A user asks an open-ended question after “How can I help you?” The assistant fails to recognize it, requests a rephrase, and risks frustration or immediate opt-out. Conversational search addresses this failure by supporting broader natural-language questions instead of relying only on predefined intents. section 11.3.1
Conversational search retrieves information through dialogue and returns a conversational answer. It can search a knowledge base and use an LLM to generate an answer, rely on an LLM’s built-in knowledge, or combine search with predefined intents in a hybrid design. section 1.1.3
Choose the interaction pattern according to the breadth and predictability of user questions:
- Use intent-and-response matching for recurring, well-defined questions. section 1.1.3
- Use conversational search or retrieval-augmented generation for broad domains and question-answering use cases. section 11.3.1
- Use a hybrid model when common questions deserve intents while less predictable questions route to search or generative AI. section 1.1.3
- Keep training relevant to current user needs so the assistant does not become stale or repeatedly ask users to rephrase. section 11.3.1
From fixed intents to broader questions
A store-hours request may use a simple API that returns the same answer for everyone. A broader company question can instead search a knowledge base and generate a response; frequent questions can remain mapped to intents while other questions use search. section 1.1.3
