Mistral

Status: experimental provider path, live-validated through Haystack Product Factory.

Witdem has no dedicated Mistral client adapter. The verified path uses MistralChatGenerator inside Haystack and the normal Haystack integration. Direct clients can use the generic provider wrapper.

from witdem_sdk.integrations.generic import instrument

call = instrument(
    existing_mistral_call,
    operation_name="mistral.chat",
    provider="mistral",
    model="mistral-small-2603",
    observe_result=lambda response: {
        "response_model": response.model,
        "input_tokens": response.usage.prompt_tokens,
        "output_tokens": response.usage.completion_tokens,
    },
)

Adjust the observer to the actual client response shape.

Evidence and cost

mistral, mistralai, and common Mistral model prefixes normalize to mistral. The bundled catalog includes selected Mistral Small, Medium, Large, Ministral, and Codestral model IDs. See Pricing catalog for the exact allowlist and refresh process.

Product Factory exercises Mistral through Haystack in live matrices. There is no standalone Mistral tutorial or native client proxy in this release.