Guides
Written from engineering practice rather than from a content calendar. Where a guide names a number, it is sourced; where it recommends something, it says when not to.
Verified 21 September 2026
Agent retrieval
Web search for AI agents
Agent retrieval inverts the assumptions of consumer search: precision matters more than recall, metadata beats full text, and absence has to be explainable.
5 sections · 3 questions answered
Build a research agent
A research agent that searches, reads, and searches again spends most of its budget rebuilding structure the index already has. Graph traversal collapses that loop.
6 sections · 3 questions answered
Multilingual search for agents
One embedding space for 101 languages means a query in one language retrieves relevant passages in another — no translation, no per-language index.
5 sections · 3 questions answered
Trust & accuracy
Stop agents hallucinating from search results
Retrieval reduces hallucination without eliminating it. Two structural fixes, corroboration counting and explainable absence, reach what prompting cannot.
4 sections · 3 questions answered
Auditable retrieval for regulated work
In regulated work the question after an answer is what it missed. Four capabilities make that answerable, and ranking quality is not one of them.
5 sections · 3 questions answered
Measuring an index before you depend on it
Coverage claims are unverifiable by construction. Here is a measurement you can run yourself in an afternoon, on any provider, using your own URLs.
7 sections · 4 questions answered
Getting recency right in agent retrieval
Four different dates hide behind the word recent, and filtering on the wrong one is how a correct retrieval stack returns confidently outdated answers.
7 sections · 4 questions answered
Grounding a sovereign model without sending the question abroad
A model kept inside a jurisdiction that grounds itself through a search API outside it has moved the question, not the data. What a search call discloses, what a provider can and cannot promise, and how to keep the record inside the boundary.
5 sections · 4 questions answered
Cost
Reduce the token cost of web-searching agents
Most of what an agent spends on retrieval is not the search call — it is the context consumed by results it did not need and the reasoning spent rejecting them.
6 sections · 3 questions answered
Search API pricing explained
Credits, per-request fees, per-token charges and processor tiers make headline rates non-comparable. Here is how to normalise them.
5 sections · 3 questions answered
Migration
Choosing a web search API
Five questions that decide the choice: index ownership, response shape, trust signals, cost at your volume, and what happens when the vendor changes.
6 sections · 3 questions answered
Retrieval that survives a provider shutting down
Three providers changed terms or disappeared in eighteen months. Teams that moved cheaply had drawn one boundary in advance, and it was not an HTTP wrapper.
7 sections · 4 questions answered
How these are written
Every number about a competitor comes from that vendor's own page or a funding announcement, with the date it was read recorded next to it. Where a vendor does not publish a comparable figure, we say so rather than estimating one. Nothing here cites a benchmark we did not run or a customer we cannot name, and the multiplier claims — how many times cheaper, how many times faster — are computed from the underlying data at build time rather than typed into a sentence where they would slowly become false.
The corollary is that a guide will sometimes tell you not to use this API. The comparison pages name where a competitor genuinely wins, and the migration pages list what you lose by moving. A guide that cannot do either is advertising.
Where to start
If you are choosing a provider, start with the five questions that actually decide it and then run the coverage measurement against your own URLs before you compare prices — a coverage gap is the one finding that cannot be tuned away afterwards. If you are already running an agent and it costs too much, the token-cost guide is the one with the largest immediate effect, because retrieval spend is dominated by context the agent read and rejected rather than by the search call itself.
If it is answer quality rather than cost, the two structural fixes for search-induced hallucination and the piece on what current actually means cover the failure modes that prompting cannot reach. Each guide links theglossary terms it relies on, so you can follow a definition rather than guess at one.
Put it into practice
10,000 free credits a month, no card. Everything in these guides works on the free tier.