December 2024

In the last month, Sensible upgraded its document-type classification endpoints to use large language model (LLM) completions instead of embeddings-based scorings, updated its GTP-4 vision model for better multimodal data extraction results, enabled uploading an unlimited number of reference documents to the web app, and added advanced functionality to several layout-based and computed field methods.

Improvement: Classify documents by type using LLMs

The classify endpoints now use LLM completions instead of embeddings-based scoring to classify documents by type. This change improves accuracy and makes configuration easier for users. Instead of uploading a variety of reference documents to improve classification performance, now you can optionally provide a brief description of the document type, e.g. "This type of document is a car rental agreement."

Improvement: GPT-4 Vision model version update

The Query Group method's Multimodal Engine, which enables you to extract data from non-text images, now uses GPT-4o mini.

UX improvement: Unlimited number of reference documents

For each document type you define in your account, you can now upload an unlimited number of reference documents.

Improvement: Zip computed field method

Sensible has improved the consistency of the Zip computed field method's behavior when you input a mix of arrays, tables, and sections. As a result you can now:

• Zip multiple tables together. The previous workaround was to first convert the tables to sections.
• Zip a mix of tables, section groups, and arrays. The previous behavior was to discard all non-section source IDs in a sources_id list that included a section, or discard non-table sources in a sources_id list that included table and array sources.

For more information, and for edge-case behavior retained for backward compatibility, see the Zip computed field method.

Improvement: Capture multiple intersections on a horizontal line

With the Intersection method's new Horizontal Anchor parameter, you can extract multiple intersections along a horizontal line, in addition to the existing behavior of capturing multiple intersections along a vertical line.

Improvement: API extraction response now includes configuration_version and content_type

The extract endpoints now return the configuration_version of the configuration used to perform the extraction and the content_type of the document, for example, application/pdf.

UX improvement: Compact view of extraction history

The Extraction history tab now displays an improved, more compact view of past extractions by introducing new icons, including:

• a new extraction status icon next to the document name. If an extraction fails, you can click the icon to view the error message.
• a new review status icon next to the document name. If the extraction needs review, you can click the icon and open the review page.