NutritionRAG — Grounded Nutrition Chatbot

Full RAG pipeline with a web app — hybrid BM25 + ANN retrieval over a textbook corpus stored in pgvector, with inline citations.

Self Project · Sept – Oct 2025

Overview

A grounded nutrition Q&A chatbot that answers questions using a real textbook corpus rather than relying on parametric LLM knowledge — reducing hallucination and adding source traceability.

System Design

  1. Ingestion: nutrition textbook chapters semantically chunked into passages; embeddings stored in pgvector
  2. Retrieval: hybrid BM25 + ANN search over the corpus — BM25 for keyword precision, ANN for semantic recall; scores fused with Reciprocal Rank Fusion
  3. Generation: retrieved passages passed as context to an LLM; answers generated with inline citations and source snippets
  4. Web App: clean interface for asking nutrition questions and viewing sourced answers

Why Hybrid Retrieval

Pure dense retrieval misses exact-match keywords (nutrients, dosage values). Pure BM25 misses semantic paraphrases. Hybrid gets both — this is the core motivation behind my thesis work on multi-expert retrieval too.

Stack

Python · pgvector · PostgreSQL · BM25 · FAISS · Hugging Face · Flask