All projects

OCRFlow

The open-source canvas for document AI — drag open OCR models onto a canvas, wire them together, and turn any PDF into validated JSON on your own hardware.

WebsiteSource Code
OCRFlow
Ready-to-run nodes
30+
OCR engines
7
Files per API call
50
Cost per page
Free

Overview

A closed cloud API, or a pile of glue scripts

Document AI today usually means choosing between a metered cloud OCR service that sees every invoice, contract and medical record you send it, or a stack of one-off scripts wrapped around a single library.

OCRFlow puts the open models and the tooling around them in one place: a canvas to build and inspect pipelines, and an API to run them in production — all on your own machines.

Typed wires, no silent breakage

Every node does one thing and every port carries a type — PageArtifact, Region[], TextLine[], JSON. Nodes only connect when the output actually fits the next input, so broken pipelines are caught while you draw them, not in production.

Best model per step
layout with Docling, text with Surya, tables with Paddle, fields with Qwen
Inspect every step
layout boxes, recognized lines with confidence, final JSON
Schema-validated output
every LLM/VLM response checked before it moves on
Ship headless
promote a canvas to a pipeline and POST documents to it

Your documents stay home

All default models are open-weight and run locally. NVIDIA, AMD and Apple Silicon GPUs are detected automatically, engines that aren't running show up greyed out in the palette, and the whole stack can run fully air-gapped.

Product tour

How it works

A lightweight control plane orchestrates independent model services. The gateway forwards typed requests to whichever engine owns the model, and each engine loads its weights only when first used — you run only what you need.

  1. Canvas · REST API

    build visually, call the same graph headless

  2. Control plane

    FastAPI gateway, Celery worker, Postgres, Redis

  3. Model services

    Surya, Docling, Paddle, Liquid, Unlimited-OCR, Ollama

Built with

  • Next.js 16 + React Flow
    the typed canvas
  • FastAPI gateway
    never imports PyTorch
  • Celery + Redis
    batch jobs and runs
  • PostgreSQL
    projects, pipelines, traces
  • Docling · Surya · PaddleOCR
    layout, text, tables
  • Ollama · LFM2.5-VL · hosted LLMs
    schema-validated extraction

Run it yourself

Docker (recommended)
$ git clone https://github.com/baselhusam/OCRFlow.git && cd OCRFlow
$ cp backend/docker/.env.example backend/docker/.env
$ make up
$ make db-migrate
OCR engines
$ make ocr-up        # every engine, GPU auto-detected
$ make ocr-surya     # ...or one at a time
$ make detect        # which accelerator overlay is used

Details

Status
Active — v0.1.0
Started
Jul 2026
Domain
Document AI / Developer Tools
License
MIT
Hosting
100% self-hosted
Hardware
NVIDIA · AMD · Apple Silicon · CPU
Deploy
Docker Compose · air-gapped
Access
REST API + keys
baselhusam/OCRFlow