OCR

OCR is a full-stack application that ingests PDF documents, runs them through an asynchronous OCR pipeline, and delivers structured archives back to the user.

Production app: ocr.tuturu.io

What The Project Does

• Authenticated document upload
• Process creation and live status tracking
• PDF split into pages
• OCR transcription on generated page images
• Post-processing and downloadable archive delivery

Architecture

The repository is a pnpm monorepo split into a few clear areas:

• apps/web: the single application server — React, Vite, TanStack Start / Router / Query UI and the server layer (server functions and route handlers) that calls the business services in memory
• workers/*: queue-driven pipeline workers for splitting, transcription, post-processing, and cleanup
• packages/*: shared business services, infra utilities, and common types
• db: Drizzle schema, migrations, and database tooling
• doc: RFC notes

There is no separate API process: the TanStack Start server owns Postgres, Redis, S3, and RabbitMQ access directly.

Stack

• Frontend: React, TypeScript, Vite, TanStack Router, TanStack Query, Radix UI
• Server: TanStack Start server functions and route handlers (Node.js, TypeScript)
• Database: PostgreSQL with Drizzle ORM
• Messaging and cache: RabbitMQ, Redis
• Object storage: Garage / S3-compatible storage
• Tooling: pnpm, Biome, Docker Compose

End-To-End Flow

1. A user uploads a PDF from the web app.
2. The Start server creates a process and stores the source file.
3. A split worker turns the PDF into page images.
4. OCR workers transcribe the pages.
5. A post-processing worker assembles the final result.
6. The browser receives live process updates over SSE (GET /api/processes/status) and exposes download actions when the run is complete.

Local Development

Install dependencies:

pnpm install

Prepare environment variables from one of the provided templates:

• .env.exemple for local development
• .env.docker.example for Docker-based runs

WEB_ALLOWED_HOSTS can be used to add extra hostnames accepted by vite preview in the web container. Use a comma-separated list when internal probes or reverse proxies access the web app with hostnames that differ from localhost or BETTER_AUTH_URL.

Start local infrastructure:

docker compose up -d

Bootstrap the local Garage bucket after the first start, or after wiping the garage-data volume:

pnpm garage:init

This imports S3_ACCESS_KEY / S3_SECRET_KEY, creates S3_BUCKET, and grants read/write access to that key.

Run the full app in development mode:

pnpm dev

pnpm dev first builds the runtime packages, then starts the web server and all workers concurrently.

Useful Commands

pnpm dev
pnpm build
pnpm lint
pnpm db:generate
pnpm db:migrate
pnpm garage:init

Docker And Deployment

Development infrastructure is defined in docker-compose.yaml.

Production services are defined in docker-compose.prod.yaml:

docker compose --env-file .env.docker \
  -f docker-compose.prod.yaml \
  up -d --build --remove-orphans

Main exposed production ports:

• Web / TanStack Start application server: 3010
• PostgreSQL: 5436

Production Redis, RabbitMQ, and object storage are provided by the shared stack in ../infra. OCR containers join the external redis-network, rabbitmq-network, and garage-network, and connect to redis-prod, rabbitmq-prod, and garage-prod.

Provision the production Garage key and bucket before the first deployment:

../infra/garage/provision-project.sh ocr-prod ocr-prod "$PWD/.env.docker"

Set REDIS_OCR_USERNAME=ocr, copy the dedicated REDIS_OCR_PASSWORD from the secret vault, and configure AMQP_URL with the existing OCR RabbitMQ user and /ocr vhost in .env.docker.

The reverse proxy should target only the web container on 127.0.0.1:3010. There is no backend upstream anymore. Keep BETTER_AUTH_URL aligned with the public HTTPS origin and add any internal proxy/probe hostnames to WEB_ALLOWED_HOSTS as a comma-separated list.

Observability And Docs

• Project notes: doc/rfc.md

The web app exposes /metrics under the ocr_web_ prefix; if it is served through vite preview, make sure the probe hostname is allowed through WEB_ALLOWED_HOSTS when needed. Monitoring dashboards and scrape configuration are handled separately.

License

This project is licensed under the ISC License.

• Local license file: LICENSE
• SPDX reference: ISC