DocuMind

An open source document intelligence API. Upload any PDF and ask questions — DocuMind finds the most relevant information, reranks it for accuracy, and returns a cited answer with source page numbers.

Real world use cases

• Law firm uploads contracts → asks "what are the termination clauses?"
• Student uploads textbook → asks "explain Newton's third law"
• Company uploads HR policy → employees ask questions about leave rules

Architecture

PDF → pdfplumber (load + chunk) → HuggingFace embeddings → Qdrant vector store → Cohere reranking → Groq LLaMA 3.3 → answer + citations

Tech Stack

• Python, FastAPI
• LangChain — LLM application framework
• Qdrant — vector database (local, no Docker needed)
• HuggingFace — free local embeddings (all-MiniLM-L6-v2)
• Cohere Rerank — improves retrieval accuracy
• Groq API — free, fast LLM inference (LLaMA 3.3 70B)

Why this stack

• Qdrant over ChromaDB => production-grade, better performance
• Cohere reranking => separates basic RAG from accurate RAG
• HuggingFace embeddings => free, no API cost, runs locally

API Endpoints

• GET /health — health check
• POST /upload — upload a PDF and process it
• POST /query — ask a question, get answer + sources

How to Run

1. Clone the repo
2. Create virtual environment: python -m venv venv
3. Activate: venv\Scripts\activate
4. Install: pip install -r requirements.txt
5. Copy .env.example to .env and add your API keys
6. Run: uvicorn api:app
7. Open http://127.0.0.1:8000/docs to test

API Keys Required (all free)

• GROQ_API_KEY — console.groq.com
• GOOGLE_API_KEY — aistudio.google.com
• COHERE_API_KEY — dashboard.cohere.com

Known Limitations

• Scanned/image PDFs not supported (no OCR)
• Table extraction is basic (complex tables may lose structure)
• No hybrid search (vector + keyword) — planned improvement
• Single user only - concurrent access requires Qdrant server mode