When it searches
The AI model decides when to search, the way it decides to use any of its tools. It reads your message, chooses what to look up, and searches, even halfway through an answer. A question it can answer well from what it knows, like a rewrite or a sum, gets no search.
The Search button at the left, under the message box, turns searching on or off for the chat, and stays as you left it. Send without searching, under the arrow beside the send button, sends one message with no search. Settings › Answers › Web search turns it off everywhere.
How much it searches
There's no number of searches or pages to set, and no limit on them. The model keeps searching and opening pages until it can answer with confidence: a question with one answer usually takes one search, and a comparison or a roundup takes as many as it needs.
Searches the model asks for together run at the same time, and each search reads its best pages side by side, so a question that needs a lot of looking up still comes back quickly. A long run of searching counts toward the answer's steps. When it reaches them, Keep going carries on from where it stopped.
Where the results come from
Each search asks several sources at once and merges what comes back:
- SearXNG, which gathers results from ordinary search engines without an account or tracking. Mosaik keeps a list of public servers, remembers which ones answer quickly, and skips the ones that refuse.
- DuckDuckGo and Bing.
- Google News and Bing News, for what's happening now and for reviews and explainers on any subject.
- Stack Overflow for programming, Hacker News for tech, arXiv and Semantic Scholar for research papers, YouTube for videos, and Wikipedia for what something is, when the question belongs to them.
- Mojeek, as a last resort when nothing else answered.
- Fast web search: Parallel's search engine, reached through OpenRouter on your key. It answers in about a second with the passages of each page that are about your question, so the answer starts sooner. Each search costs about $0.0013, shown as Web search in your costs. It is on by default; Settings › Answers › Web search › Fast web search turns it off, and the free sources answer on their own.
The best pages are then read side by side, and only the passages about your question are handed to the model. Mosaik only fetches ordinary web addresses, never anything on your own network.
With Smart set to Fastest, Faster or Quick and Fast web search on, the model searches with OpenRouter's own web search inside the same request, so a simple lookup is one request. The pages it finds are the answer's sources. This costs about $0.005 for an answer that searches, billed with the answer. With Smart's Cost at Free, Mosaik uses only the free search engines, so a chat costs nothing.
Asked to find something specific, like a video, a post or a document, Mosaik searches for it by the details you give, opens the likeliest result to check it matches, and searches again with other words or on the site itself before saying it couldn't find it. A YouTube link is read from the video's own page: its title, channel, date, length, views and description.
Live prices, rates and weather
A share price, an exchange rate or tomorrow's forecast changes too often for a search result to hold it. When a question asks for one, Mosaik asks the services that publish the figure, with no key or account:
- Yahoo Finance for shares, funds, indexes and currency pairs, with the time of the quote.
- Frankfurter, the European Central Bank's daily reference rates, for currency pairs Yahoo Finance doesn't quote.
- gold-api.com for gold, silver, platinum, palladium and copper.
- CoinGecko for coins.
- Open-Meteo for the weather now and for the next seven days. When it doesn't answer, Nominatim from OpenStreetMap finds the place and MET Norway gives the forecast.
Only the thing asked about is sent: the currencies, the company or fund, the metal, the coin or the place, never the question itself. Each figure is cited like any other source, linked to the service it came from. Turn Live prices, rates and weather off in Settings › Web search and the answer uses the ordinary search results alone.
Pictures in answers
Ask what something looks like, or for photos of a place, and the answer can show the pictures themselves. The AI searches Bing Images and Wikimedia Commons, looks over what comes back, and puts the ones that fit in its answer. Only the pictures an answer shows are kept with the chat.
The answer lays its pictures out to suit what it says, for pictures it found and pictures it made alike: a grid, one large picture as wide as the answer, a row you scroll sideways with arrows at its sides, small tiles four across, or cards that hold each picture in a box with its words. A picture can have a short title and a line of text under it.
Click any picture in a chat to open it large. An arrow shows when the pointer is near the left or right side, and the ← and → keys step too, through every picture in the chat: the ones you sent, the ones the AI made and the ones it found. In the picture's top corner are Edit with AI and a … menu, which saves a copy, opens it in another app, or opens the page a picture from the web came from. Turn Pictures from the web in answers off in Settings › Web search to keep the AI from searching for pictures.
Search by picture
When an answer means naming something specific in a picture, like a painting and its artist, a place, a product or model, a plant or a dog's breed, Mosaik first searches the web with the picture itself. Yandex Images and TinEye, asked at once, find the pages that show it, and the answer names it from those pages, not from a guess. If nothing matches, the answer says it couldn't find it. This works whether or not your model can see pictures, and the answer's work line shows each picture search.
The picture is only sent when a question asks for it. Turn Search by picture off in Settings › Web search and only your AI model sees your pictures.
How sources are shown
A paragraph or list item that rests on the web ends with a chip carrying the logos of the sites it cites. Point at the chip to light up the text it covers and see the exact page: its title, the site, the passage, and whether the page was read in full or only its summary.
The work line above the answer lists every search in order, with its query, its results and the pages that were read. Sources are saved with the message, so an old chat shows the same chips when you open it again.
Site logos are fetched once, only for the sites an answer cites, from each site itself and kept on your computer, so no logo service learns what you read. Search results the answer does not cite are never contacted for a logo.
Using your own SearXNG
Public SearXNG servers are shared and sometimes slow or gone. If you run your own, put its address in Settings › Answers › Web search. It's tried first and never rate-limited. The quickest way to start one is Docker:
docker run -d -p 8080:8080 searxng/searxngThen use http://localhost:8080. Under Advanced you can keep
every search inside your SearXNG by turning off the other engines, and Test
search runs a real query and says who answered and how fast.
Models that can't use tools
Search works best with models that can call tools, because they can decide to look something up halfway through an answer. Models that can't still get searches: before the answer starts, a quick call to the same model decides what to search for, and Mosaik searches again for anything the results leave out. While one of those models is selected, a Limited web search note shows in the bottom corner. Click it to open the model browser, where those models are marked No tools.
When the web can't be reached
If no search engine answers at all, the model is told the web is unreachable and answers from what it knows, saying what it couldn't check. If engines answer but about something else, the model is told to try again with better keywords. The work line says which happened.