Agentic Search
Searching iteratively, judging results, and searching again.
Retrieval where the model runs a search, reads the results, decides they are insufficient and searches again with a better query — rather than receiving one set of results up front. The loop is the point.
It buys quality on genuine research questions and costs latency, because each round trip is a model call plus a network call. The distinction that matters when reading API docs is between a single lookup — cheap, fast, deterministic — and the agentic mode, which is none of those.
The cost profile is what to design around: an agentic search is several model calls plus several network calls, so it belongs behind a user who expects to wait, or in an asynchronous job. For interactive lookups, a single search call is usually the right shape.
Choosing the mode by patience
A chat feature that answers questions about current events.
A single lookup answers most of them in about a second. Reserve the iterative mode for the question that could not be phrased well, and tell the user it is going to take longer — a slow answer with no explanation reads as broken.
- docsWeb search ↗
Distinguishes a quick lookup from agentic search, where a reasoning model searches, judges the results and searches again. Also covers source annotations.
- docsDeep research ↗
Long-horizon research workflows combining web search, MCP, file search and code execution — the shape of a task measured in minutes, not milliseconds.
Tool Call
The model asking for a specific function with specific arguments.
Tool
A function the model is allowed to call to affect the world.
Agent Loop · Agentic loop
Think, act, observe — repeated until the task is finished.
Just-in-Time Retrieval
Fetch what the task needs when it needs it, not everything in advance.
RAG · Retrieval-Augmented Generation
Look things up, then put what you found in the prompt.