How kyryl-opens-ml/no-ocr works
Project type and purpose
• This is a full-stack AI-driven document exploration tool.
• Frontend: React/TypeScript app (“no-ocr-ui”).
• Backend: FastAPI service (“no-ocr-api”) for document ingestion, Hugging Face-style dataset creation, and vector search ingestion.
• LLM microservices: Python scripts under “no-ocr-llms” deployed via Modal.
• Data stores: local PDF storage, LanceDB vector store, Supabase for auth and metadata.
File structure
• Top-level directories:
• no-ocr-ui — React UI, component hierarchy, Supabase config, Vite build
• no-ocr-api — FastAPI code, data ingestion modules, tests, Dockerfile
• no-ocr-llms — standalone scripts for loading and serving vision-language models
• Config files: docker-compose.yml orchestrates UI, API, LanceDB, Supabase emulator
• CI pipelines under .github/workflows validate tests and linting
• Tests under no-ocr-api/tests exercise API endpoints and ingestion logic
• pyproject.toml and requirements.txt define backend dependencies
README for additional insights
• Key technologies: FastAPI, LanceDB (via SearchClient), HF_Dataset, Qwen2-VL, Supabase, Modal, Docker.