← Back to projects
SGI Handbook RAG Chatbot
Hybrid retrieval — dense + sparse vectors; math-aware grounding, not just API glue.
High-precision QA over the SGI Driver’s Handbook — hybrid retrieval to cut hallucinations and retrieval latency.
Manual PDF structuring for vectors, FAISS + TF-IDF with RRF, Cross-Encoder re-ranking, sub-2s responses.
Mission
Remove friction from searching dense driver handbooks with a context-aware, high-precision QA system.
Innovation
Avoided the classic LLM hallucination trap with a hybrid retrieval pipeline instead of “chat the PDF” only.
Technical deep dive
- Manually curated and indexed the handbook PDF into structured text for better vector search.
- Retrieval: FAISS dense search + TF-IDF sparse search combined with Reciprocal Rank Fusion (RRF).
- Re-ranking: Cross-Encoder on top-k chunks — cutting manual lookup time by about 60%.
- Stack: Python, LangChain, Llama-3.1, FastAPI.
- Python
- LangChain
- FAISS
- Llama-3.1
- FastAPI
- RRF