text-extract-api
FastAPI service to convert documents into structured JSON or Markdown using local vision models and remove PII.
Extract tables, text, and JSON from insurance documents using local vision LLMs and PyTorch OCR.
This tool is a self-hosted API service for extracting structured data and removing PII from scanned documents and PDFs. Insurance technical teams can process claims, invoices, and medical records entirely on-premise using Ollama and PyTorch vision models. It provides a queue-backed architecture designed for high-throughput backend automation.
What it is
It converts PDFs, images, and Office files into clean Markdown or JSON format while handling tabular data and specialized forms. The core service pairs EasyOCR or local vision LLMs with Celery workers and Redis caching to execute extraction pipelines without sending sensitive data outside your network.
Who it's for
It suits insurance developers and data engineers who need to parse sensitive claim forms and anonymize customer records on local hardware. It does not suit non-technical teams looking for a ready-made graphical user interface out of the box.
Key features
Local vision OCR processing
Runs PyTorch OCR models and local LLMs through Ollama to keep sensitive insurance records inside your environment.
Automated PII redaction
Extracts document text while identifying and stripping out personally identifiable information before downstream storage.
Asynchronous queue with Celery
Uses Celery workers and Redis caching to process high-volume document batches concurrently.
Flexible output parsing
Converts complex tables, invoices, and reports into formatted Markdown or structured JSON schemas.
Remote engine integration
Supports external OCR wrappers like Marker-pdf via API endpoints for complex multi-language layouts.
Up and running in 3 steps
Our honest verdict
What's great
- Zero external cloud dependencies keep sensitive claim files strictly on local infrastructure.
- Redis caching prevents redundant processing when running repetitive LLM extraction tasks.
- Supports both standard OCR engines and vision-capable LLMs like Llama 3.2 Vision and MiniCPM-V.
- Includes a dedicated command-line interface for testing prompt files against complex documents.
Where it falls short
- Requires dedicated hardware or GPU resources to run vision LLMs with acceptable performance.
- Setting up workers, Redis, and Ollama models requires significant DevOps and Python expertise.
- Using GPL-licensed OCR engines like Marker requires separate server deployment to avoid license contamination.
- Llama 3.2 Vision strategy can be very slow when processing large multi-page document batches.
Pricing & license
The software is free to use under the MIT license. The only cost comes from operating the compute servers and local GPU resources required to process documents.
Open remote AI-training roles
880 open remote AI-training roles, updated November 15, 2026; most pay $20–$175 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
Use it, or have us set it up
FAQ
Can I run this without sending data to external cloud providers?
Yes, the complete pipeline including PyTorch OCR, Ollama models, Redis, and FastAPI runs entirely on your own local machine or private server.
How does the tool strip sensitive client data?
You can pass custom prompt files during submission to instruct the integrated LLM to scrub personally identifiable information before returning the JSON output.
Can it parse complex tables in medical or claims reports?
Yes, the underlying vision models and OCR strategies convert complex tabular data into structured Markdown or JSON.
