-agent via Ollama et renvoie la trace dès que la boucle converge (20-40s).',
+    'reasoning.runQueryLabel': 'Question',
+    'reasoning.runQueryPlaceholder': 'Ex : Quelles sont les obligations du fournisseur ?',
+    'reasoning.runModelLabel': 'Modèle (optionnel)',
+    'reasoning.runModelPlaceholder': 'gpt-oss:20b',
+    'reasoning.runModelSub':
+      'Nom du modèle Ollama. Laisser vide pour utiliser le défaut serveur (RAG_MODEL_ID).',
+    'reasoning.runSubmit': 'Lancer',
+    'reasoning.running': 'docling-agent tourne... (20-40s)',
+    'reasoning.runErrUnknown': 'Erreur inconnue lors de l\u2019appel à docling-agent.',
+    'reasoning.cancel': 'Annuler',
+    'reasoning.retry': 'Réessayer',
+    'reasoning.pickAnother': 'Choisir un autre document',
+    'reasoning.prepError': 'Préparation impossible',
+    'reasoning.prepErrAnalysis': "L'analyse Docling a échoué ou n'a pas produit de document_json.",
+    'reasoning.prepErrTimeout': "L'analyse prend trop de temps — réessaye plus tard.",
+    'reasoning.prepErrUnknown': 'Erreur inconnue.',
     'chunking.settings': 'Chunking',
     'chunking.chunkerType': 'Type de chunker',
     'chunking.maxTokens': 'Tokens max',
@@ -189,6 +281,7 @@ const messages: Messages = {
     'nav.studio': 'Studio',
     'nav.documents': 'Documents',
     'nav.history': 'History',
+    'nav.reasoning': 'Reasoning',
     'nav.settings': 'Settings',
     'nav.collapse': 'Collapse sidebar',
     'nav.expand': 'Expand sidebar',
@@ -253,6 +346,17 @@ const messages: Messages = {
     'results.elements': 'Elements',
     'results.markdown': 'Markdown',
     'results.images': 'Images',
+    'results.graph': 'Graph',
+    'results.graphLoading': 'Loading graph…',
+    'results.graphEmpty': 'No graph yet for this document (enable Neo4j).',
+    // GraphView — node details panel & interactions
+    'graph.nodeDetails': 'Node details',
+    'graph.close': 'Close',
+    'graph.page': 'Page',
+    'graph.text': 'Text',
+    'graph.provenances': 'Provenances ({n})',
+    'graph.contains': 'Contents ({n})',
+    'results.retry': 'Retry',
     'results.pageOf': 'Page {current} of {total}',
     'results.noElements': 'No elements detected on this page',
     'results.noImages': 'No images detected in this document',
@@ -279,6 +383,84 @@ const messages: Messages = {
 
     'studio.prepare': 'Prepare',
     'studio.ingest': 'Ingest',
+    'studio.maintain': 'Maintain',
+    // Reasoning trace (R&D v1 — overlays a docling-agent RAGResult on the graph)
+    'reasoning.importBtn': 'Import reasoning trace',
+    'reasoning.importTitle': 'Import reasoning trace',
+    'reasoning.importHint':
+      'Drop a RAGResult JSON produced by docling-agent (or by the experiments/reasoning-trace R&D script).',
+    'reasoning.drop': 'Drop a .json file here',
+    'reasoning.dropSub': 'or click to pick one',
+    'reasoning.parsing': 'Parsing file...',
+    'reasoning.pasteToggle': 'Paste JSON instead',
+    'reasoning.pastePlaceholder': 'Paste a RAGResult JSON payload here...',
+    'reasoning.pasteSubmit': 'Load',
+    'reasoning.close': 'Close',
+    'reasoning.errJson': 'Invalid JSON: {msg}',
+    'reasoning.errShape': "File doesn't look like a RAGResult (answer, converged, iterations).",
+    'reasoning.panelTitle': 'Reasoning trace',
+    'reasoning.focus': 'Focus',
+    'reasoning.focusHint': 'Dim non-visited elements to make the reasoning path stand out.',
+    'reasoning.reimport': 'Re-import',
+    'reasoning.clear': 'Clear',
+    'reasoning.query': 'Question',
+    'reasoning.converged': 'Converged',
+    'reasoning.notConverged': 'Max iterations',
+    'reasoning.resolved': 'sections resolved',
+    'reasoning.answerLabel': 'Answer',
+    'reasoning.copy': 'Copy',
+    'reasoning.copied': 'Copied ✓',
+    'reasoning.copyAnswer': 'Copy answer to clipboard',
+    'reasoning.reasonPlaceholder': '— no structured rationale',
+    'reasoning.missingWarn':
+      '{n} section(s) missing from the graph. The document may have been re-analyzed — re-run Maintain or regenerate the trace.',
+    'reasoning.graphNotLoadedWarn':
+      "This document's Neo4j graph isn't loaded — iterations are shown but can't be positioned on the structure. Run prime_neo4j or trigger a fresh analysis.",
+    'reasoning.iterationsTitle': 'Iterations',
+    'reasoning.noIterations': 'Agent visited no section (document without headers?).',
+    'reasoning.statusAnswered': 'Answered',
+    'reasoning.statusMore': 'More needed',
+    'reasoning.statusMissing': 'Missing',
+    'reasoning.charsLabel': '{n} chars',
+    // Reasoning page (standalone tunnel)
+    'reasoning.pageTitle': 'Reasoning Trace',
+    'reasoning.pageSubtitle':
+      "Drop a PDF, then import a RAGResult trace from docling-agent to visualize the reasoning path on the document's graph.",
+    'reasoning.dropPdf': 'Drop a PDF',
+    'reasoning.dropPdfHint': 'or click to pick one',
+    'reasoning.uploading': 'Uploading document...',
+    'reasoning.existingDocs': 'Previously analyzed documents',
+    'reasoning.noAnalyzedDocs':
+      'None of your existing documents have been analyzed yet — run one from Studio, or drop a new PDF above.',
+    'reasoning.pagesCount': '{n} pages',
+    'reasoning.changeDoc': 'Change document',
+    'reasoning.modeSwitchLabel': 'View mode',
+    'reasoning.modeGraph': 'Graph',
+    'reasoning.modeDocument': 'Document',
+    'reasoning.docNoContent': 'No rendered content available for this document.',
+    'reasoning.analyzing': 'Analyzing document...',
+    'reasoning.analyzingHint':
+      'Docling is parsing the PDF with default settings. May take 1–3 minutes depending on size.',
+    'reasoning.runBtn': 'Run reasoning',
+    'reasoning.runTitle': 'Run docling-agent',
+    'reasoning.runHint':
+      'Ask a question against this document. The backend calls docling-agent over Ollama and returns the trace once the loop converges (20–40s).',
+    'reasoning.runQueryLabel': 'Question',
+    'reasoning.runQueryPlaceholder': 'e.g. What are the supplier obligations?',
+    'reasoning.runModelLabel': 'Model (optional)',
+    'reasoning.runModelPlaceholder': 'gpt-oss:20b',
+    'reasoning.runModelSub':
+      'Ollama model name. Leave empty to use the server default (RAG_MODEL_ID).',
+    'reasoning.runSubmit': 'Run',
+    'reasoning.running': 'docling-agent is thinking… (20–40s)',
+    'reasoning.runErrUnknown': 'Unknown error while calling docling-agent.',
+    'reasoning.cancel': 'Cancel',
+    'reasoning.retry': 'Retry',
+    'reasoning.pickAnother': 'Pick another document',
+    'reasoning.prepError': 'Preparation failed',
+    'reasoning.prepErrAnalysis': 'Docling analysis failed or produced no document_json.',
+    'reasoning.prepErrTimeout': 'Analysis is taking too long — try again later.',
+    'reasoning.prepErrUnknown': 'Unknown error.',
     'chunking.settings': 'Chunking',
     'chunking.chunkerType': 'Chunker type',
     'chunking.maxTokens': 'Max tokens',

frontend/src/shared/types.ts

@@ -67,6 +67,10 @@ export interface PageElement {
   bbox: [number, number, number, number]
   content: string
   level: number
+  /** Docling `self_ref` — "#/texts/12", "#/tables/3", etc. Empty string for
+   * items that don't have one (rare). Lets callers correlate a bbox with
+   * the matching graph node without fuzzy bbox matching. */
+  self_ref?: string
 }
 
 // Backend serializes with snake_case (dataclasses.asdict)

frontend/src/shared/ui/AppSidebar.vue

@@ -62,6 +62,23 @@
         <span class="nav-label">{{ t('nav.search') }}</span>
       </RouterLink>
 
+      <RouterLink
+        v-if="reasoningEnabled"
+        to="/reasoning"
+        class="nav-item"
+        data-e2e="nav-reasoning"
+        :class="{
+          active: route.name === 'reasoning' || route.name === 'reasoning-doc',
+        }"
+      >
+        <svg class="nav-icon" viewBox="0 0 20 20" fill="currentColor">
+          <path
+            d="M10 2a5 5 0 00-5 5c0 1.72.87 3.24 2.2 4.14.47.32.8.84.8 1.4V14a1 1 0 001 1h2a1 1 0 001-1v-1.46c0-.56.33-1.08.8-1.4A5 5 0 0010 2zM8 17a1 1 0 011-1h2a1 1 0 110 2H9a1 1 0 01-1-1z"
+          />
+        </svg>
+        <span class="nav-label">{{ t('nav.reasoning') }}</span>
+      </RouterLink>
+
       <RouterLink
         to="/history"
         class="nav-item"
@@ -138,6 +155,7 @@ import { useIngestionStore } from '../../features/ingestion/store'
 const featureStore = useFeatureFlagStore()
 const ingestionStore = useIngestionStore()
 const ingestionEnabled = computed(() => featureStore.isEnabled('ingestion'))
+const reasoningEnabled = computed(() => featureStore.isEnabled('reasoning'))
 const version = computed(() => featureStore.appVersion)
 const route = useRoute()
 const { t } = useI18n()

mkdocs.yml

@@ -57,7 +57,7 @@ nav:
   - Audit:
       - Audit Master: audit/master.md
       - Audits:
-          - Clean Architecture: audit/audits/01-clean-architecture.md
+          - Hexagonal Architecture: audit/audits/01-clean-architecture.md
           - DDD: audit/audits/02-ddd.md
           - Clean Code: audit/audits/03-clean-code.md
           - KISS: audit/audits/04-kiss.md