Web search & fetch
Search through Brave, read pages, get structured live data, and answers with numbered sources.
Web Search
Web Search uses the Brave Search API with your own key. Add it in Settings → Built-in Tools → Web Search → Brave Search API Key (how to get one). When a question needs current information the model searches by itself; the results become the citations in its answer.
What one search brings back
A search asks Brave's LLM Context endpoint for the pages it would ground the query on, with the text already pulled out of each one. The model gets about ten sources, numbered, and at most two from one site: results of one site mostly say the same thing. When ten are not enough, the model searches again with a sharper query, or reads one source in full with Web Fetch.
Numbers run through the whole conversation. If the first search of a chat
gave sources 1–10, the next one starts at 11, and a [3] later in the chat
still means the third source of the first search.
The model also chooses, per search:
| Choice | When the model uses it |
|---|---|
| Strict | A follow-up when the first results were noisy or off-topic: only pages that clearly match are kept |
| Images / videos | You asked for pictures or videos, or it is building something visual |
Structured data
For some questions Brave has more than pages: the data itself. Every search asks Brave whether it has any; Brave decides, and when it does, the data arrives as the first source of the search. It does for:
- weather and forecasts
- stocks and market indices, cryptocurrencies, exchange rates
- sports scores and fixtures — football, American football, basketball, baseball, ice hockey, cricket, Formula 1
- calculations, unit conversions, Unix timestamps, definitions, parcel tracking
The source is named after the provider Brave took the data from, and it holds the figures as they are — today's high and low, an index's price and its intraday series, the score and the scorers — rather than a page's sentence about them. The model reasons from those numbers and cites them like any other source, and when it draws a chart or a dashboard as an artifact, it draws it from them.
Try "how is the KOSPI doing today?" or "chart the weather in Beijing this week".
Pictures and videos
Ask for pictures or videos ("show me photos of …") and they come back as a gallery and video cards under the answer. Each picture belongs to the page it was found on, so the citation under it is that page. A picture from a page the search did not read is a source of its own, marked as such.
The gallery shows the thumbnail Brave keeps, never the site's own file, so a site that refuses outside requests, or a picture that is gone, does not leave a hole.
Options
Under the tool's row in Settings → Built-in Tools:
| Option | Effect |
|---|---|
| Country | Bias results towards a region |
| Languages | Keep results in the language you choose (Brave uses the first one) |
| Max Results | Sources a search brings back, 1–20 (default 10) |
| Deep recall | Searches three phrasings plus a broader pass — more sources, more tokens, slower |
| Recency Filter | Past 24 hours, week, month or year (past hour searches the past 24 hours) |
| Domain Filter | Comma-separated; prefix with - to exclude (nature.com, .edu, -reddit.com) |
What a search costs
Brave counts requests. A search is two: the pages, and the question of whether Brave has structured data for it. The rest are added only when they are used:
| Part of a search | Requests |
|---|---|
| The search | 2 |
| Structured data | +1 when Brave has data |
| Pictures, videos | +1 each |
| Deep recall | 6 in all (three phrasings, each twice) |
Deep recall also asks your chat model for the two extra phrasings. Brave's free plan answers one request a second; Actovia waits and tries once more when it is told to slow down.
Sources in a long chat
A search's sources stay whole for the reply that made them and the next one, so a follow-up can dig into the same results. After that the model keeps one line per source — its number, title and link — and can read any of them back in full when it needs to, without searching again. That is what keeps a long chat with many searches from filling up the model's context.
When a search fails
A search that fails is not "no results". When Brave rejects the key, the plan's quota is used up, the rate limit is reached or Brave cannot be reached, the model is told which, and says so — a wrong key is something to fix in Settings, not a topic nobody has written about.
Web Fetch
Give the model a link ("summarise https://…") and it reads the page itself — an article, documentation, an API response, a PDF or a file on GitHub. A fetched page is a source like a search result and can be cited.
Pages are read through Jina Reader, which renders pages built with JavaScript; when it cannot, Actovia reads the page itself and keeps its main content.
Web Fetch reads only public addresses: it will not reach your local network or Actovia's own API, and every redirect is checked again before it is followed.