Case studyAI / ML
VedaBot
Ancient knowledge. Modern retrieval.
- Year
- 2025
- Focus
- RAG pipeline from scratch.
Overview
A retrieval-augmented chatbot that answers Ayurvedic health questions grounded in real reference books — no hallucinations.
The challenge
Generic LLMs hallucinate on domain-specific health topics. I wanted grounded answers — ones that come from actual Ayurvedic texts, not from the model's imagination. So I built the retrieval layer myself.
How it works
Built a complete RAG pipeline — PDF extraction, text chunking, embedding generation with all-MiniLM-L6-v2, and vector storage in MongoDB.
Achieved 0.65 cosine similarity on relevant passage retrieval, beating the baseline non-RAG approach by 27%.
Ran systematic evaluation comparing RAG vs non-RAG answers against source text — proving retrieval augmentation meaningfully improves accuracy.
Technology stack
- Python
- Gemini
- Sentence-Transformers
- MongoDB
- Streamlit
- PyMuPDF
Highlights
- cosine similarity
- 0.65
- RAG vs non-RAG improvement
- 27%
- pipeline from scratch
- RAG