What it costs ​

Two people pay for an answer, and they pay in different currencies. You pay requests, tokens and bandwidth. Your reader pays attention — and gets back something they can check.

This page is the arithmetic behind both, with every figure read off this site's own index and this project's own defaults rather than estimated.

What a reader downloads ​

Retrieval runs in the reader's browser, so the corpus has to get there. Opening the panel fetches the index once and caches it:

The vectors are one sign bit per dimension, which is what keeps that number in kilobytes: at float32 the same blob would be 13.4 MB. A reader who never opens the panel downloads none of it at all.

That is the cost nobody budgets for and the one that scales with your corpus rather than with your traffic. It is worth knowing before you index a thousand pages.

What a question costs you ​

A turn is one embedding request plus one or more model calls:

Measured at the shipped maxIterations: 2 with an 8192-token context: 5.9k prompt tokens and 0.7k output per turn.

A follow-up costs no more than a first question. It is scored on two queries — itself, and itself glued to what it follows — and both texts ride a single embedding request rather than buying a vector each. The same holds when the antecedent reaches two turns back, which is what keeps and on Docusaurus? from retrieving nothing.

Where the money actually goes ​

At the model, not at the retriever. One query embedding is about ten tokens and 212 ms, on the order of a dollar a year at ten thousand questions a day; the corpus embedding is a build cost paid once, and a content-addressed cache means editing one page re-embeds that page rather than the corpus.

Measured on the rebuild that added this very page: of 580 chunks, 561 vectors came back from the cache and 19 were bought — one batched request. The empty-state openers cost nothing at all that build, because their answers were written by hand and the questions had not changed.

So when you multiply anything by your provider's price list, multiply the 5.9k and the 0.7k. Everything else on this page is rounding.

What you are not paying for ​

There is no server. No vector database, no ingestion queue, no retrieval service, nothing to keep warm and nothing to page anybody about at three in the morning. The index is a static directory beside your HTML, and the only backend a production deployment needs is a proxy that attaches your API key — plus an endpoint of your own if you want to collect votes.

That absence is most of the economics. It is why the numbers above are the whole bill and not the visible part of one.

What the reader gets for it ​

An answer they can check. Every citation marker resolves to a chunk this host put in front of the model on that turn, verified against a set the host keeps — an id the model invented is stripped, a link to a route that is not in the index is de-linked, and an answer with no surviving citations is not shown at all. See what it guarantees.

Their own language over your corpus. The dense channel is not a ranking refinement, it is the entire cross-language capability: a reader typing Russian at English documentation has no keyword overlap to fall back on, and without an embedder that reader is refused outright.

A conversation that stays a conversation. A follow-up is composed against the question it follows, and against the one that followed when it was itself an ellipsis — so a chain of and on React?, and on Docusaurus? keeps its subject. The turn is primed with the sections the previous answer cited, including in a thread reopened from the archive. Measured on this site's own corpus, reaching the second hop moves recall@8 from 0.750 to 0.900 on the questions that need it and leaves every other question untouched.

Nothing of theirs leaves the browser. Conversations live in the reader's own localStorage — twenty of them, capped at 512 KB — and are never sent anywhere. The one thing that travels is a vote, and only when you have configured an endpoint for it. See conversation history.

Answers, not spinners. The search itself is local, so the only network the reader waits on is the embedding and the model.

What raises the bill ​

• maxIterations. Every accumulated observation is re-sent on every step, so a turn's cost grows with the square of its steps. At 20 the worst case is roughly 138k tokens for one question.
• A pool that rotates. A model that refuses, returns nothing, or times out costs its request anyway, and the next member is tried.
• Site function tools. Each call is a search step: one more request, and every later step re-sends its reply. A repeated call in the same turn is free, and actions cost nothing, because they ride the answer call. See Tools and actions.
• An off-topic question. guard.mode ships as 'off', because no single threshold survives every language a site's readers type in — so today an unanswerable question still spends a request finding out. Calibrate a threshold for your corpus and set 'calibrated', and it costs zero model calls instead.