The problem: The Gita has 701 verses. Finding applicable wisdom for a specific situation requires either deep familiarity or hours of reading.
How it works: 1. User describes their ethical dilemma 2. Query is embedded using sentence-transformers 3. ChromaDB retrieves top-k semantically similar verses 4. LLM generates structured output: 3 options with tradeoffs, implementation steps, verse citations
Tech stack: - Backend: FastAPI, PostgreSQL, Redis - Vector DB: ChromaDB with all-MiniLM-L6-v2 embeddings - LLM: Ollama (qwen2.5:3b) primary, Anthropic Claude fallback - Frontend: React + TypeScript + Tailwind
Key design decisions: - RAG to prevent hallucination — every recommendation cites actual verses - Confidence scoring flags low-quality outputs for review - Structured JSON output for consistent UX - Local LLM option for privacy and zero API costs
What I learned: - LLM JSON extraction is harder than expected. Built a three-layer fallback (direct parse → markdown block extraction → raw_decode scanning) - Semantic search on religious texts works surprisingly well for ethical queries - Smaller models (3B params) work fine when constrained by good prompts and retrieved context
GitHub: https://github.com/geetanjaliapp/geetanjali
Happy to discuss the RAG architecture or take feedback.
