Classifying documents by type

Sensible "classifies" documents in the following senses:

You can use the Sensible API to classify a document by its similarity to high-level document types you define in your Sensible account, without making an extraction request. For example, classify a document as a 1040s document type or a pay_stubs document type. For more information, see the following sections.
When you extract data from a single document, you classify the document by specifying one of the high-level document types you defined in your account, for example 1040s or pay_stubs. Sensible then automatically classifies the document by its subtype, or "config", in the document type, for example, the 1040s_2018 version or the 1040_20 version of a 1040s document. For more information, see DevOps platform.
When you extract data from multiple documents in a single request, you specify a list of possible document types, and Sensible classifies both high-level document types and subtypes automatically. For example:
1. Sensible classifies, or "segments", each document in a multi-document file, or "portfolio". For example, for a loan_application_bundle.pdf document containing a pay_stubs document, a 1040 document, and a bank_statements document, you can segment each document by its page range in the file, and return its extracted data separately. You can configure LLM- or fingerprint-based segmentation. For more information, see Multi-document extractions.
2. Classify each attached document in an email by document type, then return each document's extracted data separately. For more information, see Getting started with email extraction.

  flowchart TD
        A([Classification]) --> D["Classify-only API"]
        A --> B["Single-doc extraction"]
        A --> C["Multi-doc request"]

        D --> D1["No extraction. Sensible scores doc similarity against your defined types"]
        D1 --> D2["Returns document type (e.g. 1040s or pay_stubs)"]

        B --> B1["You specify document type (e.g. 1040s)"]
        B1 --> B2["Sensible auto-classifies subtype / config (e.g. 1040s_2018)"]
        B2 --> B3[Extract data]

        C --> C1["You specify list of possible document types"]
        C1 --> C2["Portfolio (multi-doc file)"]
        C1 --> C3["Email attachments"]
        C2 --> C2a["Sensible segments each doc by page range, classifies type + subtype,returns each doc's data"]
        C3 --> C3a["Sensible classifies each attachment by type + subtype, returns each doc's data"]

The following sections cover using the Sensible API's Classify endpoints to return a document's high-level type.

Classify endpoints

When you call Sensible's Classify API endpoints, Sensible classifies a document by comparing it to the types you define in your account. For example, you can classify 1040 forms and bank statements if you define the following types in your account:

a bank statements type
a 1040s type

Sensible uses a document type's name and its description for LLM-based classification:

Sensible can classify documents into your document types even if the document type is empty (lacks a config or reference document). For example, if you lack a citibank config or reference document in your bank_statements type, Sensible can still classify a 2023-1-1_citbank_statement_jon_doe.pdf document as a bank statement.
If Sensible doesn't find an existing document type to which to match your document in your account, it returns an error.

To optionally improve classification results, describe each document type in your account in its Settings tab. For examples of descriptions, see Document type descriptions. By default, Sensible classifies a document using all the types you define in your account. You can optionally define a subset of document types for classifying a document.

Use cases for classification endpoints

Use cases for the Classify endpoints include the following examples:

Prior to an extraction workflow. Determine which documents to extract before calling a Sensible extraction endpoint.
Independent from an extraction workflow. Determine where to route each document or to label each document in a system of record.