Contexta

Contexta is a structure-aware, retrieval-augmented Q&A system designed to deliver high-precision answers with exact document citations.

Features

• Document ingestion with layout-aware parsing
• Semantic chunking with rich metadata
• Vector storage and retrieval via Qdrant
• Hybrid retrieval with reranking
• LLM-powered answer generation with guardrails
• End-to-end RAG pipeline
• Evaluation framework for retrieval and generation quality
• React frontend for document upload, chat, and source inspection

Architecture

backend/
  app/
    ingestion/   → load, parse, and layout documents
    chunking/    → split documents into chunks with metadata
    vector/      → embeddings and Qdrant vector store
    retrieval/   → retriever and reranker
    llm/         → LLM client, prompts, and guardrails
    rag/         → end-to-end RAG pipeline
    models/      → Pydantic models
    evaluation/  → retrieval/generation evaluators
frontend/
  src/
    components/  → Upload, Chat, Sources
tests/
  test_ingestion.py, test_chunking.py, test_vector.py,
  test_retrieval.py, test_rag.py

Tech Stack

• Backend: FastAPI, Python
• Vector DB: Qdrant
• LLM: OpenAI / compatible providers
• Frontend: React
• Testing: pytest

Getting Started

# Backend
cd backend
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate
pip install -r requirements.txt
uvicorn app.main:app --reload

# Frontend
cd frontend
npm install
npm run dev

API Endpoints

• POST /api/upload — upload documents
• POST /api/chat — send a query and get a RAG answer
• GET /api/sources/{query_id} — retrieve cited sources

Testing

pytest tests/ -v

License

MIT