Conversational search allows a visitor to ask a question in natural language instead of constructing a keyword query. An agent interprets the question, checks whether it is within the topics configured for the agent, and returns a concise answer based on the matching website content.
The answer is generated from your website content. It is not a general-purpose assistant and should not be expected to answer questions that are outside the agent’s configured topics or the information available on the website.
How an agent answers a question
The following steps describe the answer process at a high level:
- The visitor submits a question.
- The agent validates the question and checks its scope against the topics connected to the agent.
- The agent searches the question-and-answer knowledge created from those topics.
- The agent generates an answer from the relevant content and can provide the sources used.
- If the question is invalid, outside the configured scope, or cannot be answered with sufficient confidence, the agent returns its configured default response instead of guessing.
The exact response can vary as your monitored topics and website content change. Test important questions again after you change an agent or its connected topics.
Conversational search and keyword search
Keyword search returns matching pages or results for the terms entered by a visitor. Conversational search interprets the intent of a question and presents a natural-language answer, with supporting sources where available.
These experiences can complement each other. Use conversational search when visitors want an answer to a question, and provide links to source pages so visitors can verify the answer or continue exploring.
Topics define the scope
Topics are high-level categories that group related pages on your website. An agent uses only the topics connected to it. For example, an agent for a university website might connect to topics such as Admissions, Student accommodation, and Fees.
To change what an agent can answer, update its connected topics rather than adding unrelated instructions to the answer guidelines. See Connect topics to an agent.
Answers, sources, and confidence
A response can include:
- the generated answer
- links to the source pages used to create the answer
- a confidence level or score
- a validation result indicating whether the question was answerable
Treat generated answers as a convenient way to find information, not as a replacement for the source page. Review source links and keep the underlying website content accurate and current.
Single-turn and threaded conversations
An agent can support follow-up questions in the same conversation. A follow-up uses the conversation history as context, so a visitor can ask a related question without repeating all the details.
A new conversation starts when the conversation is reset or the connection is lost. Your website implementation should make this clear to visitors and should not assume that conversation history is available after a new connection is opened.