OllaMan DocsWeb Search
Let agents search the web and read pages for up-to-date answers
Overview
Local models only know what was in their training data. With Web Search, an agent can look things up online and read the pages it finds, so it can answer questions about recent events, current documentation, or anything it doesn't already know.
Web search gives the model two tools:
- web_search: searches the web and returns titles, links, and short snippets
- web_fetch: reads the main text of a web page, such as a result from a search
The model decides when to use them. For questions it can answer on its own, it replies directly without searching.
Works with or without Agent Mode
Web search does not require Agent Mode. In a normal chat, the model gets only the two web tools: it cannot read or write files or run commands.
Turning Web Search On or Off
Find the Search Button
At the bottom of the chat input area, the Search button sits between the Think button and the Agent Mode button.
Toggle It for This Agent
Click Search to turn web search on or off. The button is highlighted when it's on.
Chat as Usual
Ask your question. When the model searches or reads a page, the step appears inline in the reply, showing the search query or page URL. Click a step to expand its results.
On by default, saved per agent
Web search is on by default. The switch is stored separately for each agent, so you can keep it on for a research assistant and turn it off for an agent that should never go online.
Button Hints
An amber warning icon on the Search button means search is on but won't run:
- Provider not set up: the selected provider needs an API key or instance URL. Fill it in under Settings → Agent → Web Search.
- Model doesn't support tool calling: in a normal chat, web search needs a model with tool calling (the Tools capability). Other models answer without searching. Switch to a model that supports tools, or turn on Agent Mode.
How It Works with Agent Mode
| Agent Mode off | Agent Mode on | |
|---|---|---|
| Available tools | web_search, web_fetch | Web tools plus read, write, edit, bash |
| Approval | Web tools run automatically | Follows Agent Mode's Auto Execute setting |
| Model requirement | Tool calling support | Any model (tool calling recommended) |
Choosing a Search Provider
Open Settings → Agent → Web Search to pick the provider. The provider is shared by all agents.
Built-in
Free and needs no setup. Uses DuckDuckGo web results and falls back to Brave. Heavy use may be rate limited.
Ollama
Ollama's official web search and page fetch. Requires an API key from a free ollama.com account.
SearXNG
Your own SearXNG instance. Enter its URL and enable the json format under search.formats in its settings.yml.
Tavily
A search API built for AI agents. Requires a Tavily API key; each search uses credits (Basic: 1, Advanced: 2).
Other Settings
- Results per search: how many results each search returns, from 1 to 10 (default 10). Fewer results keep the context smaller.
- Test: runs one search to check that the provider works.
Open Web Search Settings
Go to Settings → Agent and find the Web Search section.
Select a Provider
Choose Built-in, Ollama, SearXNG, or Tavily. Use Get API key to open the provider's key page when one is needed.
Enter Credentials
Paste the API key or instance URL. It is saved when you leave the field.
Test
Click Test. "Search works." means the provider is ready.
Privacy and Safety
What is sent where
When web search is on, the search queries and page URLs that the model chooses are sent to the selected provider. Your conversation itself is not sent. API keys are stored locally on your device and are never shown in logs.
- Page content is treated as untrusted data. The model is told not to follow instructions found in web pages.
- With the Built-in provider, pages are fetched directly from your device. Local and private network addresses (such as
localhost,192.168.x.x, or.localhosts) and non-standard ports are blocked, so a web page cannot trick the model into reaching services on your machine or network. - Only text pages are read, up to 3 MB each.