In short. Mistral's built-in Websearch returns source URLs and titles in tool_reference chunks inside message.output.content.
Definition: As of 17 October 2026, Mistral's built-in Websearch returns source links in
tool_referencechunks interleaved with answer text inside amessage.outputentry. Each reference carries a sourceurlandtitle.
Key facts
- Mistral's Websearch documentation supports the built-in tools through the Conversations API and Agents API, and excludes Chat Completions.
- Mistral's Web Search with References cookbook shows how to inspect built-in search responses through
response.outputs. - The Citations & References guide separately demonstrates custom function results whose source records are retrieved using
reference_ids.
Where does the source URL appear?
The Websearch output explanation describes message.output.content as a list containing text and reference chunks. A reference's type is tool_reference; tool identifies web_search, title names the source, and url holds its address. The documentation also lists a source field.
A separate tool.execution entry records the search execution. Its metadata includes the tool name, an identifier and timestamps. Reading only that entry misses the links embedded in the generated answer. An application displaying the answer therefore needs to process the content chunks, preserving the references alongside the text.
How can an application inspect the response?
The built-in search section of Mistral's Web Search with References cookbook starts a conversation with tools=[{"type": "web_search"}]. It supplies a model directly, so this example begins without creating an agent first.
For an existing response, the following adaptation of its output loop prints each complete serialized output entry. The cookbook truncates each printed entry; removing that truncation makes the full content available for inspection.
for output in response.outputs:
print(output.model_dump_json())
The relevant fields to inspect are the output's type and, for a generated message, the chunks in content. For a website audit, keeping that serialized response preserves the returned source addresses together with the answer being examined.
Why do other examples use reference IDs?
Mistral's Citations & References guide demonstrates a custom get_information function. It returns JSON source records containing URLs, titles and snippets. The application adds that result to the chat history as a tool message before requesting the final answer.
In that example, the formatter collects ReferenceChunk.reference_ids, looks up each identifier in the function result, and prints the matching title and URL. These IDs belong to the custom tool's reference data. The built-in Websearch response described above exposes URLs directly in its tool-reference chunks. A parser needs to follow the format used by its chosen API workflow.
Related terms
FAQ
Can I enable built-in Websearch through Chat Completions?
Mistral's Websearch documentation places web_search and web_search_premium in the Conversations and Agents APIs. It states that Chat Completions responses lack the search-result references required by those tools.
How do I turn custom reference IDs into links?
Follow the lookup in Mistral's Citations & References example: collect the IDs from reference chunks, resolve them against the tool's JSON result, and use the corresponding url and title fields.
Sources
- Mistral AI, "Websearch", undated, accessed 17 October 2026
- Mistral AI, "Web Search with References", undated, accessed 17 October 2026
- Mistral AI, "Citations & References", undated, accessed 17 October 2026
Check how a page reads to answer engines. Open the GEO score analyzer