How matching behaves
Search is fast — matches return in less than 0.5 seconds — and returns zero, one, two or three agents, depending on how confident the match is.
What comes back
- One agent when the query clearly belongs to that agent.
- Two or three agents when descriptions overlap or the query is ambiguous. The best match comes first.
- No agents when the query does not carry enough information, or lies outside every agent's description. The user sees that nothing matched. A wrong agent does not run.
What the search box does with the results is the search behavior. By default the best match runs at once. Other settings offer two or three matches as choices, or show the results without running anything.
What it matches on
Matching is semantic: a query does not have to share words with a description. Search works from the meaning of the description, and from the situations, needs and outcomes the description implies. A description written in terms of what the user gets therefore matches better than a description written in terms of how the agent works.
Follow-ups
A follow-up query in a conversation is searched too, so a user can move to a related agent in the middle of a conversation. By default the top match runs. When nothing matches, the previous agent answers again. Other follow-up settings offer the matches as choices with the previous agent among them, or keep the whole conversation on the agent that started it.
From inside an agent
The search command gives agents the same search. An orchestrating agent can find the right specialist and invoke that specialist.