docRefinery vs Unstract

Prompts written by engineers, or layouts reviewed once

Unstract, from Zipstack, is an LLM document extraction platform released under AGPL-3.0. Its centre is Prompt Studio, where you describe each field in natural language, test the prompts against sample documents and deploy the result as an API or an ETL pipeline into a warehouse. LLMWhisperer, its text preparation service, keeps layout and reads checkboxes and handwriting before an LLM sees the page.

docRefinery takes a different route to the same output. Its artefact is not a prompt someone maintains. Each new layout is reviewed once, stored as configuration your organisation governs, and recalled whenever that sender's documents come back. Remembered, not relearned.

Where Unstract is the better choice

• Self-hosting without a sales call. The code is on GitHub under AGPL-3.0. An engineering team can run it on its own hardware this afternoon. Our on-premises option is an Enterprise contract.
• Model choice. Unstract lets you pick and compare LLMs, and its LLMChallenge page recommends pairing models from different providers so that one checks the other.
• Pipelines into a warehouse. ETL pipelines that land extracted data in a data warehouse are part of the product. We return data through an API and export to JSON, CSV or XLSX; webhooks and integrations are in design and not yet available.
• Compliance on Enterprise. Unstract lists SOC 2 Type II, ISO 27001, HIPAA and GDPR with a DPA, plus SSO/SAML and audit logs, on its Cloud and On-Prem Enterprise tiers. Its Starter and Growth tiers do not include them. docRefinery's SOC 2 Type II audit is in progress with no report yet, and SSO is planned for Enterprise but not available yet.
• Low-cost text preparation. LLMWhisperer runs from $1 to $15 per 1,000 pages depending on mode, with 100 pages a day free.

What docRefinery does that Unstract does not, or does differently

docRefinery fits when no one wants to maintain prompts, and when control and audit matter. These are the differences against Unstract.

• A workflow from one sample. Drop one document and every field it prints is found, ready to keep, relabel or switch off, so nobody starts from a blank prompt.
• You define the output. Each field has a plain-language instruction, a "Return as" name you choose as its output key, a type and a required flag, in a form an operations lead can edit.
• One output format for every document. Every field returns text, alias, field_confidence and location, and every table returns rows, so one parser handles every document type.
• Layout memory. Each layout is learned once and remembered, and re-trained when a form changes. For purchase orders or invoices from the same suppliers every month, that is what keeps review effort falling.
• A ready catalogue. 20 document types and 785 field definitions to start from, and custom types trained, hosted and re-trained at no charge.
• Rules, versions and replay. Business rules such as line items adding up to the total run on every document, and every change is a version you can replay against past runs before it goes live.
• Governed fields. Owner, admin and member roles decide what may be extracted, and a caller cannot widen it.
• Null, scored and cited. A field missing from the page comes back null, and each value carries a confidence score and a page and box. Anything below its threshold waits for a person.
• Corrections that become fixes. A reviewer's correction becomes a suggested fix, tested on that layout's own documents before anyone applies it.
• Long files, segmented. Long, multi-page files are handled at scale and separated into the documents inside them.

A note on cost

Unstract publishes its prices: Document Studio Starter at $499 a month with 5,000 pages and $0.10 a page beyond that, Growth at $2,249 a month with 25,000 pages, and Enterprise on quote. The open-source edition costs nothing to license, though you run the models and infrastructure. docRefinery is free up to 100 pages a month, and paid plans are custom, quoted by sales to your volume.