Many RAG pipelines are designed for single-turn questions: embed the current query, retrieve chunks, generate an answer. Real users do not interact that cleanly. They ask follow-ups, use pronouns, change scope, correct themselves, and switch topics.

Conversational RAG adds state around retrieval. The system keeps recent conversation history, rewrites ambiguous follow-ups into standalone search queries, and decides when history helps or hurts retrieval.

Conversation makes the retrieval problem harder. You now have to manage context windows, memory, query rewriting, topic shifts, stale assumptions, and grounded answers over multiple turns.

In this chapter, we will explore how to design RAG systems that support multi-turn conversations, so users can ask follow-ups naturally and still get grounded answers.