danvics/docling-studio
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

feat: enable chunking in remote (Docling Serve) mode

Browse source 
Hybrid approach: reuse LocalChunker to chunk the DoclingDocument JSON
returned by Serve, so chunking works identically in both local and
remote modes without calling Serve's chunk endpoint.

Backend:
- _build_chunker() always returns LocalChunker (remove engine guard)
- Use docling-core[chunking] extra for required dependencies
- Skip client-side batching in remote mode (Serve manages its own
  resources, and batching discards document_json needed for chunking)
- Fix Serve form fields: remove generate_page_images (not a Serve
  field), use repeated form keys for to_formats and page_range
- Log Serve error response body on 4xx/5xx for diagnosis
- Fix FastAPI 204 DELETE routes missing response_model=None

Frontend:
- Update chunking feature flag to enable Prepare UI in remote mode

Closes #51
This commit is contained in:
Pier-Jean Malandrino 2026-04-16 15:01:15 +02:00
parent d0c77b7618
commit e0f8e81b63
12 changed files with 140 additions and 23 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
document-parser/api/analyses.py
View file

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

2
document-parser/api/documents.py
View file

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

2
document-parser/api/ingestion.py
View file

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

20
document-parser/infra/serve_converter.py
View file

@ -96,6 +96,13 @@ class ServeConverter:
headers=self._headers(),
)
if response.status_code >= 400:
logger.error(
"Docling Serve error %d: %s (form_data=%s)",
response.status_code,
response.text[:500],
{k: v for k, v in form_data.items()},
)
response.raise_for_status()
result_data = response.json()
@ -122,8 +129,12 @@ def _build_form_data(
) -> dict[str, str | list[str]]:
"""Build form fields matching Docling Serve's multipart form contract.
Array fields (to_formats) are sent as lists — httpx encodes them as
repeated form keys (to_formats=md&to_formats=html&to_formats=json).
Serve uses FastAPI's ``Form()`` parsing — list/tuple fields are sent
as **repeated form keys** (httpx encodes Python lists this way
automatically: ``to_formats=md&to_formats=html&to_formats=json``).
Note: ``generate_page_images`` is a PdfPipelineOptions field, NOT a
ConvertDocumentsOptions field — sending it causes a 422.
"""
data: dict[str, str | list[str]] = {
"to_formats": ["md", "html", "json"],
@ -135,11 +146,12 @@ def _build_form_data(
"do_picture_classification": str(options.do_picture_classification).lower(),
"do_picture_description": str(options.do_picture_description).lower(),
"include_images": str(options.generate_picture_images).lower(),
"generate_page_images": str(options.generate_page_images).lower(),
"images_scale": str(options.images_scale),
}
if page_range is not None:
data["page_range"] = f"{page_range[0]}-{page_range[1]}"
# Serve expects page_range as two repeated form fields:
# page_range=1&page_range=10
data["page_range"] = [str(page_range[0]), str(page_range[1])]
return data

13
document-parser/main.py
View file

@ -57,12 +57,15 @@ def _build_converter():
def _build_chunker():
"""Build the chunker adapter — only available in local mode."""
if settings.conversion_engine == "local":
from infra.local_chunker import LocalChunker
"""Build the chunker adapter.
return LocalChunker()
return None
Uses LocalChunker in all modes — in remote mode it chunks the
DoclingDocument JSON returned by Docling Serve, so docling-core
(lightweight) is the only local dependency needed.
"""
from infra.local_chunker import LocalChunker
return LocalChunker()
def _build_repos() -> tuple[SqliteDocumentRepository, SqliteAnalysisRepository]:

2
document-parser/requirements.txt
View file

@ -1,4 +1,4 @@
docling-core>=2.0.0,<3.0.0
docling-core[chunking]>=2.0.0,<3.0.0
fastapi>=0.115.0,<1.0.0
uvicorn[standard]>=0.32.0,<1.0.0
python-multipart>=0.0.12

20
document-parser/services/analysis_service.py
View file

@ -324,11 +324,18 @@ class AnalysisService:
file_path: str,
options: ConversionOptions,
) -> ConversionResult | None:
"""Run batched or single conversion. Returns None if the job was deleted mid-batch."""
"""Run batched or single conversion. Returns None if the job was deleted mid-batch.
Batching is only used for local mode — it limits memory usage when
Docling runs in-process. In remote mode the Serve instance manages
its own resources, and batching would discard document_json (needed
for chunking).
"""
total_pages = _count_pdf_pages(file_path)
batch_size = self._config.batch_page_size
is_remote = self._is_remote_converter()
if batch_size > 0 and total_pages > batch_size:
if batch_size > 0 and total_pages > batch_size and not is_remote:
return await self._run_batched_conversion(
job_id, file_path, options, total_pages, batch_size
)
@ -337,6 +344,15 @@ class AnalysisService:
timeout=self._conversion_timeout,
)
def _is_remote_converter(self) -> bool:
"""Check if the converter is a remote (Serve) adapter."""
try:
from infra.serve_converter import ServeConverter
return isinstance(self._converter, ServeConverter)
except ImportError:
return False
async def _finalize_analysis(
self,
job_id: str,

65
document-parser/tests/test_chunking.py
View file

@ -465,3 +465,68 @@ class TestRechunkEndpoint:
},
)
assert resp.status_code == 422
# ---------------------------------------------------------------------------
# Remote chunking path — hybrid local chunking from Serve's document_json
# ---------------------------------------------------------------------------
class TestRemoteChunkingPath:
"""Verify that chunking works on document_json produced by Serve (remote mode)."""
@pytest.mark.asyncio
async def test_rechunk_with_serve_document_json(self):
"""AnalysisService.rechunk() works with a LocalChunker even in remote mode."""
from infra.local_chunker import LocalChunker
from services.analysis_service import AnalysisService
chunker = LocalChunker()
analysis_repo = AsyncMock()
document_repo = AsyncMock()
converter = AsyncMock() # ServeConverter mock — not used for rechunking
service = AnalysisService(
converter=converter,
analysis_repo=analysis_repo,
document_repo=document_repo,
chunker=chunker,
)
# Simulate a completed job with document_json from Serve
job = AnalysisJob(id="j-remote", document_id="d1")
job.mark_running()
job.mark_completed(
markdown="# Title\nParagraph text here.",
html="<h1>Title</h1><p>Paragraph text here.</p>",
pages_json="[]",
document_json=json.dumps({
"schema_name": "DoclingDocument",
"version": "1.0.0",
"name": "test",
"origin": {
"mimetype": "application/pdf",
"filename": "test.pdf",
"binary_hash": 0,
},
"furniture": {"self_ref": "#/furniture", "children": [], "content_layer": "furniture"},
"body": {"self_ref": "#/body", "children": [], "content_layer": "body"},
"groups": [],
"texts": [],
"pictures": [],
"tables": [],
"key_value_items": [],
"form_items": [],
"pages": {},
}),
)
analysis_repo.find_by_id = AsyncMock(return_value=job)
analysis_repo.update_chunks = AsyncMock(return_value=True)
chunks = await service.rechunk(
"j-remote",
{"chunker_type": "hybrid", "max_tokens": 512},
)
assert isinstance(chunks, list)
analysis_repo.update_chunks.assert_called_once()

29
document-parser/tests/test_serve_converter.py
View file

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

4
frontend/src/features/feature-flags/store.test.ts
View file

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

2
frontend/src/features/feature-flags/store.ts
View file

@ -32,7 +32,7 @@ interface FeatureFlagContext {
const featureRegistry: Record<FeatureFlag, FeatureFlagDef> = {
chunking: {
description: 'Document chunking for RAG preparation',
isEnabled: (ctx) => ctx.engine === 'local',
isEnabled: (ctx) => ctx.engine !== null,
},
disclaimer: {
description: 'Show shared-instance disclaimer banner',

2
frontend/src/features/feature-flags/useFeatureFlag.test.ts
View file

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