DocuMind

Upload a document, ask questions, get answers with cited sources.

Built as a production-grade RAG system. Proper chunking, hybrid retrieval, reranking, and citations.

Live: https://docu-mind-h-tech.vercel.app

How it works

Upload
  └─ extract text (PDF/DOCX/images/Excel/PPTX)
  └─ chunk with overlap
  └─ embed via OpenAI
  └─ store in Pinecone

Query
  └─ embed question
  └─ vector search (Pinecone) + BM25 keyword search (in-memory)
  └─ merge candidate pools, rerank top results
  └─ generate answer via OpenAI GPT-4o-mini
  └─ return answer with inline citations

Hybrid search matters: pure vector search misses exact keyword matches ("Python", "clause 4.2"). BM25 catches those. Both pools get merged and reranked before the LLM sees them.

Stack

Layer	Choice
LLM	OpenAI GPT-4o-mini
Embeddings	OpenAI text-embedding-3-small (1536d)
Vector DB	Pinecone (serverless)
Reranker	Cohere (BM25 fallback if no key)
Document parsing	PyMuPDF, python-docx, pandas, python-pptx
API	FastAPI + uvicorn
Frontend	React (Create React App)
CI	GitHub Actions (ruff + pytest + docker build)
Backend deployment	Render
Frontend deployment	Vercel

Running locally

With Docker:

cp .env.example .env   # fill in your keys
docker-compose up

Frontend at localhost:3000, API at localhost:8000.

Without Docker:

pip install -r requirements.txt
cp .env.example .env   # fill in your keys
python documind_main.py

API at localhost:8000, Swagger docs at /docs.

Required environment variables:

GENERATION_OPENAI_API_KEY=sk-...
PINECONE_API_KEY=pcsk_...
VECTOR_DB_PROVIDER=pinecone

Optional (for Cohere reranking):

COHERE_API_KEY=...
RERANKER_PROVIDER=cohere

API

Method	Path	Description
GET	/health	Health check
POST	/api/upload	Upload a document (file or text)
POST	/api/query	Ask a question about a document
GET	/api/documents	List uploaded documents
DELETE	/api/documents/{id}	Delete a document
GET	/docs	Swagger UI

Upload a file:

curl -X POST https://documindhtech-1.onrender.com/api/upload \
  -F "file=@report.pdf"

Query:

curl -X POST https://documindhtech-1.onrender.com/api/query \
  -H "Content-Type: application/json" \
  -d '{"document_id": "...", "question": "What are the key findings?"}'

Deployment

Backend (Render): render.yaml is pre-configured. Connect the repo, set the env vars below as secrets, and deploy.

GENERATION_OPENAI_API_KEY=...
PINECONE_API_KEY=...
VECTOR_DB_PROVIDER=pinecone

Frontend (Vercel): frontend/vercel.json is pre-configured. Import the repo, set Root Directory to frontend/, and deploy. No env vars needed.

Chunking

RecursiveCharacterTextSplitter with 1000-char chunks and 150-char overlap (~15%). Overlap keeps sentence context intact across boundaries, which matters for citation accuracy.

Project structure

├── documind_main.py          # Entry point — FastAPI app, mounts router + SPA
├── services/
│   ├── routes.py             # API router (/api/*)
│   ├── api_service.py        # Pipeline orchestrator (upload + query)
│   ├── chunking_service.py   # RecursiveCharacterTextSplitter wrapper
│   ├── cloud_vector_service.py  # Pinecone / Weaviate / Qdrant / Supabase
│   ├── reranker_service.py   # Cohere / BM25 reranker
│   ├── citation_service.py   # Inline citation mapping
│   ├── chat_service.py       # OpenAI chat wrapper
│   ├── simple_gemini_embedding_service.py  # OpenAI embeddings
│   └── text_extract.py       # Multi-format extraction
├── frontend/                 # React app
├── tests/
│   └── test_main.py          # pytest suite (mocked, no credentials needed)
├── Dockerfile.backend        # Multi-stage Python image (non-root)
├── Dockerfile.frontend       # Multi-stage nginx image
├── docker-compose.yml
└── render.yaml

CI

GitHub Actions runs on every push to main:

1. Backend: ruff lint → pytest → docker build
2. Frontend: eslint → jest → docker build

Tests are mocked — CI passes without API keys.