Unstructured review

It's worth being precise about what this product is, because the category matters. Unstructured splits documents into elements, paragraphs, titles, tables, so a search or retrieval system can index them, and its platform is built around moving those elements from source connectors to destination connectors. It doesn't, on its own, return a filled-in schema. To score it fairly we applied the same field-filling step to its output that every other system got, so what we measured is how much of the useful structure survives its parse. On easy pages, enough. On hard ones, not enough.

Where it held up was clean English text. Plain text files, spreadsheets, and single-page English invoices came through nearly intact, and on several of them it matched DocuPipe at the top of the table. It also handled Hebrew and Arabic text direction better than its overall rank would suggest. If your documents are mostly English prose and simple tables and your goal is search rather than extraction, nothing in our results argues against it.

Where it fell down was everywhere structure had to be kept rather than read. On long registers and catalogs the row structure didn't survive the parse, and the extraction had nothing to rebuild from. Chinese and Japanese scans came back nearly empty. A one-page table that every other system read perfectly came back with nothing, and so did an engineering drawing whose data lived in a title block. In each case the text may well have been captured, but the shape of the document wasn't. That, plus the fact that you have to supply the extraction step yourself, is why it scores low on resilience for this job.

The price is honest and the terms are clean. Parsing costs a cent and a half a page after 10,000 free pages, no card required, pay-as-you-go with no minimum. Zero data retention is stated on every plan, and SOC 2 Type 2, HIPAA, and GDPR are all on the pricing page, with dedicated instances and VPC on the Business plan. Just remember this is the price of parsing, not extraction. To get structured fields out you add your own model step, and its cost and engineering are yours.

The product is a real platform for what it's designed to do. The hosted UI builds workflows from source connectors through partitioning, chunking, and enrichment to destination connectors. There's a Python SDK, a REST API, an MCP server, workflow and job management endpoints, and the open-source library that a great many teams start with. What it doesn't have is a schema editor, a review screen, or an extraction assistant, because that isn't the job.

So if what you need is structured fields out of a document, this is the wrong tool for the task, and its score reflects that mismatch more than any lack of quality. If what you need is a parse to feed a search index, our test doesn't answer that question, and you shouldn't read it as if it did.