RAG-Based Conversational Product Catalog Management: Architecture, Structured Output as FSM, and Empirical Evaluation

Retrieval-Augmented Generation (RAG) has emerged as a principled solution to the knowledge limitations of large language models (LLMs). We design and implement a complete RAG pipeline built on LangChain, ChromaDB with HNSW indexing, BAAI/bge-small-en-v1.5 embeddings, and LLaMA 3.3 70B deployed on Groq LPU infrastructure. Our central contribution is the formalization of structured LLM output as a finite-state machine (FSM) controller for guided product attribute collection. We report Precision@1 of 75%, average end-to-end latency of ~979ms per dialogue turn, and a 58 percentage point improvement over a BM25 lexical baseline.