feat(neo4j): Day 1 — compose service, driver, schema bootstrap

Add Neo4j as an optional graph-native storage layer (ingestion profile). Introduces infra/neo4j with a singleton async driver wrapper and an idempotent bootstrap of constraints + indexes, wired into the FastAPI lifespan. Integration tests skip when no live Neo4j is reachable. Refs #186
2026-04-17 15:18:22 +02:00 · 2026-04-17 15:18:22 +02:00 · 712fc3f1cd
commit 712fc3f1cd
parent e0f8e81b63
15 changed files with 752 additions and 1 deletions
--- a/.env.example
+++ b/.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
--- a/docker-compose.dev.yml
+++ b/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:
--- a/docker-compose.yml
+++ b/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:
--- a/docs/design/neo4j-integration.md
+++ b/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)
+  - FastAPI endpoint `/api/documents/{doc_id}/graph` returns nodes + edges around a scope (whole doc or subtree)
+  - 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 ?**
+   - Rec: **cytoscape** — lighter, better layout API, used by Neo4j itself.
+
+4. **Graph endpoint: return full doc or paginate ?**
+   - Rec: full doc for v0.5 (reasonable cap at 200 pages). Pagination 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.*
--- a/document-parser/infra/neo4j/init.py
+++ b/document-parser/infra/neo4j/init.py
@ -0,0 +1,10 @@
+"""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.driver import Neo4jDriver, close_driver, get_driver
+from infra.neo4j.schema import bootstrap_schema
+
+__all__ = ["Neo4jDriver", "bootstrap_schema", "close_driver", "get_driver"]
--- a/document-parser/infra/neo4j/driver.py
+++ b/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")
--- a/document-parser/infra/neo4j/schema.py
+++ b/document-parser/infra/neo4j/schema.py
@ -0,0 +1,51 @@
+"""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 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)",
+)
+
+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),
+    )
--- a/document-parser/infra/settings.py
+++ b/document-parser/infra/settings.py
@ -25,6 +25,9 @@ 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"
    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"
@ -102,6 +105,9 @@ class Settings:
            batch_page_size=int(os.environ.get("BATCH_PAGE_SIZE", "10")),
            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"),
            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"),
--- a/document-parser/main.py
+++ b/document-parser/main.py
@ -93,6 +93,28 @@ def _build_analysis_service(
    )


+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 Neo4jDriver, 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() -> IngestionService | None:
    """Build the ingestion service — only if embedding + opensearch are configured."""
    if not settings.embedding_url or not settings.opensearch_url:
@ -150,8 +172,15 @@ async def lifespan(app: FastAPI) -> AsyncIterator[None]:
    if ingestion_service is not None:
        app.include_router(ingestion_router)
        logger.info("Ingestion router mounted")
+    app.state.neo4j = await _init_neo4j()
    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(
--- a/document-parser/requirements.txt
+++ b/document-parser/requirements.txt
@ -8,3 +8,4 @@ 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
--- a/document-parser/tests/neo4j/init.py
+++ b/document-parser/tests/neo4j/init.py
--- a/document-parser/tests/neo4j/conftest.py
+++ b/document-parser/tests/neo4j/conftest.py
@ -0,0 +1,36 @@
+"""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
+
+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()
--- a/document-parser/tests/neo4j/test_document_roundtrip.py
+++ b/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
--- a/document-parser/tests/neo4j/test_driver.py
+++ b/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
--- a/document-parser/tests/neo4j/test_schema.py
+++ b/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"