A self-hosted REST API for partitioning documents with the unstructured open-source library. The API detects the uploaded file type, selects the corresponding partitioner, and returns document elements as JSON or CSV.

This repository is for running the Partition Endpoint yourself. For Unstructured's managed API, current platform features, and hosted quickstarts, see the Unstructured API documentation.

• The partitionable document types supported by the pinned unstructured[all-docs] dependency. See the current open-source supported file types.
• PDF and image strategies: auto, fast, hi_res, and ocr_only.
• Optional table and image-block extraction, OCR language selection, page breaks, coordinates, deterministic or UUID element IDs, and gzip uploads.
• by_title chunking, including overlap controls and optional original-element metadata.
• Single-file and multi-file requests, with JSON, CSV, or multipart/mixed responses.

The generated API reference lists the complete request schema. After starting the server, open http://localhost:8000/general/docs, or read the OpenAPI document at http://localhost:8000/general/openapi.json.

• Python 3.12
• uv
• The system dependencies required by the document types you plan to process. See the unstructured installation guide.

Install the locked runtime and test dependencies, then start the development server:

make install
make run-web-app

The API listens on http://localhost:8000. Verify it is ready:

curl --fail http://localhost:8000/healthcheck

Partition a sample document:

curl --request POST \
  --url http://localhost:8000/general/v0/general \
  --header 'accept: application/json' \
  --form 'files=@sample-docs/family-day.eml'

The response is a list of document elements:

[
  {
    "element_id": "db1ca22813f01feda8759ff04a844e56",
    "text": "Hi All,",
    "type": "UncategorizedText",
    "metadata": {
      "filename": "family-day.eml"
    }
  }
]

The default PDF/image strategy is auto. For layout-aware partitioning, request hi_res:

curl --request POST \
  --url http://localhost:8000/general/v0/general \
  --header 'accept: application/json' \
  --form 'files=@sample-docs/layout-parser-paper.pdf' \
  --form 'strategy=hi_res'

Use languages to specify OCR languages. The older ocr_languages parameter remains available for compatibility but is deprecated by the underlying library.

curl --request POST \
  --url http://localhost:8000/general/v0/general \
  --header 'accept: application/json' \
  --form 'files=@sample-docs/english-and-korean.png' \
  --form 'strategy=ocr_only' \
  --form 'languages=eng' \
  --form 'languages=kor'

Chunk the output by document title. Original elements are included in chunk metadata by default; set include_orig_elements=false when you want a smaller response:

curl --request POST \
  --url http://localhost:8000/general/v0/general \
  --header 'accept: application/json' \
  --form 'files=@sample-docs/layout-parser-paper-fast.pdf' \
  --form 'chunking_strategy=by_title' \
  --form 'max_characters=1500' \
  --form 'include_orig_elements=false'

Return CSV instead of JSON:

curl --request POST \
  --url http://localhost:8000/general/v0/general \
  --header 'accept: text/csv' \
  --form 'files=@sample-docs/family-day.eml' \
  --form 'output_format=text/csv'

Multi-platform images for amd64 and arm64 are published to Quay:

docker pull quay.io/unstructured-io/unstructured-api:latest
docker run --rm --name unstructured-api -p 8000:8000 \
  quay.io/unstructured-io/unstructured-api:latest

Pin a release or commit tag instead of latest for reproducible deployments.

make docker-build
make docker-start-api

Parallel mode splits PDFs into page ranges, sends those ranges to another Partition Endpoint, and combines the results. It is most useful for hi_res workloads.

Alternatively, the official Python client can split PDFs client-side with split_pdf_page=True.

Install the locked development environment:

make install

Run the checks and test suite:

make check
make test

See the Unstructured contribution guide before opening a pull request.

Report vulnerabilities through this repository's security policy.

Licensed under the Apache License 2.0.