WHAT IS IT?
Unstract takes unstructured documents - PDFs, scans, images, DOCX, spreadsheets - and turns them into clean JSON using LLMs. You describe what to extract in plain English inside Prompt Studio, then ship it as a REST API or as an ETL pipeline feeding your data warehouse. The point: replace per-vendor templates and homegrown regex with a single prompt that absorbs layout variations.
WHY IS IT INTERESTING?
- Prompt Studio, no templates: the extraction schema is written in natural language and tested live against real documents. A new invoice type stops being days of development.
- Two deployment modes: either a REST API you POST a file to and get JSON back, or an ETL pipeline that pulls documents from a bucket and loads results into your database.
- LLM provider agnostic: OpenAI, Anthropic, Bedrock, Gemini, Mistral, Ollama locally. Swap models without rewriting prompts, and stay on-premise if you need to.
- Connectors already wired: S3, MinIO, GCS, Azure Blob, Google Drive, Dropbox, SFTP on the way in; Snowflake, BigQuery, Redshift, PostgreSQL, MySQL, SQL Server on the way out.
- MCP server and n8n node: extraction is exposed to AI agents via Model Context Protocol, or drops straight into an existing n8n workflow.
- Self-hosted in one command:
./run-platform.shbrings up the whole Docker Compose stack. Django, Celery, FastAPI, PostgreSQL, Redis and RabbitMQ, no manual assembly.
USE CASES
- Automatic extraction from heterogeneous supplier invoices into an ERP, with no per-issuer template.
- Processing KYC files and identity documents for compliance, with auditable JSON output.
- Ingesting contracts and insurance policies into a warehouse for analysis.
- Digitizing medical reports or tax forms into a structured database.
- Backing a custom AI agent that queries a document corpus through the MCP server.
GOOD TO KNOW
The core is AGPL-3.0, but several accuracy and operations features are reserved for the Cloud and Enterprise tiers: dual-LLM verification (LLMChallenge), SinglePass extraction to cut token costs, a human-in-the-loop review interface, SSO and RBAC. The open core is production-usable - you just need to know where the line sits. Budget 8 GB of RAM minimum.