Extracting Content
Understand how to convert PDF Documents to Markdown for use in AI Agents
With the Document Parsing API, you can parse a document with a single API call and you will always get in return:
- A markdown version of the document, including:
- Tables encoded as Markdown or HTML
- [Optional] Markdown chunks based on page, section, or fragment
- A complete document layout, including:
- Page numbers
- Bounding boxes for each fragment found in the document (e.g. signature, key-value pair, figure)
- [Optional] Structured data based on the schema you define
- See the Structured Data Extraction documentation for more information.
Your data is NOT sent to a third party service(OpenAI, Anthropic, etc), and uses our own models to parse the document.
Core Document Parsing Workflow
With a file_id
and an api_key
, you can quickly parse a document with a single API call. And more importantly, you can control
how the document is parsed.
Call the `/parse` endpoint
The /parse
endpoint will create a parse job with the following request payload:
file_id
: Either thefile_id
returned from uploading a file to Tensorlake Cloud, or a pre-signed URL or any HTTP URL that can be used to download the filesettings
: Settings related to how the document will be parsed. See the main configuration settings below, and get a full list in the API Reference.
The endpoint will return:
parse_id
: The unique ID Tensorlake uses to reference the specific parsing job. This ID can be used to get the output when the parsing job is completed and re-visit previously used settings.
Query the status using the `/parse/{parse_id}` endpoint
The /parse/{parse_id}
endpoint will return:
status
: The status of the parsing job. This can beFAILURE
,PENDING
,PROCESSING
, orSUCCESSFUL
.- If the parsing job is
PENDING
orPROCESSING
, you should wait a few seconds and then check again by re-calling the endpoint.
Retrieve the output using the `/parse/{parse_id}` endpoint
If the /parse/{parse_id}
endpoint returns as SUCCESSFUL
status, the response payload will include an Output
object:
outputs.chunks
: An array of objects that contain a chunk number and the markdown content for that chunk.outputs.document
: A JSON representation of the entire document, including any errors that happened while parsing.outputs.num_pages
: An integer representing the number of pages that were parsed.- [optionally]
outputs.structured_data
: An array of objects that contain the structured data as JSON and the page number where it was found.
Explore Parse Configuration Settings
The main configuration settings for the parsing job are:
Setting | Options | Default Value |
---|---|---|
pages | Select the range of pages to parse. | None (all pages) |
tableOutputMode | Choose between Markdown, , or JSON. | "markdown" |
tableParsingMode | Choose between or . | "tsr" |
chunkStrategy | Choose between , Page, Section, or Fragment. | None (a single md document) |
detectStrikethrough | remove lines that have a strikethrough. | false |
deliverWebhook | Deliver a webhook when the parsing job finishes. Learn how to configure webhooks here. | false |
Get a full list of the configuration setting options on the /parse
section of the API reference.
Use the /parse
API
Calling the /parse
enpoint will create a new document parsing job, starting in the pending
state. It will transition to the processing
state and then to the successful
state when it’s parsed successfully.
If you are using the Python SDK, all the configuration options described above are expressed through
the ParsingOptions
class.
If you are using the Python SDK, all the configuration options described above are expressed through
the ParsingOptions
class.
The HTTP API for parsing is thoroughly documented here. Here is an example of how to initiate a parsing job:
Retrieve Output from the Parsing Job
The parsed document output can be retrieved using the /parse/{parse_id}
endpoint, or using the get_job
SDK function.
The response is a JSON object if you are using the REST API, and a Job
object if you are using the Python SDK.
Understand the Parsing Output
The outputs
attribute of the response contains the following fields which returns the parsed document.
- outputs.num_pages: An integer representing the number of pages that were parsed.
- outputs.chunks: An array of objects that contain a chunk number and the markdown content for that chunk. See more below.
- outputs.document: A JSON representation of the entire document, including any errors that happened while parsing. See more below.
- [optionally] outputs.structured_data: An array of objects that contain the structured data as JSON and the page number where it was found.
- outputs.errors: The errors encountered while parsing the document.
The Outputs class has been documented in the Python SDK and in the REST API.
Markdown Chunks
The markdown content of the document is available in the outputs.chunks
attribute of the JSON response. The number of chunks
depends on the chunking strategy you chose.
Chunking Strategy Options
- None - The whole document is returned as a single chunk. This allows you to use your own chunking logic.
- Page - Each page is returned as a separate chunk. You should receive as many chunks as the number of pages in the document.
- Section - The document is split into chunks based on the section headers detected in the document.
- Fragment - Every page frament (e.g. table, figure, paragraph) is returned as a separate chunk. You will most likely have to merge these chunks based on your use-case.
Document Layout and Bounding Boxes
The entire document layout is available in the outputs.document
attribute of the JSON response. This object has a list of Pages, each
encoded as a JSON object. Each outputs.document.pages[x]
contains the following attributes:
page_number
- The page number of the page.dimensions
- The width and height of the page in pixels.page_fragments
- The list of objects on the page. Each page fragment has the following attributes:fragment_type
- The type of the object. Currently we detect the following types:- Section Header
- Text
- Table
- Form
- Formula
- Figure
- Signature
reading_order
- The reading order of the page fragments. This is the order in which the fragment would be read by a human.bounding_box
- The bounding box of the page fragment, in the format[x1, y1, x2, y2]
.content
- The actual content that is found on that fragment of the page.
Explore Advanced Capabilities
This page covered the basic reading capabilities of the Document Parsing API. In addition to this, you can also use Document Ingestion for more advanced parsing. Explore these options:
Structured Data Extraction
By simply specifying a schema, you can extract exactly the data you need from any document, all with the same API call as basic parsing.
Summarization
By setting a few extra Settings, you can ensure all tables, figures, and charts are summarized.
Signature Detection
Setting detect_signatures
to true
will ensure all signatures are detected throughout your document.
Page Classification
[Coming Soon] Specify on what types of pages certain structured data can be found for more accurate data retrieval.