Dharma RAG

An AI-powered Q&A application that retrieves relevant passages from the Bhagavad Gita and generates grounded responses through a Flask API backed by PostgreSQL.

General-purpose models answer scriptural questions without reliable source grounding. The goal was a focused AI system that retrieves relevant text first, then generates answers from that context.

I built a RAG pipeline using SentenceTransformers for semantic search and LLaMA 3.1 for response generation. Flask APIs expose query and retrieval endpoints, with PostgreSQL storing documents and supporting efficient lookup.

Text is embedded with SentenceTransformers and stored in PostgreSQL. At query time, the question is embedded, similar passages are retrieved, and LLaMA 3.1 generates a response from the selected context. Flask handles routing, request validation, and API responses.

• Semantic search over Bhagavad Gita text
• Context retrieval with SentenceTransformers
• Response generation with LLaMA 3.1
• Flask APIs for query and document retrieval
• PostgreSQL for document storage

• Retrieval quality for verse-level text

  Passages needed to stay semantically complete when chunked. I tuned embedding and retrieval to return coherent context for generation.

• Connecting retrieval and generation cleanly

  The API had to pass retrieved context into LLaMA 3.1 in a consistent format so responses stayed tied to source material.

• PostgreSQL-backed document flow

  Indexed documents and retrieval results needed a reliable storage layer. PostgreSQL kept document data and API access in one place.

• SentenceTransformers for semantic search

  Used embedding-based retrieval to match natural-language questions to relevant Gita passages.

• LLaMA 3.1 for grounded generation

  Generation runs on retrieved context rather than open-ended prompting alone.

• Flask + PostgreSQL API layer

  Flask exposed retrieval and generation endpoints with PostgreSQL handling document storage and lookup.

• RAG quality depends heavily on retrieval before generation.
• A clear API boundary makes it easier to iterate on embedding and model choices.
• PostgreSQL works well as the persistence layer for document-backed Q&A.

• Citation display for retrieved verses
• Improved chunking for longer passages
• Streaming responses from the API