Overview

Sensible is a developer-first platform for extracting structured data from documents (also called document parsing), for example:

• business forms in PDF format or Microsoft Word
• email bodies and attachments
• spreadsheets
• PNGs and JPEGS of documents

Use Sensible to build document-automation features into your vertical SaaS products.

With Sensible's SenseML language, you can write deterministic or LLM-based extraction queries for any type of document:

And get back key facts as JSON:

{
    "street_address": {
        "value": "1234 ABC COURT",
        "type": "address"
    },
    "included_appliances": [
        {
            "value": "washers",
            "type": "string"
        },
        {
            "value": "dryers",
            "type": "string"
        },
        {
            "value": "refrigerators",
            "type": "string"
        }
    ]
}

Sensible is highly configurable. You can extract data in minutes by leveraging large language models (LLMs), or you can get fine-grained control with Sensible's visual, layout-based rules. By combining layout- and LLM-based extraction methods, Sensible supports the entire document landscape, from consistently laid-out, highly structured business forms to free-form, variable legal contracts:

Configurable data extraction

Configure your extractions using SenseML, Sensible's document-specific query language. SenseML extracts document primitives like rows, tables, checkboxes, sections, and more as JSON.

With SenseML, you can:

• Preprocess documents deterministically by correcting layout metadata problems, removing unwanted pages, and more, so that Sensible has a clean, standardized text representation of the document from which to extract structured data in a later step. For more information, see Preprocessors.
• Use "methods" to extract document primitives, like rows, columns, tables, boxes, checkbox status, and more. You can also filter and format extracted data types like currencies, dates, addresses, or your custom types. You can mix and match LLM and deterministic methods in the same extraction config. For more information, see Layout-based methods.
• Handle document variations in a set of similar documents (a "document type") by conditionally executing SenseML methods based on the document's contents.
• Post-process extracted document data. For example:
  • Write logical validations like customer ID is 9 digits to throw custom errors and warnings about your extracted data.
  • Manipulate the extracted data schema with computed methods like concat, split, and custom logic. Or, completely transform Sensible's standardized parsed_document output into any schema to fit your data consumption needs using postprocessors.
  • Get measures of accuracy for LLMs with confidence scores, and get overall measures of extraction completeness with extraction coverage scores.

A field is the basic SenseML query unit for extracting a piece of document data. The output of a field is a JSON key-value pair that structures the extracted data. SenseML is the basis for Sensible's extraction workflow.

Here's an example of a field that extracts a table:

For more information about SenseML, see SenseML reference introduction.

DevOps platform for document data extraction

See the following image for a high-level overview of Sensible's document data extraction workflow:

For more information about this diagram, see DevOps platform.

Learn more

To use the Sensible platform, you'll:

• Learn to extract data, or use out-of-the-box supported document types. See Getting started and Getting started with layout-based extractions.
• Integrate using Sensible's API, SDKs, quick-extract UI, or other tools.
• Validate extracted data by writing rules for custom errors like extracted zip code is invalid format.
• Monitor extraction metrics in production.
• Review and correct extracted data at the field level using the human-in-the-loop review UI.