feat(reasoning): live docling-agent runner + UX polish

Backend — live runner
- New `POST /api/documents/:id/rag` endpoint. Loads `document_json` from
  SQLite, reconstructs the DoclingDocument, wraps the model id in
  `ModelIdentifier(ollama_name=...)`, and calls `agent._rag_loop`
  off-thread (blocking sync call). Returns a `RAGResult` in the shape
  the existing v1 import path already consumes, so the frontend overlay
  is fully reused.
- `_rag_loop` is private upstream; we call it because `run()` wraps the
  answer in a synthetic DoclingDocument and drops the iteration trace.
- Settings: `RAG_ENABLED`, `OLLAMA_HOST`, `RAG_MODEL_ID`. Router mounts
  unconditionally; handler 503s when the flag is off or deps aren't
  installed. `rag_available` surfaced in `/api/health`.
- Maps known docling-agent bugs to readable HTTP errors: 502 with
  "the model couldn't produce a parseable answer" when `_rag_loop`
  raises `IndexError` from `find_json_dicts([])[0]` after 3 + 3
  rejection-sampling retries (model-dependent).
- Tests: 11 cases (flag off, query empty, no analysis, happy path,
  model_id wrap, Ollama env, IndexError → 502, other errors → 500,
  deps missing → 503).

Backend — bug fix
- Default `BATCH_PAGE_SIZE` flipped from `10` to `0` to match the
  dataclass default. The old default silently dropped `document_json`
  (see `domain/services.merge_results`) for any doc > 10 pages, which
  broke the reasoning tunnel. Set `BATCH_PAGE_SIZE>0` explicitly on
  memory-constrained deploys if batching is wanted.

Frontend — runner UX
- `features/reasoning/api.ts:runReasoning()` — POST wrapper.
- `RunReasoningDialog.vue` — query textarea + optional model_id
  override. Blocks close while running, 20-40s loading state,
  synthesises a sidecar-shaped envelope so the panel surfaces query +
  model the same way an imported trace would.
- `ReasoningWorkspace.vue` — primary "Run reasoning" button; "Import
  trace" relegated to ghost secondary.
- Store: `runDialogOpen`, `running`, `setRunning`.

Frontend — answer polish
- Answer rendered through `marked` + DOMPurify (models emit markdown
  lists; `pre-wrap` rendered them as plain "1. …" strings).
- Dedicated answer block with orange border, "ANSWER" label, "Copy"
  button (clipboard + "Copied ✓" feedback).
- IterationCard: drop the duplicate `response` block (the main answer
  is authoritative); style reasons equal to `"fallback"` (docling-agent
  `select_from_failure` placeholder) as italic muted "— no structured
  rationale".

Frontend — node details contents
- Clicking a SectionHeader (or any node with compound children) lists
  its contained elements in `NodeDetailsPanel` under a new "Contents"
  block. Children come from the same `parentMap` used for Cytoscape
  compound parenting (explicit PARENT_OF + synthetic section scope),
  inverted once and cached as a computed.
- Click a child row → pan the viewport to it + swap the selection.

Housekeeping
- `cytoscape-navigator` removed from `package-lock.json` (follow-up
  from the minimap removal in the previous commit).

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

@@ -35,6 +35,10 @@ class HealthResponse(_CamelModel):
     max_page_count: int | None = None
     max_file_size_mb: int | None = None
     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/infra/settings.py

@@ -28,6 +28,12 @@ class Settings:
     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"
@@ -102,12 +108,20 @@ 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"),

document-parser/main.py

@@ -221,6 +221,13 @@ 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:
@@ -243,4 +250,18 @@ async def health() -> HealthResponse:
         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,
         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/requirements.txt

@@ -9,3 +9,7 @@ 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/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"]

frontend/package-lock.json

@@ -11,7 +11,6 @@
         "cytoscape": "^3.30.0",
         "cytoscape-dagre": "^2.5.0",
         "cytoscape-expand-collapse": "^4.1.1",
-        "cytoscape-navigator": "^2.0.2",
         "dompurify": "^3.3.3",
         "marked": "^17.0.4",
         "pinia": "^2.3.0",
@@ -1876,14 +1875,6 @@
         "cytoscape": "^3.3.0"
       }
     },
-    "node_modules/cytoscape-navigator": {
-      "version": "2.0.2",
-      "resolved": "https://registry.npmjs.org/cytoscape-navigator/-/cytoscape-navigator-2.0.2.tgz",
-      "integrity": "sha512-TZFBLFWEMW858UOt4rzusOjtDj7YT5vNx2uCwpUuicUYbaWCHHcUROBZWO+hiuSPWpVhvGLFlOq3NBcAVYOAgw==",
-      "peerDependencies": {
-        "cytoscape": "^2.6.0 || ^3.0.0"
-      }
-    },
     "node_modules/dagre": {
       "version": "0.8.5",
       "resolved": "https://registry.npmjs.org/dagre/-/dagre-0.8.5.tgz",

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

@@ -49,7 +49,12 @@
             {{ tooltip.text }}
           </div>
         </div>
-        <NodeDetailsPanel :node="selectedNode" @close="closeDetails" />
+        <NodeDetailsPanel
+          :node="selectedNode"
+          :contents="selectedNodeContents"
+          @close="closeDetails"
+          @navigate="navigateToNode"
+        />
       </div>
     </template>
   </div>
@@ -57,7 +62,7 @@
 
 <script setup lang="ts">
 import type { Core } from 'cytoscape'
-import { onMounted, onBeforeUnmount, ref, watch, nextTick } from 'vue'
+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'
@@ -98,6 +103,32 @@ 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) ?? []
+})
+
 // Hover tooltip: position (px within .graph-canvas) + text. Null hides it.
 const tooltip = ref<{ x: number; y: number; text: string } | null>(null)
 
@@ -182,11 +213,13 @@ async function renderGraph(): Promise<void> {
   // 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).
-  const parentMap = mergeParentMaps(
+  // 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) => ({
@@ -202,7 +235,7 @@ async function renderGraph(): Promise<void> {
         // 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: parentMap.get(n.id),
+        parent: computedParentMap.get(n.id),
         raw: n,
       },
     })),
@@ -399,6 +432,22 @@ function closeDetails(): void {
   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 })
+  }
+}
+
 function disposeGraph(): void {
   if (cy.value) {
     cy.value.destroy()

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

@@ -57,6 +57,31 @@
         </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>
 
@@ -66,8 +91,20 @@ import { computed } from 'vue'
 import { useI18n } from '../../../shared/i18n'
 import type { GraphNode, GraphProvenance } from '../graphApi'
 
-const props = defineProps<{ node: GraphNode | null }>()
-defineEmits<{ close: [] }>()
+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()
 
@@ -133,6 +170,35 @@ 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>
@@ -275,4 +341,69 @@ function fmt(n: number | null | undefined): string {
   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/reasoning/api.ts

@@ -1,5 +1,6 @@
 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
@@ -13,3 +14,28 @@ import type { GraphPayload } from '../analysis/graphApi'
 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/store.ts

@@ -39,6 +39,11 @@ function isRAGResult(x: RAGResult | undefined): boolean {
 
 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)
@@ -62,6 +67,18 @@ export const useReasoningStore = defineStore('reasoning', () => {
     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
@@ -98,12 +115,16 @@ export const useReasoningStore = defineStore('reasoning', () => {
     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,
@@ -118,6 +139,9 @@ export const useReasoningStore = defineStore('reasoning', () => {
     // actions
     openImportDialog,
     closeImportDialog,
+    openRunDialog,
+    closeRunDialog,
+    setRunning,
     setResult,
     setOverlay,
     setActiveIteration,

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

@@ -20,9 +20,8 @@
         {{ statusLabel }}
       </span>
     </div>
-    <p v-if="iteration.reason" class="it-reason">{{ iteration.reason }}</p>
-    <p v-if="iteration.response && iteration.canAnswer" class="it-response">
-      {{ iteration.response }}
+    <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">
@@ -52,6 +51,15 @@ const statusLabel = computed(() => {
   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>
@@ -151,16 +159,9 @@ const statusLabel = computed(() => {
   color: var(--text);
 }
 
-.it-response {
-  margin: 6px 0 0;
-  font-size: 12px;
-  line-height: 1.45;
-  color: var(--text-secondary);
+.it-reason.placeholder {
+  color: var(--text-muted);
   font-style: italic;
-  padding: 6px 8px;
-  border-left: 2px solid #ea580c;
-  background: rgba(234, 88, 12, 0.04);
-  border-radius: 2px;
 }
 
 .it-meta {

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

@@ -40,14 +40,28 @@
 
     <section v-if="result" class="rp-answer">
       <div class="rp-answer-header">
-        <span class="rp-converged" :class="{ yes: result.converged, no: !result.converged }">
-          {{ result.converged ? t('reasoning.converged') : t('reasoning.notConverged') }}
+        <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>
-      <p class="rp-answer-text">{{ result.answer }}</p>
     </section>
 
     <section v-if="store.missingCount > 0" class="rp-warn" data-e2e="reasoning-missing-warn">
@@ -76,7 +90,9 @@
 
 <script setup lang="ts">
 import type { Core } from 'cytoscape'
-import { computed, watch } from 'vue'
+import DOMPurify from 'dompurify'
+import { marked } from 'marked'
+import { computed, ref, watch } from 'vue'
 
 import { useI18n } from '../../../shared/i18n'
 import {
@@ -104,6 +120,34 @@ 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".
@@ -273,11 +317,12 @@ function onClear(): void {
 .rp-answer {
   display: flex;
   flex-direction: column;
-  gap: 6px;
-  padding: 10px 12px;
-  background: var(--accent-muted, rgba(234, 88, 12, 0.04));
-  border: 1px solid var(--border-light);
-  border-radius: var(--radius-sm);
+  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 {
@@ -287,6 +332,20 @@ function onClear(): void {
   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;
@@ -306,18 +365,97 @@ function onClear(): void {
   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-text {
-  margin: 0;
-  font-size: 13px;
-  line-height: 1.5;
+.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);
-  white-space: pre-wrap;
+}
+
+.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 {

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

@@ -8,18 +8,27 @@
         {{ docFilename ?? docId }}
       </div>
       <button
-        class="rw-action-btn"
+        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">
       <GraphView ref="graphViewRef" :doc-id="docId" :fetcher="fetchReasoningGraph" />
       <ReasoningPanel :cy="graphCy" />
     </div>
+
+    <RunReasoningDialog :doc-id="docId" :doc-filename="docFilename" />
   </div>
 </template>
 
@@ -31,6 +40,7 @@ import { useI18n } from '../../../shared/i18n'
 import { fetchReasoningGraph } from '../api'
 import { useReasoningStore } from '../store'
 import ReasoningPanel from './ReasoningPanel.vue'
+import RunReasoningDialog from './RunReasoningDialog.vue'
 
 const props = defineProps<{
   docId: string
@@ -115,6 +125,19 @@ onBeforeUnmount(() => reasoningStore.reset())
   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;
+}
+
 .rw-body {
   flex: 1 1 auto;
   min-height: 0;

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/shared/i18n.ts

@@ -91,6 +91,7 @@ const messages: Messages = {
     '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',
@@ -147,6 +148,11 @@ const messages: Messages = {
     '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':
@@ -172,6 +178,19 @@ const messages: Messages = {
     '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',
@@ -332,6 +351,7 @@ const messages: Messages = {
     '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',
@@ -383,6 +403,11 @@ const messages: Messages = {
     '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':
@@ -408,6 +433,19 @@ const messages: Messages = {
     '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',