Unstructured Transform

Not yet independently verified. The comparison with general-purpose models is Unstructured's own test on 224 pages it chose, and the post does not say how the hallucination rate or table score were measured; we have not seen an independent test. Prices come from Unstructured's pricing page, read on 25 October 2026. We will update this when it can be confirmed, and remove this note.

Unstructured Transform is the document parsing service from Unstructured, the company behind the open-source unstructured library. It turns PDFs, scans and other business documents into structured output that an application or an AI agent can use. Unstructured announced a rebuilt version on 21 October 2026, with new parsing pipelines and a new generation of trained models aimed at the hardest parts of a document: tables with merged, multi-level headers, forms with dozens of paired fields, and pages degraded by repeated scanning.

What it does

Transform reads a page and keeps its structure. Tables come back as grids with rows, columns and merged cells preserved; forms come back as typed key-value pairs for fields the user defines once, such as the covenants and rates in a credit agreement or the fields of an ACORD form; charts, diagrams and photos get plain-text descriptions; and every value carries its page and location so it can be checked against the original. Heading levels are kept and captions are linked to their figures, which Unstructured says helps an agent find the section it needs.

The maker's approach is to split the page into regions and give each the context it needs, processing a dense financial table on its own rather than as one part of a crowded page.

In its own test on 224 pages of real-world documents, Unstructured reports a table score of 0.896 and a hallucination rate of 5.0% for Transform, against 0.845 to 0.764 and 7.4% to 7.6% for three general-purpose models, while Transform had the lowest text recognition rate of the four (95.8%) and one of the models was cheaper per page. The figures are the vendor's own.

Who it is for

Unstructured names lenders, insurers, laboratories and manufacturers among the teams that use it to pull the terms of a loan, the details of a claim, the results of a lab test or the parts list for a product. A document can be uploaded in the browser and inspected as a formatted page, Markdown or JSON, and a Get Code button produces a Python, TypeScript or cURL snippet that runs the same configuration in an application. It is offered as an API and through MCP.

Key facts

Maker: Unstructured Announced: 21 October 2026 Pricing: 10,000 free pages with no card required; then $0.015 a page on pay-as-you-go; Business plan with custom pricing, multi-user accounts and a dedicated instance, VPC or multi-tenant SaaS Access: browser, API and MCP; also listed on AWS and Azure marketplaces Deployment: SaaS, cloud-hosted dedicated instance, in-VPC on AWS, Azure or GCP, or bare metal (the last three on the Business plan)