In short. Mistral offers web_search and web_search_premium through Conversations and Agents, returning answer text interleaved with source references.

Definition: Mistral web search API refers to Mistral's built-in tools for retrieving web information within an AI assistant's conversation. As of 17 October 2026, Mistral documents web_search and web_search_premium, with source references returned alongside answer text.

Key facts

Mistral's Websearch documentation describes ordinary search through web_search; web_search_premium adds access to news articles through integrated news provider verification.
The agent tools overview places execution of built-in tools in Mistral's environment and supports their use through Conversations or Agents.
The Web Search with References cookbook includes a built-in search example that starts a conversation without first creating an agent.

How is web search enabled?

A tool is a capability an assistant can call while handling a request. Mistral's agent tools overview instructs developers to include built-in tools in the tools parameter when creating an agent or calling the Conversations API. For ordinary web search, the configuration fragment is:

{"tools": [{"type": "web_search"}]}

The supported conversation endpoint is /v1/conversations. Built-in web search is also available through the Agents API. The overview excludes these search tools from the Chat Completions endpoint, /v1/chat/completions.

Mistral's Websearch guide leaves the model free to decide whether to search for a request. Adding the tool makes search available; an application checks the returned execution entries to see whether it ran.

What does a search response contain?

The Websearch guide separates tool.execution, the record of the search operation, from message.output, the assistant's answer. Execution metadata includes the tool name and start and completion timestamps.

The answer's content is a sequence of chunks, meaning separate pieces of message content. Text chunks and tool_reference chunks are interleaved. A reference supplies fields including tool, title, url, and source. This structure lets an application associate answer text with linked material while retaining the search execution record separately.

How can an application display sources?

Mistral's Citations & References guide shows a separate implementation using a developer-defined function that supplies documents. Its formatter handles TextChunk and ReferenceChunk objects, collects reference_ids, and resolves those IDs against the supplied reference data to print titles and URLs.

The Web Search with References cookbook develops this custom workflow with Wikipedia search results and text snippets. Its streaming example maps incoming reference IDs back to stored source records and formats clickable links. For the built-in search integration, the relevant response structure is the tool_reference chunk described above.

FAQ

Do I need to create an agent first?

No. Mistral's agent tools overview allows built-in tools directly in Conversations, including web search.

Why does the citations guide use Chat Completions?

The citations guide demonstrates a developer-defined function returning source material. That example uses Chat Completions; the built-in web_search tools require Conversations or Agents.

Sources

Check how a page reads to answer engines. Open the GEO score analyzer