Turn PDFs, scans and office documents into structured JSON with language models, using local models that need no API key.
This product is available in the build below. Open it for the operating system, cloud and version detail, or read its deployment guide.
| Product | Cloud | Operating system | Version | |
|---|---|---|---|---|
| Unstract on Ubuntu 24.04 | Azure | Ubuntu 24.04 | Standard | View · Guide |
Unstract is an open source, AGPL-3.0 licensed platform for extracting structured data from unstructured documents such as invoices, contracts, statements and forms. In Prompt Studio you describe the fields you want as prompts, check every answer against sample documents side by side, and then deploy the project as a REST API or as an ETL pipeline that reads documents from cloud storage and writes JSON to a database or file store.
It works with hosted models from Azure OpenAI, OpenAI, Anthropic, Google, Mistral and others, or with models running on your own server. It suits finance, operations and data teams that want document automation they can run and audit themselves.
The image works the moment you sign in: a local document parser and two small local models (IBM Granite 4.0 Tiny for answers and Nomic Embed for retrieval, on the CPU through Ollama) are already set up as the defaults, so you can build and test extractions with no API key and add a hosted model later. First boot generates the sign in password, encryption key, database, queue, storage and vector database credentials and the TLS certificate uniquely for each VM, and nothing serves until it has proven the published default login is refused. The databases and parser sit on private networks with no internet access, containers cannot reach the Azure metadata service, product analytics are switched off, and a step by step deployment guide and 24/7 support come with it.