Agentic Search

Agentic search is a retrieval strategy in which a model issues a search, reads the results, reformulates the query, and decides for itself when it has gathered enough evidence to answer. Classic RAG makes one retrieval call per question with a fixed top-k and hands whatever comes back to the generator; agentic search makes an unbounded number of calls, and the sequence of queries is chosen at runtime by the model rather than fixed in the pipeline.

It is worth separating this from agentic extraction, which is a different job that often runs in the same system. Agentic extraction is about pulling structured fields out of a page once you have decided to read it. Agentic search is about deciding which pages to read at all. A pipeline can do either without the other, and the failure modes do not overlap: bad extraction gives you a wrong value from the right page, bad search gives you a confident answer built from the wrong pages.

The mechanism is a tool-use loop. The search API is exposed to the model as a callable function, the model invokes it, sees the results, and invokes it again with a narrower phrasing, a different entity name, or a follow-up question that only became obvious after the first round. Multi-hop questions are where this earns its cost. "Which of this company's investors also backed its closest competitor" cannot be answered by any single query, because the second half of the question depends on the answer to the first. The corresponding risk is that the stopping condition is a judgement call by the model, so runaway loops are a real production failure. Most teams shipping this cap the number of steps and set a per-request spend limit.

The tension that decides both quality and cost is cached index versus live fetch. Vendors running their own crawl, such as Exa and Parallel, serve results from a stored index, which is why sub-second latency is achievable and why the marginal query is cheap. The cost is freshness bounded by recrawl cadence. Live fetching a page at query time is current but slow and billed per document. Most AI search APIs meter these separately, which is the detail that surprises people on their first invoice: Exa publishes distinct line items for search requests and for page contents, Tavily charges one credit for a basic search and two for an advanced one, and Linkup publishes €5 per 1,000 standard searches with a higher deep tier above it.

The practical consequence is that per-call pricing stops being a useful comparison. Under agentic search a single user question becomes as many billed searches and content fetches as the model chooses to make, so the unit that matters is cost per resolved user question, measured on your own traffic. Run that measurement before committing, cap the loop, and log the query sequence. When an answer turns out wrong, the query trace is usually the only artifact that explains why.