Skip to content

ContextualWisdomLab/newsdom-api

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

161 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

NewsDOM API

OpenSSF Scorecard

NewsDOM API is a language-agnostic PDF-to-DOM sidecar. It converts any PDF into a canonical JSON document tree — pages, sections, headings, body blocks, images, captions, and bounding boxes — using MinerU. It runs as a standalone FastAPI service and is designed to be embedded as a git submodule / sidecar in a larger system.

Features

  • Primary engine: MinerU pipeline backend
  • Service wrapper: FastAPI
  • Output: canonical JSON with pages, sections, section headings, body blocks, images, captions, bounding boxes, and quality metadata
  • Language-selectable: language defaults to MinerU's multilingual ch model; official language families and compatibility aliases such as japan remain available
  • Parsing mode defaults to auto so born-digital text PDFs skip forced OCR
  • Optional bearer auth on /parse; unauthenticated /health liveness probe

Quickstart

Install

Install uv first if it is not already available in your PATH, then sync the repository-managed virtual environment:

uv sync --frozen --all-extras

To enable real parsing with MinerU, install the MinerU CLI separately in the same .venv that uv sync created:

uv pip install --python .venv/bin/python "mineru[pipeline]==3.4.4"

On Windows, replace .venv/bin/python with .venv\Scripts\python.exe.

Run

uv run uvicorn --app-dir src newsdom_api.main:app --reload

Docker

docker build -t newsdom-api .
docker run -p 8000:8000 newsdom-api

The default image exposes the REST API on port 8000 as a multi-arch service image. It is suitable for linux/amd64 and linux/arm64, including Apple Silicon hosts running the API service inside Docker.

The default image ships the API service only and does not bundle the MinerU runtime. /parse requires a compatible MinerU runtime to be available inside the container image or exposed through NEWSDOM_MINERU_BIN.

Readiness caveat: /health reports process liveness only. Because the default image does not bundle MinerU, a container can report a green /health while /parse still returns 503 until a MinerU runtime is reachable. Treat MinerU availability as a separate readiness concern when wiring this sidecar into a larger system.

docker compose

A docker-compose.yml is provided for standalone/sidecar use. Its healthcheck targets the unauthenticated /health endpoint:

docker compose up --build

Uncomment NEWSDOM_MINERU_BIN (path to a MinerU executable) and/or NEWSDOM_API_TOKEN (bearer secret) in the compose environment: block to make /parse functional and/or protected.

Building a MinerU-bundled image

To ship an image where /parse works behind a green /health, bundle a MinerU runtime. Either build the NVIDIA variant below, or extend the default image and install MinerU into the same virtualenv, for example:

FROM newsdom-api:latest
USER root
RUN uv pip install --python /app/.venv/bin/python "mineru[pipeline]==3.4.4"
USER newsdom

Alternatively mount a MinerU executable and point NEWSDOM_MINERU_BIN at it, so /parse no longer returns 503 behind a healthy /health.

For heavier parsing deployments, build the optional NVIDIA-oriented variant:

docker build -f Dockerfile.nvidia -t newsdom-api:nvidia .
docker run --gpus all -p 8000:8000 newsdom-api:nvidia

Dockerfile.nvidia is intended for Linux/NVIDIA environments and is linux/amd64-only. Apple Silicon can run the lean API image, but Docker Desktop does not expose Apple GPU acceleration to Linux containers, so real GPU-accelerated parsing should stay on a native Apple Silicon path instead of the containerized runtime.

The NVIDIA variant is linux/amd64-only and is meant for hosts that can provide the CUDA user-space/runtime stack required by MinerU.

Parse a PDF

curl -F "file=@sample.pdf" http://127.0.0.1:8000/parse

/parse accepts multipart/form-data with a required file part (application/pdf) and two optional form fields:

Field Default Values Maps to
language ch MinerU 3.4.4 public family or alias (ch, en, japan, korean, arabic, east_slavic, cyrillic, devanagari, …) MinerU -l
mode auto auto, ocr, txt MinerU -m

mode=auto lets born-digital (text-layer) PDFs skip forced OCR; ocr forces optical recognition and txt extracts only the embedded text layer. Invalid values return 422. The previous Japanese-newspaper behavior is still available explicitly:

curl -F "file=@sample.pdf" -F "language=japan" -F "mode=ocr" \
  http://127.0.0.1:8000/parse

The accepted language contract follows the official MinerU 3.4.4 CLI implementation. MinerU canonicalizes en, japan, chinese_cht, and latin to ch; this sidecar performs the same normalization before launching the subprocess.

Authentication

/parse is unauthenticated by default (development). Set the sidecar's own NEWSDOM_API_TOKEN environment variable to require a bearer token:

NEWSDOM_API_TOKEN=$(openssl rand -hex 32) \
  uv run uvicorn --app-dir src newsdom_api.main:app
curl -F "file=@sample.pdf" -H "Authorization: Bearer $NEWSDOM_API_TOKEN" \
  http://127.0.0.1:8000/parse

Requests without a matching Authorization: Bearer <token> header receive 401. /health stays unauthenticated so orchestrators can always probe it. Supply the token from your deployment's secret store rather than committing it.

Each request is written to a request-scoped temporary directory before MinerU runs, and those temporary files are removed after the response completes. Sanitized parse failures return 503 MinerU runtime unavailable when the runtime cannot be executed and 502 MinerU output was incomplete when MinerU finishes without the required output artifacts.

Run tests

uv run pytest

Fuzzing smoke

uv run python fuzzers/dom_builder_fuzzer.py --smoke tests/fixtures/mineru_sample.json
uv run python fuzzers/schema_response_fuzzer.py --smoke fuzzers/corpus/schema_response_fuzzer/valid_parse_response.json
uv run python fuzzers/equivalence_metrics_fuzzer.py --smoke fuzzers/corpus/equivalence_metrics_fuzzer/structural_metrics.json

Every fuzzers/*_fuzzer.py target is coverage-guided under Atheris and is picked up automatically by the ClusterFuzzLite workflow, which runs a bounded budget on each pull request. Targets cover the untrusted-input boundaries: the MinerU DOM normalizer (build_dom), the ParseResponse schema validator, and the equivalence metrics normalizer. See docs/papers/ for background.

The repository also enforces a quality-gate workflow with 100% source coverage and docstring audit coverage.

Fixtures and provenance

This repository ships only synthetic test fixtures and derived structural baselines. For fixture provenance and regeneration notes, see tests/fixtures/README.md.

Development

Development setup, fixture handling rules, and local-only baseline maintenance are documented in CONTRIBUTING.md.

Mechanical branch updates and merges are attributed to github-actions[bot]. Scratch PoC files are not committed. Failed GitHub Checks are not reviewed as URL lists. OpenCode Review, Strix Security Scan, and PR Review Merge Scheduler are provided by the organization-level required workflows in ContextualWisdomLab/.github, not copied into this repository.

Security reporting guidance is documented in SECURITY.md. Version tags trigger a GitHub-native release workflow that builds distribution artifacts, checksums, and provenance attestations.

Project history is tracked in CHANGELOG.md.

Repository branch workflow is documented in docs/workflow/git-flow.md.

Repository layout

  • src/newsdom_api/: API, MinerU wrapper, DOM builder, synthetic fixture generator
  • tests/: unit tests and committed synthetic fixtures
  • tools/: local maintenance utilities

About

스캔된 일본어 신문 PDF를 기사·제목·본문·이미지 구조의 DOM형 JSON으로 파싱하는 API.

Topics

Resources

License

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Packages

 
 
 

Contributors

Languages