DocuMind

A RAG-based document intelligence system. Upload a document, ask questions, get answers with inline citations.

---

How it works

Upload:
  document → text extraction → chunking (1000 tok / 150 overlap)
           → Gemini embeddings → Pinecone vector DB

Query:
  question → embed → vector search (top-10) → rerank (top-5)
           → Gemini LLM → answer with [1][2] citations

---

Architecture

┌─────────────┐        ┌──────────────────────────────────────┐
│   Browser   │        │              Backend (FastAPI)       │
│  React SPA  │◄──────►│  /api/upload   /api/query   /health  │
└─────────────┘        │                                      │
      │                │  services/                           │
   nginx               │  ├── chunking_service.py             │
   proxy               │  ├── simple_gemini_embedding_service │
                       │  ├── cloud_vector_service.py         │
                       │  ├── reranker_service.py             │
                       │  ├── chat_service.py                 │
                       │  └── citation_service.py             │
                       └──────────────────────────────────────┘

---

Run locally

Prerequisites: Docker, Docker Compose, API keys

cp .env.example .env
# fill in at minimum: GEMINI_API_KEY, PINECONE_API_KEY, COHERE_API_KEY

docker-compose up

Service	URL
Frontend	http://localhost:3000
Backend	http://localhost:8000
API docs	http://localhost:8000/docs

---

Run without Docker

# Backend
pip install -r requirements.txt
python documind_main.py

# Frontend (separate terminal)
cd frontend && npm install && npm start

---

Tech stack

Layer	Tech
Backend	FastAPI, Python 3.11, uvicorn
Frontend	React 18, served via nginx
Embeddings	Google Gemini embedding-001 (768d)
Vector DB	Pinecone (primary) - Weaviate / Qdrant / Supabase supported
Reranker	Cohere Rerank-3 - Jina / Voyage / BGE supported
LLM	Gemini 2.0 Flash (primary), OpenAI GPT-4 (fallback)
Chunking	LangChain RecursiveCharacterTextSplitter

---

Environment variables

See .env.example. Minimum required to run the full pipeline:

GEMINI_API_KEY
PINECONE_API_KEY
COHERE_API_KEY
VECTOR_DB_PROVIDER=pinecone
RERANKER_PROVIDER=cohere

---

API endpoints

Method	Path	Description
POST	/api/upload	Upload document or text
POST	/api/query	Query with a question
GET	/api/documents	List uploaded documents
GET	/health	Health check
GET	/docs	Swagger UI