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.
Category | Document Types |
---|---|
Plaintext | .txt , .eml , .msg , .xml , .html , .md , .rst , .json , .rtf |
Images | .jpeg , .png |
Documents | .doc , .docx , .ppt , .pptx , .pdf , .odt , .epub , .csv , .tsv , .xlsx |
Zipped | .gz |
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
To extract the table structure from PDF files using the hi_res
strategy, ensure that the pdf_infer_table_structure
parameter is set to true
. This setting includes the table's text content in the response. By default, this parameter is set to false
to avoid the expensive reading process.
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 'pdf_infer_table_structure=true' \
| jq -C . | less -R
Currently, we provide support for enabling and disabling table extraction for file types other than PDF files. Set parameter skip_infer_table_types
to specify the document types that you want to skip table extraction with. By default, we skip table extraction
for PDFs and Images, which are pdf
, jpg
and png
. Again, please note that table extraction only works with hi_res
strategy. For example, if you don't want to skip table extraction for images, you can pass an empty value to skip_infer_table_types
with:
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=[]' \
| 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
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
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 withmake 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' \
'https://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 <[email protected]>"
],
"sent_to": [
"Mallori Harrell <[email protected]>"
],
"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' \
'https://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,text,element_id,filename,page_number,url,sent_from,sent_to,subject,sender\n
UncategorizedText,\"Hi,\",bc50944723f014607ad612b6983944a7,alert.eml,1,,['Mallori Harrell <[email protected]>'],['Mallori Harrell <[email protected]>'],ALERT: Stolen Lunch,Mallori Harrell <[email protected]>\n
NarrativeText,\"It has come to our attention that as of 9:00am this morning, Harold's lunch is missing. If this was done in error please return the lunch immediately to the fridge on the 2nd floor by noon.\",51944d1f63f9472edb165fb3c9e5c525,alert.eml,1,,['Mallori Harrell <[email protected]>'],['Mallori Harrell <[email protected]>'],ALERT: Stolen Lunch,Mallori Harrell <[email protected]>\n
NarrativeText,\"If the lunch has not been returned by noon, we will be reviewing camera footage to determine who stole Harold's lunch.\",8e8f9e2e50e39e072fda08d277aa77b9,alert.eml,1,,['Mallori Harrell <[email protected]>'],['Mallori Harrell <[email protected]>'],ALERT: Stolen Lunch,Mallori Harrell <[email protected]>\n
NarrativeText,The perpetrators will be PUNISHED to the full extent of our employee code of conduct handbook.,736a826679b971f594103fd9751e5c8f,alert.eml,1,,['Mallori Harrell <[email protected]>'],['Mallori Harrell <[email protected]>'],ALERT: Stolen Lunch,Mallori Harrell <[email protected]>\n
UncategorizedText,\"Thank you for your time,\",3eeae5f64dab54c52dd5fff779808071,alert.eml,1,,['Mallori Harrell <[email protected]>'],['Mallori Harrell <[email protected]>'],ALERT: Stolen Lunch,Mallori Harrell <[email protected]>\n
Title,Unstructured Technologies,d5b612de8cd918addd9569b0255b65b2,alert.eml,1,,['Mallori Harrell <[email protected]>'],['Mallori Harrell <[email protected]>'],ALERT: Stolen Lunch,Mallori Harrell <[email protected]>\n
Title,Data Scientist,46b174f1ec7c25d23e5e50ffff0cc55b,alert.eml,1,,['Mallori Harrell <[email protected]>'],['Mallori Harrell <[email protected]>'],ALERT: Stolen Lunch,Mallori Harrell <[email protected]>\n"
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 totrue
to process individual pdf pages remotely, default isfalse
.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 is3
.UNSTRUCTURED_PARALLEL_MODE_SPLIT_SIZE
- the number of pages to be processed in one request, default is1
.UNSTRUCTURED_PARALLEL_RETRY_ATTEMPTS
- the number of retry attempts on a retryable error, default is2
. (i.e. 3 attempts are made in total)
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 configurable with UNSTRUCTURED_MEMORY_FREE_MINIMUM_MB
.
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 --port 8000 --host 0.0.0.0
See our security policy for information on how to report security vulnerabilities.
Section | Description |
---|---|
Unstructured Community Github | Information about Unstructured.io community projects |
Unstructured Github | Unstructured.io open source repositories |
Company Website | Unstructured.io product and company info |
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.