We are thrilled to announce our newly launched Unstructured API. While access to the hosted Unstructured API will remain free, API Keys are required to make requests. To prevent disruption, get yours here now and start using it today! Check out the readme here to get started making API calls.

We are releasing the beta version of our Chipper model to deliver superior performance when processing high-resolution, complex documents. To start using the Chipper model in your API request, you can utilize the hi_res strategy. Please refer to the documentation here.

As the Chipper model is in beta version, we welcome feedback and suggestions. For those interested in testing the Chipper model, we encourage you to connect with us on Slack community.

This repo implements a pre-processing pipeline for the following documents. Currently, the pipeline is capable of recognizing the file type and choosing the relevant partition function to process the file.

Try our hosted API! It's freely available to use with any of the filetypes listed above. This is the easiest way to get started. If you'd like to host your own version of the API, jump down to the Developer Quickstart Guide.

 curl -X 'POST' \
  'https://api.unstructured.io/general/v0/general' \
  -H 'accept: application/json' \
  -H 'Content-Type: multipart/form-data' \
  -H 'unstructured-api-key: <YOUR API KEY>' \
  -F 'files=@sample-docs/family-day.eml' \
  | jq -C . | less -R

Four strategies are available for processing PDF/Images files: hi_res, fast, ocr_only and auto. fast is the default strategy and works well for documents that do not have text embedded in images.

On the other hand, hi_res is the better choice for PDFs that may have text within embedded images, or for achieving greater precision of element types in the response JSON. Please be aware that, as of writing, hi_res requests may take 20 times longer to process compared to the fast option. See the example below for making a hi_res request.

 curl -X 'POST' \
  'https://api.unstructured.io/general/v0/general' \
  -H 'accept: application/json' \
  -H 'Content-Type: multipart/form-data' \
  -F 'files=@sample-docs/layout-parser-paper.pdf' \
  -F 'strategy=hi_res' \
  | jq -C . | less -R

The ocr_only strategy runs the document through Tesseract for OCR. Currently, hi_res has difficulty ordering elements for documents with multiple columns. If you have a document with multiple columns that do not have extractable text, we recommend using the ocr_only strategy. Please be aware that ocr_only will fall back to another strategy if Tesseract is not available.

For the best of all worlds, auto will determine when a page can be extracted using fast or ocr_only mode, otherwise it will fall back to hi_res.

The hi_res strategy supports different models, and the default is detectron2onnx. You can also specify hi_res_model_name parameter to run hi_res strategy with the chipper model while using the host API:

 curl -X 'POST' \
  'https://api.unstructured.io/general/v0/general' \
  -H 'accept: application/json' \
  -H 'Content-Type: multipart/form-data' \
  -F 'files=@sample-docs/layout-parser-paper.pdf' \
  -F 'strategy=hi_res' \
  -F 'hi_res_model_name=chipper'  \
  | jq -C . | less -R

We also support models to be used locally, for example, yolox. Please refer to the using-the-api-locally section for more information on how to use the local API.

Note: This kwarg will eventually be deprecated. Please use languages. You can also specify what languages to use for OCR with the ocr_languages kwarg. See the Tesseract documentation for a full list of languages and install instructions. OCR is only applied if the text is not already available in the PDF document.

curl -X 'POST' \
  'https://api.unstructured.io/general/v0/general' \
  -H 'accept: application/json' \
  -H 'Content-Type: multipart/form-data' \
  -F 'files=@sample-docs/english-and-korean.png' \
  -F 'strategy=ocr_only' \
  -F 'ocr_languages=eng'  \
  -F 'ocr_languages=kor'  \
  | jq -C . | less -R

You can also specify what languages to use for OCR with the languages kwarg. See the Tesseract documentation for a full list of languages and install instructions. OCR is only applied if the text is not already available in the PDF document.

curl -X 'POST' \
  'https://api.unstructured.io/general/v0/general' \
  -H 'accept: application/json' \
  -H 'Content-Type: multipart/form-data' \
  -F 'files=@sample-docs/english-and-korean.png' \
  -F 'strategy=ocr_only' \
  -F 'languages=eng'  \
  -F 'languages=kor'  \
  | jq -C . | less -R

When elements are extracted from PDFs or images, it may be useful to get their bounding boxes as well. Set the coordinates parameter to true to add this field to the elements in the response.

 curl -X 'POST' \
  'https://api.unstructured.io/general/v0/general' \
  -H 'accept: application/json' \
  -H 'Content-Type: multipart/form-data' \
  -F 'files=@sample-docs/layout-parser-paper.pdf' \
  -F 'coordinates=true' \
  | jq -C . | less -R

Currently, we provide support for enabling and disabling table extraction for all file types. Set parameter skip_infer_table_types to specify the document types that you want to skip table extraction with. By default, we enable table extraction for all file types (skip_infer_table_types=[]). Again, please note that table extraction only works with hi_res strategy. For example, if you want to skip table extraction for images, you can pass a list with matching image file types:

 curl -X 'POST' \
  'https://api.unstructured.io/general/v0/general' \
  -H 'accept: application/json' \
  -H 'Content-Type: multipart/form-data' \
  -F 'files=@sample-docs/layout-parser-paper-with-table.jpg' \
  -F 'strategy=hi_res' \
  -F 'skip_infer_table_types=["jpg"]' \
  | jq -C . | less -R

You can specify the encoding to use to decode the text input. If no value is provided, utf-8 will be used.

curl -X 'POST' \
 'https://api.unstructured.io/general/v0/general' \
 -H 'accept: application/json'  \
 -H 'Content-Type: multipart/form-data' \
 -F 'files=@sample-docs/fake-power-point.pptx' \
 -F 'encoding=utf_8' \
 | jq -C . | less -R

You can send gzipped file and api will un-gzip it.

curl -X 'POST' \
 'https://api.unstructured.io/general/v0/general' \
 -H 'accept: application/json'  \
 -H 'Content-Type: multipart/form-data' \
 -F 'gz_uncompressed_content_type=application/pdf' \
 -F 'files=@sample-docs/layout-parser-paper.pdf.gz' 

If field gz_uncompressed_content_type is set, the API will use its value as content-type of all files after uncompressing the .gz files that are sent in single batch. If not set, the API will use various heuristics to detect the filetypes after uncompressing from .gz.

When processing XML documents, set the xml_keep_tags parameter to true to retain the XML tags in the output. If not specified, it will simply extract the text from within the tags.

curl -X 'POST' \
 'https://api.unstructured.io/general/v0/general' \
 -H 'accept: application/json'  \
 -H 'Content-Type: multipart/form-data' \
 -F 'files=@sample-docs/fake-xml.xml' \
 -F 'xml_keep_tags=true' \
 | jq -C . | less -R

For supported filetypes, set the include_page_breaks parameter to true to include PageBreak elements in the output.

curl -X 'POST' \
 'https://api.unstructured.io/general/v0/general' \
 -H 'accept: application/json'  \
 -H 'Content-Type: multipart/form-data' \
 -F 'files=@sample-docs/layout-parser-paper-fast.pdf' \
 -F 'include_page_breaks=true' \
 | jq -C . | less -R

By default, the element ID is a SHA-256 hash of the element text. This is to ensure that the ID is deterministic. One downside is that the ID is not guaranteed to be unique. Different elements with the same text will have the same ID, and there could also be hash collisions. To use UUIDs in the output instead, set unique_element_ids=true. Note: this means that the element IDs will be random, so with every partition of the same file, you will get different IDs. This can be helpful if you'd like to use the IDs as a primary key in a database, for example.

curl -X 'POST' \ 
 'https://api.unstructured.io/general/v0/general' \
 -H 'accept: application/json'  \
 -H 'Content-Type: multipart/form-data' \
 -F 'files=@sample-docs/layout-parser-paper-fast.pdf' \
 -F 'unique_element_ids=true' \
 | jq -C . | less -R

Use the chunking_strategy form-field to chunk text into larger or smaller elements. Defaults to None which performs no chunking. The available chunking strategies are basic and by_title.

The basic strategy combines whole consecutive document elements to maximally fill chunks of max_characters length. A single element that by itself exceeds max_characters is divided into two or more chunks by text-splitting (on a word boundary).

The by_title strategy has the same behaviors except document section boundaries are respected, meaning elements from two different sections never occur in the same chunk. A Title (section heading) element introduces a new section, hence the name.

Additional Parameters (all optional):

`max_characters`
  The hard maximum chunk size. No chunk will exceed this length. Defaults to 500.

`new_after_n_chars`
  A chunk of this length or greater is considered "full" and will not receive an additional element, even if it would fit within `max_characters`.
  This "soft-maximum" defaults to `max_characters`.

`overlap`
  Specifies the length of a string ("tail") to be drawn from each chunk and prefixed to the
  next chunk as a context-preserving mechanism. By default, this only applies to split-chunks
  where an oversized element is divided into multiple chunks by text-splitting.

`overlap_all`
  Default: `False`. When `True`, apply overlap between "normal" chunks formed from whole
  elements and not subject to text-splitting. Use this with caution as it entails a certain
  level of "pollution" of otherwise clean semantic chunk boundaries.

`combine_under_n_chars`
  Combines elements (for example a series of titles) until a section reaches a
  length of n characters. Defaults to 500. Only operative for the "by_title"
  strategy.

`multipage_sections`
  If True, sections can span multiple pages. Defaults to True. Only operative for
  the "by_title" strategy.

curl -X 'POST' 
 'https://api.unstructured.io/general/v0/general' \
 -H 'accept: application/json'  \
 -H 'Content-Type: multipart/form-data' \
 -F 'files=@sample-docs/layout-parser-paper-fast.pdf' \
 -F 'chunking_strategy=by_title' \
 | jq -C . | less -R

• Using pyenv to manage virtualenv's is recommended

  • Mac install instructions. See here for more detailed instructions.

    • brew install pyenv-virtualenv
    • pyenv install 3.10.12

  • Linux instructions are available here.

  • Create a virtualenv to work in and activate it, e.g. for one named document-processing:

    pyenv virtualenv 3.10.12 unstructured-api
pyenv activate unstructured-api

See the Unstructured Quick Start for the many OS dependencies that are required, if the ability to process all file types is desired.

• Run make install
• Start a local jupyter notebook server with make run-jupyter
  OR
  just start the fast-API locally with make run-web-app

After running make run-web-app (or make docker-start-api to run in the container), you can now hit the API locally at port 8000. The sample-docs directory has a number of example file types that are currently supported.

For example:

 curl -X 'POST' \
  'http://localhost:8000/general/v0/general' \
  -H 'accept: application/json' \
  -H 'Content-Type: multipart/form-data' \
  -F 'files=@sample-docs/family-day.eml' \
  | jq -C . | less -R

The response will be a list of the extracted elements:

[
  {
    "element_id": "db1ca22813f01feda8759ff04a844e56",
    "coordinates": null,
    "text": "Hi All,",
    "type": "UncategorizedText",
    "metadata": {
      "date": "2022-12-21T10:28:53-06:00",
      "sent_from": [
        "Mallori Harrell <mallori@unstructured.io>"
      ],
      "sent_to": [
        "Mallori Harrell <mallori@unstructured.io>"
      ],
      "subject": "Family Day",
      "filename": "family-day.eml"
    }
  },
...
...

The output format can also be set to text/csv to get the data in csv format rather than json:

 curl -X 'POST' \
  'http://localhost:8000/general/v0/general' \
  -H 'accept: application/json' \
  -H 'Content-Type: multipart/form-data' \
  -F 'files=@sample-docs/family-day.eml' \
  -F 'output_format="text/csv"'

The response will be a list of the extracted elements in csv format:

type,element_id,text,filename,sent_from,sent_to,subject,languages,filetype
UncategorizedText,db1ca22813f01feda8759ff04a844e56,"Hi All,",family-day.eml,['Mallori Harrell <mallori@unstructured.io>'],['Mallori Harrell <mallori@unstructured.io>'],Family Day,['eng'],message/rfc822
NarrativeText,a663c393a5e143c01ef2bb5c98efa2c1,Get excited for our first annual family day! ,family-day.eml,['Mallori Harrell <mallori@unstructured.io>'],['Mallori Harrell <mallori@unstructured.io>'],Family Day,['eng'],message/rfc822
NarrativeText,ce65ca3bef59957d3f1c2bab5725c82f,"There will be face painting, a petting zoo, funnel cake and more.",family-day.eml,['Mallori Harrell <mallori@unstructured.io>'],['Mallori Harrell <mallori@unstructured.io>'],Family Day,['eng'],message/rfc822
NarrativeText,d7bcf988af9f06042d83e25c531e5744,Make sure to RSVP!,family-day.eml,['Mallori Harrell <mallori@unstructured.io>'],['Mallori Harrell <mallori@unstructured.io>'],Family Day,['eng'],message/rfc822
Title,5550577db69c2c8aabcd90979698120a,Best.,family-day.eml,['Mallori Harrell <mallori@unstructured.io>'],['Mallori Harrell <mallori@unstructured.io>'],Family Day,['eng'],message/rfc822
Title,ca1c571d993b6c1ed8ef56a06c16ba22,Mallori Harrell,family-day.eml,['Mallori Harrell <mallori@unstructured.io>'],['Mallori Harrell <mallori@unstructured.io>'],Family Day,['eng'],message/rfc822
Title,d5b612de8cd918addd9569b0255b65b2,Unstructured Technologies,family-day.eml,['Mallori Harrell <mallori@unstructured.io>'],['Mallori Harrell <mallori@unstructured.io>'],Family Day,['eng'],message/rfc822
Title,2e0b9e8ee04b9594a9c26d8535b818ff,Data Scientist,family-day.eml,['Mallori Harrell <mallori@unstructured.io>'],['Mallori Harrell <mallori@unstructured.io>'],Family Day,['eng'],message/rfc822

As mentioned above, processing a pdf using hi_res is currently a slow operation. One workaround is to split the pdf into smaller files, process these asynchronously, and merge the results. You can enable parallel processing mode with the following env variables:

• UNSTRUCTURED_PARALLEL_MODE_ENABLED - set to true to process individual pdf pages remotely, default is false.
• UNSTRUCTURED_PARALLEL_MODE_URL - the location to send pdf page asynchronously, no default setting at the moment.
• UNSTRUCTURED_PARALLEL_MODE_THREADS - the number of threads making requests at once, default is 3.
• UNSTRUCTURED_PARALLEL_MODE_SPLIT_SIZE - the number of pages to be processed in one request, default is 1.
• UNSTRUCTURED_PARALLEL_RETRY_ATTEMPTS - the number of retry attempts on a retryable error, default is 2. (i.e. 3 attempts are made in total)

Due to the overhead associated with file splitting, parallel processing mode is only recommended for the hi_res strategy. Additionally users of the official Python client can enable client-side splitting by setting split_pdf_page=True.

You may also set the optional UNSTRUCTURED_API_KEY env variable to enable request validation for your self-hosted instance of Unstructured. If set, only requests including an unstructured-api-key header with the same value will be fulfilled. Otherwise, the server will return a 401 indicating that the request is unauthorized.

Some documents will use a lot of memory as they're being processed. To mitigate OOM errors, the server will return a 503 if the host's available memory drops below 2GB. This is configured with the environment variable UNSTRUCTURED_MEMORY_FREE_MINIMUM_MB, which defaults to 2048. You can lower this value to reduce these messages, that is, allow the server to use more memory. Otherwise, you can set to 0 to fully remove this check.

By default server will run for indefinitely. To change that the MAX_LIFETIME_SECONDS environmental variable can be set. If server is run with this variable set, it will enter a graceful shutdown period after MAX_LIFETIME_SECONDS from its initialization. Graceful shutdown period lasts for up to 3600 seconds and during it:

• server denies any new requests - they're met with an empty response,
• server continues processing active requests and shuts down (ending graceful period) if all of them are processed.

After the graceful period is over if server is still running, it is shutdown forcefully, cancelling all active requests and sending empty responses to each of them.

Max lifetime requires gnu timeout to be installed, available by default on most linux systems. Downloadable on macOS as gtimeout with gnu coreutils.

The following instructions are intended to help you get up and running using Docker to interact with unstructured-api. See here if you don't already have docker installed on your machine.

NOTE: we build multi-platform images to support both x86_64 and Apple silicon hardware. Docker pull should download the corresponding image for your architecture, but you can specify with --platform (e.g. --platform linux/amd64) if needed.

We build Docker images for all pushes to main. We tag each image with the corresponding short commit hash (e.g. fbc7a69) and the application version (e.g. 0.5.5-dev1). We also tag the most recent image with latest. To leverage this, docker pull from our image repository.

docker pull downloads.unstructured.io/unstructured-io/unstructured-api:latest

Once pulled, you can launch the container as a web app on localhost:8000.

docker run -p 8000:8000 -d --rm --name unstructured-api downloads.unstructured.io/unstructured-io/unstructured-api:latest

You can pass in a PORT variable to run the server on a different port in the container.

docker run -p 9500:9500 -d --rm --name unstructured-api -e PORT=9500 downloads.unstructured.io/unstructured-io/unstructured-api:latest

See our security policy for information on how to report security vulnerabilities.

We’ve partnered with Scarf (https://scarf.sh) to collect anonymized user statistics to understand which features our community is using and how to prioritize product decision-making in the future. To learn more about how we collect and use this data, please read our Privacy Policy.