From Amnesia to Awareness: Giving Retrieval-Only Chatbots Memory

Retrieval-only Q&A chatbots often fail at multi-turn conversations because they treat each query in isolation, lacking memory of the conversational context. This stateless design breaks down when users ask natural follow-up questions, forcing them to repeat full context in every query. The proposed solution is to implement a query rewriting layer using a Large Language Model (LLM). This LLM analyzes the current query along with the recent conversation history to rewrite ambiguous follow-ups into specific, standalone questions. This hybrid approach enables natural, contextual dialogue while preserving the content control and accuracy of a retrieval-based system grounded in a verified knowledge base.