How agent search works

An agent cannot answer reliably without finding the right information. This guide follows the path from a question to source-backed evidence and shows where each kind of search tool fits.

Browse the provider directory ↗Read the articles ↗

Search for agents is a system of choices, not a single API call. Some layers find new information. Others organize data you already have, rank candidates, or test whether the answer is supported.

The parts below are a useful mental model. A real product may combine several in one service or skip a layer when the task does not need it.

01Source selection and query planning

Start with the question

Determine what information is needed and how current it must be.

A question about today's market price needs a different source from a question about an internal policy. The agent first needs to identify the subject, required freshness, and whether the answer lives on the public web, in specialist records, or in a private corpus.

Explore the visual explanation
02Web and industry search APIs

Discover sources

Search the web or a specialist database for candidate material.

A web search API can find public pages. Industry search reaches filings, papers, clinical trials, and other records that general web indexes may not cover well. This stage returns places to look, not a verified answer.

Explore the visual explanation
03Crawlers, browser tools, document parsers

Acquire and parse

Turn pages and documents into material the agent can inspect.

Fetch the source, extract its useful text or structure, and retain its URL, publisher, date, and document location. A result snippet is often too thin to support a claim; the underlying document matters.

Explore the visual explanation
04Embedding models, search engines, vector databases

Index a private corpus

Make your own documents searchable before an agent needs them.

Keyword indexes match terms. Embedding models turn text into vectors that can surface semantic similarity. Vector databases and search engines store and retrieve those representations. Many systems combine lexical and vector retrieval rather than choosing one exclusively.

Explore the visual explanation
05Lexical, dense, and hybrid retrieval

Retrieve candidates

Bring back a manageable set of passages or records.

The first retrieval pass favors speed and coverage. Its job is to avoid missing useful evidence while keeping the candidate set small enough to inspect. Filters for source, date, permissions, and document type may matter as much as similarity.

Explore the visual explanation
06Reranking models

Rerank for the task

Order the candidates by their usefulness to this question.

A reranker compares the question with each candidate more closely than a first-pass index usually can. It can improve which passages reach the agent's context, but it cannot make an irrelevant or missing source trustworthy.

Explore the visual explanation
07Provenance, source checks, citation support

Verify and answer

Check the evidence and show where every material claim came from.

The agent should open decisive sources, resolve conflicts, note uncertainty, and attach citations to the claims they support. A citation is useful only when the linked source actually contains the evidence.

Explore the visual explanation
08Retrieval and answer evaluation

Evaluate the whole system

Measure whether the search process improved the final answer.

Test retrieval quality, factual support, freshness, coverage, latency, and cost using representative questions. Inspect failures by layer: a strong answer model cannot recover a source that was never found.

Explore the visual explanation
Keep going

Explore the tools behind each layer.

The directory links to product documentation for search APIs, data sources, retrieval infrastructure, models, and evaluation tools. Entries describe documented capabilities, not a ranking of quality.

Open the provider directory