🎯 Hallucination-Resistant Retrieval
Standard vector search frequently misses exact part numbers, model codes, and acronyms inside dense engineering manuals. DocMindAI fuses dense semantic vectors with sparse lexical keyword matching.
⚡ Hybrid Retrieval Architecture: FAISS + BM25 + RRF
- Dense Vector Search: On-disk FAISS index using
all-MiniLM-L6-v2(384-dimensional semantic embeddings). - Sparse Keyword Matching:
rank_bm25index to catch exact part numbers and serial codes. - Reciprocal Rank Fusion (RRF): Merges dense and sparse rankings with reciprocal rank math, preventing hallucinations and surfacing exact matches.
- Interactive CLI Demo: Includes
demo_hybrid_retrieval.pyfor side-by-side benchmarking of vector vs. keyword vs. hybrid fusion.
🛡️ Production Capabilities
- Exact Document Scoping: Restricts query context to user-selected PDF manuals or whole libraries.
- SQL Analytics: Natural language questions translate into exact SQL analytics against metadata tables.
- Containerized: Production Dockerfile and
.dockerignorefor instant container deployment.
💻 Tech Stack & Repositories
- Frameworks & Search: Python, Flask, LangChain, FAISS, rank_bm25.
- LLM Synthesis: Claude 3.5 Haiku API with strict citation grounding.
- DevOps: Docker, pytest (6/6 passing tests).
- GitHub Repository: docmindai-rag-chatbot