Conversational Search

Definition

Conversational search is an interaction model where users find information through a multi-turn dialogue rather than a single query. The system maintains context across turns, asks clarifying questions, and refines results based on ongoing conversation.

Distinct from single-query search in that the query is not a single string — it’s a conversation history.

Core Challenges

Coreference and Context Carryover

“Show me laptops under $1000" \to " Wha t ab o u t o n es w i t h 16 GBR A M ? "— t h eseco n d q u ery im pl i c i tl yre f ers t o t h e f i rs t f i lt erse t . T h esys t e mm u s t reso l v e " o n es " t o " l a pt o p s u n d er$ 1000.”

Query Rewriting

Each user turn must be rewritten into a standalone query before retrieval. LLMs have made this tractable:

Conversation history + new turn → LLM → self-contained query → retrieval

Mixed Initiative

Both user and system can take initiative:

User-driven: user narrows down step by step
System-driven: system asks clarifying questions (Query Understanding → clarification dialogues)

Architecture Patterns

RAG-Based Conversational Search

RAG with conversation history in the LLM context. User’s question + prior turns → query rewriting → retrieval → LLM answer generation. The dominant pattern for LLM-powered search.

Faceted Dialogue

System presents Faceted Search options as conversation turns: “Did you mean laptops for gaming or for work?” — structured clarification.

Agentic Search

Search agents that decompose complex questions into sub-queries, run multiple searches, and synthesize answers across turns.