Unstract – LLM Document Extraction Platform

TL;DR

TL;DR: Unstract is an open-source platform that uses LLMs to extract structured JSON from unstructured documents (PDFs, images, scans) via natural language prompts, deployable as REST APIs or ETL pipelines.

Source and Accuracy Notes

• Project page: unstract.com
• Source repository: github.com/Zipstack/unstract
• License: AGPL-3.0 (verified via GitHub API)
• Documentation: docs.unstract.com
• Source last checked: 2026-06-17 (6,663 stars, 632 forks on GitHub)

What Is Unstract?

Unstract is an open-source platform that turns unstructured documents into structured JSON data using large language models. Instead of writing regex patterns or building vendor-specific templates, you define extraction schemas with natural language prompts and deploy them as APIs or ETL pipelines.

The platform targets teams in finance, insurance, healthcare, KYC/compliance, and logistics who need to process large volumes of documents like invoices, bank statements, contracts, and tax forms.

“Unstract uses LLMs to extract structured JSON from documents — PDFs, images, scans, you name it. Define what you want to extract using natural language prompts, and deploy as an API or ETL pipeline.” — from the official README

Key Features

Prompt Studio — Define document extraction schemas with natural language. No regex or template coding required. The agentic version builds schemas, crafts prompts, and validates accuracy automatically.

API Deployment — Send a document over REST API, get structured JSON back. Integrates with n8n for workflow automation.

ETL Pipeline — Pull documents from folders (S3, GCS, local), process them, and load results to your data warehouse.

MCP Server — Connect Unstract to AI agents (Claude, GPT, etc.) via Model Context Protocol for agentic document processing.

Human in the Loop — Built-in verification interface for reviewing and correcting LLM extractions before final output.

LLM Challenge — Benchmark and compare extraction accuracy across different LLM providers on your specific documents.

Setup Workflow

Step 1: System Requirements

• Linux or macOS (Intel or M-series)
• Docker and Docker Compose installed
• 8 GB RAM minimum
• Git

Step 2: Quick Start (5 minutes)

# Clone the repository
git clone https://github.com/Zipstack/unstract.git
cd unstract

# Start the platform
./run-platform.sh

The script pulls all Docker images and starts the platform. Visit http://frontend.unstract.localhost and login with:

• Username: unstract
• Password: unstract

Step 3: Configure LLM Provider

Unstract supports multiple LLM providers:

• OpenAI (GPT-4, GPT-3.5)
• Anthropic (Claude)
• AWS Bedrock
• Ollama (local models)
• Any OpenAI-compatible API

Add your API keys in the platform settings under “Adapters”.

Step 4: Create Your First Extraction

1. Open Prompt Studio from the dashboard
2. Upload a sample document (PDF, image, or scan)
3. Define what to extract using natural language (e.g., “Extract invoice number, date, total amount, vendor name”)
4. Test the extraction on your sample
5. Deploy as API endpoint or ETL pipeline

Step 5: Deploy as API

# After creating your extraction prompt in Prompt Studio,
# click "Deploy" to get a REST API endpoint

curl -X POST https://api.unstract.localhost/api/v1/extract \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -F "[email protected]" \
  -F "prompt_id=your-prompt-id"

Response:

{
  "invoice_number": "INV-2024-001",
  "date": "2024-01-15",
  "total_amount": 1250.00,
  "vendor_name": "Acme Corp"
}

Architecture Overview

┌────────────────────────────────────────────────────────────┐
│                          Unstract                          │
├─────────────┬─────────────┬─────────────┬──────────────────┤
│  Frontend   │   Backend   │   Worker    │ Platform Service │
│  (React)    │  (Django)   │  (Celery)   │   (FastAPI)      │
├─────────────┴─────────────┴─────────────┴──────────────────┤
│                      Cache (Redis)                         │
├────────────────────────────────────────────────────────────┤
│              Document Storage (S3/MinIO/GCS)               │
└────────────────────────────────────────────────────────────┘

The platform runs as four Docker containers:

• Frontend: React UI for Prompt Studio and management
• Backend: Django API for user management and configuration
• Worker: Celery workers for async document processing
• Platform Service: FastAPI service for extraction orchestration

Deployment Options

Docker Compose (Self-Hosted)

# Default deployment
./run-platform.sh

# Specific version
./run-platform.sh -v v0.1.0

# Upgrade to latest
./run-platform.sh -u

# Build images locally
./run-platform.sh -b -v current

Managed Cloud

Unstract offers a managed cloud version with:

• No infrastructure management
• Automatic updates
• SLA guarantees
• Enterprise support

Check pricing at unstract.com/pricing.

Kubernetes

For production Kubernetes deployments, refer to the Helm charts documentation.

Security Notes

⚠️ Backup your encryption key!

Unstract encrypts adapter credentials (LLM API keys, cloud storage credentials) using an encryption key generated at first run. If you lose this key, all configured adapters become inaccessible.

Find the key in backend/.env or platform-service/.env:

ENCRYPTION_KEY=your-encryption-key-here

Store this key securely offline. There is no recovery mechanism.

Data Privacy

• Documents are processed in-memory and not retained after extraction (unless you configure persistent storage)
• LLM API calls send document content to your chosen provider (OpenAI, Anthropic, etc.)
• For sensitive documents, use local LLMs via Ollama
• Self-hosted deployments keep all data within your infrastructure

Access Control

• Role-based access control (RBAC) for team deployments
• API key scoping per prompt/endpoint
• Audit logging for all extraction requests

Practical Evaluation Checklist

When to use Unstract:

• ✅ You process 100+ documents per month with varying formats
• ✅ You need structured JSON output, not just OCR text
• ✅ Document schemas change frequently (new vendors, new form types)
• ✅ You want to avoid writing/maintaining regex patterns
• ✅ You need both API and batch ETL capabilities
• ✅ You want to use multiple LLM providers and compare accuracy

When to look elsewhere:

• ❌ You process simple, fixed-format documents (use traditional OCR/IDP)
• ❌ You need real-time extraction with sub-second latency (LLM calls take 2-10s)
• ❌ You have strict on-prem requirements with no internet access (unless using local LLMs)
• ❌ Your budget is very limited (LLM API costs add up at scale)

Cost considerations:

• Open-source platform is free (AGPL-3.0 license)
• LLM API costs: ~$0.01-0.05 per document (depends on provider, document size, model)
• Managed cloud: starts at $X/month (check current pricing)
• Infrastructure: 8GB RAM minimum, 2 CPU cores recommended

FAQ

Q: What document formats does Unstract support? A: PDFs (text-based and scanned), images (PNG, JPG, TIFF), and any format that can be converted to an image. The platform uses OCR for scanned documents and images.

Q: Can I use my own LLM or local models? A: Yes. Unstract supports OpenAI, Anthropic, AWS Bedrock, and any OpenAI-compatible API. For local deployment, use Ollama to run open-source models like Llama 2, Mistral, or Code Llama.

Q: How accurate is the extraction? A: Accuracy depends on document quality, LLM provider, and prompt design. The platform includes LLM Challenge to benchmark accuracy on your specific documents. Typical accuracy ranges from 85-98% for well-structured documents.

Q: Is there a limit on document size? A: The platform handles documents up to the token limits of your chosen LLM. For GPT-4, that’s ~128K tokens (roughly 300 pages of text). Larger documents can be split automatically.

Q: Can I deploy Unstract without Docker? A: The official deployment uses Docker Compose. Manual installation is possible but not officially supported. You need Python 3.10+, Node.js 18+, Redis, and PostgreSQL.

Q: What’s the difference between AGPL-3.0 and the enterprise edition? A: AGPL-3.0 is fully functional for most use cases. The enterprise edition adds SSO/SAML, advanced RBAC, priority support, and managed cloud hosting. Check the editions comparison for details.

Q: Does Unstract work with n8n or other automation tools? A: Yes. Unstract provides a native n8n node and REST APIs that integrate with any automation platform (Zapier, Make, custom workflows).

Conclusion

Unstract fills a real gap in the document processing space by making LLM-based extraction accessible without requiring ML expertise. The natural language prompt interface dramatically reduces the time to onboard new document types from days to minutes.

For teams drowning in manual data entry from invoices, contracts, or compliance documents, Unstract offers a practical path to automation. The open-source AGPL-3.0 license makes it accessible for startups and small teams, while the managed cloud option serves enterprises needing SLAs and support.

The MCP server integration positions Unstract well for the emerging agentic AI workflow — your AI agents can now process documents as part of larger automated workflows.

Next steps:

• Try the quick start guide
• Join the Slack community for support
• Star the GitHub repo if you find it useful