fix(ocr): file page headers and footers under the furniture content layer

CLAUDE.md

@@ -62,3 +62,6 @@ tag atomically. The version lives in three files — always change it with
 `python scripts/bump_version.py set --version X.Y.Z`, never by hand.
 
 See `docs/architecture/deployment.md` for the full pipeline
+
+## Gotchas & Rules
+- **Check the library before writing a helper.** Before adding any function that renders, serializes, parses or walks a `DoclingDocument`, or splits chunks or fuses scores.

extralit-server/docker/server/.dockerignore

@@ -0,0 +1,4 @@
+# Allowlist: the context is a wheel and one script, whatever stray build output lands here.
+*
+!scripts/*.sh
+!dist/*.whl

extralit-server/docker/server/Dockerfile

@@ -1,23 +1,34 @@
 FROM python:3.12-slim AS builder
 
-# Install uv
-COPY --from=ghcr.io/astral-sh/uv:0.7.12 /uv /uvx /bin/
-# Copying extralit distribution files
-COPY dist/*.whl /packages/
-RUN python -m venv /opt/venv
-ENV PATH="/opt/venv/bin:$PATH"
-ENV MAMBA_ROOT_PREFIX=/opt/venv
-ENV CONDA_PREFIX=/opt/venv
-ENV UV_CACHE_DIR=/home/extralit/.cache/uv
+COPY --from=ghcr.io/astral-sh/uv:0.12.6 /uv /uvx /bin/
 
-RUN --mount=type=cache,target=/home/extralit/.cache/uv \
-    apt-get update && \
+ENV VIRTUAL_ENV=/opt/venv \
+    PATH="/opt/venv/bin:$PATH" \
+    UV_CACHE_DIR=/root/.cache/uv \
+    # Pay for the .pyc files once here instead of on every worker's first import.
+    UV_COMPILE_BYTECODE=1 \
+    # The cache mount is a different filesystem from /opt/venv, so hardlinking cannot work.
+    UV_LINK_MODE=copy \
+    UV_PYTHON_DOWNLOADS=0
+
+# psycopg2 is built from source. The toolchain stays in this stage — the runtime image copies
+# nothing out of it but the venv, so there is nothing to purge afterwards.
+RUN apt-get update && \
     apt-get upgrade -y && \
     apt-get install -y --no-install-recommends libc6-dev libpq-dev gcc && \
-    for wheel in /packages/*.whl; do uv pip install "$wheel"[postgresql]; done && \
-    apt-get purge -y --auto-remove libc6-dev libpq-dev gcc && \
     apt-get clean && \
-    rm -rf /var/lib/apt/lists/* /packages
+    rm -rf /var/lib/apt/lists/*
+
+# --seed, though nothing here needs pip: without it `pip` falls through to the base image's
+# /usr/local/bin/pip, which installs where /opt/venv/bin/python cannot see it. Dropping it saves
+# 5 MB of 1847 and hands any derived image a silent wrong-interpreter install.
+RUN uv venv --seed "$VIRTUAL_ENV"
+
+# The wheel is the only input that differs between two builds of the same tree, so it enters
+# last — and as a bind mount, which leaves no copy of it behind in the stage.
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=bind,source=dist,target=/packages \
+    for wheel in /packages/*.whl; do uv pip install "$wheel"[postgresql]; done
 
 FROM python:3.12-slim
 
@@ -39,21 +50,25 @@ RUN mkdir -p "$EXTRALIT_HOME_PATH" && \
     chown extralit:extralit "$EXTRALIT_HOME_PATH" && \
     apt-get update && \
     apt-get upgrade -y && \
-    apt-get install -y --no-install-recommends libpq-dev libgl1 libglib2.0-0 poppler-utils && \
+    apt-get install -y --no-install-recommends libpq-dev libgl1 libglib2.0-0 poppler-utils tesseract-ocr-eng && \
     apt-get clean && \
     rm -rf /var/lib/apt/lists/*
 VOLUME $EXTRALIT_HOME_PATH
 
-COPY scripts/start_extralit_server.sh /home/extralit
+# liteparse would otherwise download the language data on the first scanned page it meets.
+ENV TESSDATA_PREFIX=/usr/share/tesseract-ocr/5/tessdata
+RUN test -f "$TESSDATA_PREFIX/eng.traineddata"
+
+# Layers below are ordered by how often they change, because a rebuilt layer rebuilds every layer
+# under it.
+COPY --chmod=0755 scripts/start_extralit_server.sh /home/extralit/
 # Destination folder must be the same as the builder one. Otherwise installed script won't work (since the installation fixes the path inside the script)
 COPY --chown=extralit:extralit --from=builder /opt/venv /opt/venv
 
+ENV VIRTUAL_ENV=/opt/venv
 ENV PATH="/opt/venv/bin:$PATH"
-ENV MAMBA_ROOT_PREFIX=/opt/venv
-ENV CONDA_PREFIX=/opt/venv
 
 WORKDIR /home/extralit
-RUN chmod +x start_extralit_server.sh
 
 USER extralit
 

extralit-server/pyproject.toml

@@ -77,9 +77,20 @@ dependencies = [
     "lancedb>=0.37.1",
     "pylance>=10.0.0",
     "duckdb>=1.5.4",
+    # ships lance.duckdb_extension in the wheel and pins duckdb to the release it was built for
+    "duckdb-extension-lance>=1.5.5",
     "docling-core>=2.91.0,<3.0.0",
     "pyarrow>=23.0.1",
     "pdf-inspector>=1.14.2",
+    # abi3 wheels from 2.14, so the block extractor is not tied to one CPython
+    "liteparse>=2.14.0",
+    # what docling_core.transforms.chunker imports; the [chunking] extra would add transformers
+    "semchunk>=2.2",
+    "tree-sitter>=0.25,<0.27",
+    "tree-sitter-python",
+    "tree-sitter-c",
+    "tree-sitter-javascript",
+    "tree-sitter-typescript",
     "xxhash>=3.6.0",
     "obstore>=0.11.0",
 ]

extralit-server/src/extralit_server/contexts/ocr/docling_builder.py

@@ -20,7 +20,7 @@
     Size,
     TableCell,
 )
-from docling_core.types.doc.document import DocumentOrigin, NodeItem
+from docling_core.types.doc.document import DocumentOrigin, FloatingItem, NodeItem, PictureItem, TableItem
 
 PDF_MIMETYPE = "application/pdf"
 
@@ -34,6 +34,9 @@
 #: Captions and footnotes legitimately overlap their figure, so containment must not drop them.
 CONTAINMENT_EXEMPT_LABELS = frozenset({DocItemLabel.CAPTION, DocItemLabel.FOOTNOTE})
 
+#: How far either side of a caption its figure or table may sit, nearest first, before after.
+CAPTION_REACH = (-1, 1, -2, 2)
+
 
 @dataclass(frozen=True)
 class LayoutBlock:
@@ -153,6 +156,29 @@ def sort_body_by_position(doc: DoclingDocument) -> None:
     doc.body.children.sort(key=lambda ref: _sort_key(doc, ref.resolve(doc)))
 
 
+def _page_of(item: NodeItem) -> Optional[int]:
+    provs = getattr(item, "prov", None) or []
+    return provs[0].page_no if provs else None
+
+
+def link_captions(doc: DoclingDocument) -> None:
+    """Attach each unlinked caption to the nearest table or picture on its page, looking back first.
+
+    docling serializes a linked caption inside its owner and nowhere else, so this is what keeps a
+    caption out of the prose stream and inside the table or figure that it names.
+    """
+    body = [ref.resolve(doc) for ref in doc.body.children]
+    linked = {ref.cref for item in body if isinstance(item, FloatingItem) for ref in item.captions}
+    for index, item in enumerate(body):
+        if getattr(item, "label", None) != DocItemLabel.CAPTION or item.self_ref in linked:
+            continue
+        for offset in CAPTION_REACH:
+            owner = body[index + offset] if 0 <= index + offset < len(body) else None
+            if isinstance(owner, (TableItem, PictureItem)) and _page_of(owner) == _page_of(item):
+                owner.captions.append(item.get_ref())
+                break
+
+
 def append_blocks(
     doc: DoclingDocument,
     ctx: PageContext,
@@ -185,4 +211,5 @@ def append_blocks(
             added.append(item)
 
     sort_body_by_position(doc)
+    link_captions(doc)
     return added

extralit-server/src/extralit_server/contexts/ocr/text.py

@@ -4,7 +4,7 @@
 
 from typing import Optional
 
-from docling_core.types.doc import DocItemLabel, DoclingDocument
+from docling_core.types.doc import ContentLayer, DocItemLabel, DoclingDocument
 from docling_core.types.doc.document import NodeItem
 
 from extralit_server.contexts.ocr.docling_builder import (
@@ -15,6 +15,9 @@
     make_prov,
 )
 
+# The labels docling itself files as furniture; a label alone does not keep them out of chunks.
+FURNITURE_LABELS = frozenset({DocItemLabel.PAGE_HEADER, DocItemLabel.PAGE_FOOTER})
+
 
 def add_text_block(
     doc: DoclingDocument,
@@ -36,4 +39,5 @@ def add_text_block(
     if block.label == DocItemLabel.SECTION_HEADER:
         return doc.add_heading(text=text, level=block.level or 1, prov=prov, parent=parent)
 
-    return doc.add_text(label=block.label, text=text, prov=prov, parent=parent)
+    layer = ContentLayer.FURNITURE if block.label in FURNITURE_LABELS else ContentLayer.BODY
+    return doc.add_text(label=block.label, text=text, prov=prov, parent=parent, content_layer=layer)

extralit-server/tests/integration/conftest.py

@@ -10,7 +10,7 @@
 
 import pytest
 import pytest_asyncio
-from httpx import AsyncClient
+from httpx import ASGITransport, AsyncClient
 
 from extralit_server.constants import API_KEY_HEADER_NAME
 from extralit_server.database import get_async_db
@@ -39,7 +39,7 @@ async def override_get_async_db():
 
     api_v1.dependency_overrides[get_async_db] = override_get_async_db
 
-    async with AsyncClient(app=app, base_url="http://testserver") as client:
+    async with AsyncClient(transport=ASGITransport(app=app), base_url="http://testserver") as client:
         yield client
 
     # Pop only what this fixture registered. `tests/unit/conftest.py`'s async_client writes

extralit-server/tests/unit/api/handlers/v1/test_datasets.py

@@ -1,3 +1,4 @@
+import json
 import math
 import uuid
 from datetime import datetime
@@ -82,6 +83,11 @@
     from sqlalchemy.ext.asyncio import AsyncSession
 
 
+def nan_body(payload: dict, headers: dict) -> dict:
+    """Post a body the JSON spec forbids. `json=` cannot: httpx encodes with allow_nan=False."""
+    return {"content": json.dumps(payload), "headers": {**headers, "Content-Type": "application/json"}}
+
+
 @pytest.mark.asyncio
 class TestSuiteDatasets:
     async def test_list_current_user_datasets(self, async_client: "AsyncClient", owner_auth_header: dict) -> None:
@@ -1959,7 +1965,7 @@ async def test_create_dataset_records_with_metadata_nan_values(
         }
 
         response = await async_client.post(
-            f"/api/v1/datasets/{dataset.id}/records/bulk", headers=owner_auth_header, json=records_json
+            f"/api/v1/datasets/{dataset.id}/records/bulk", **nan_body(records_json, owner_auth_header)
         )
 
         assert response.status_code == 422
@@ -3027,23 +3033,25 @@ async def test_update_dataset_records_with_metadata_nan_value(
 
         response = await async_client.put(
             f"/api/v1/datasets/{dataset.id}/records/bulk",
-            headers=owner_auth_header,
-            json={
-                "items": [
-                    {
-                        "id": str(records[0].id),
-                        "metadata": {"terms": math.nan},
-                    },
-                    {
-                        "id": str(records[1].id),
-                        "metadata": {"float": math.nan},
-                    },
-                    {
-                        "id": str(records[2].id),
-                        "metadata": {"terms": "a"},
-                    },
-                ]
-            },
+            **nan_body(
+                {
+                    "items": [
+                        {
+                            "id": str(records[0].id),
+                            "metadata": {"terms": math.nan},
+                        },
+                        {
+                            "id": str(records[1].id),
+                            "metadata": {"float": math.nan},
+                        },
+                        {
+                            "id": str(records[2].id),
+                            "metadata": {"terms": "a"},
+                        },
+                    ]
+                },
+                owner_auth_header,
+            ),
         )
 
         assert response.status_code == 422

extralit-server/tests/unit/conftest.py

@@ -3,7 +3,7 @@
 
 import pytest
 import pytest_asyncio
-from httpx import AsyncClient
+from httpx import ASGITransport, AsyncClient
 from opensearchpy import OpenSearch
 from sqlalchemy.engine.interfaces import IsolationLevel
 
@@ -95,7 +95,7 @@ async def override_get_search_engine():
         }
     )
 
-    async with AsyncClient(app=app, base_url="http://testserver") as async_client:
+    async with AsyncClient(transport=ASGITransport(app=app), base_url="http://testserver") as async_client:
         yield async_client
 
     # Clear from `api_v1` — that is where the overrides above were registered. Clearing