DeepSeek

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

Witdem has no dedicated DeepSeek client adapter. The verified path uses Haystack's OpenAI-compatible generator with DEEPSEEK_API_KEY and the normal Haystack integration. Direct clients can use the generic provider wrapper.

from witdem_sdk.integrations.generic import instrument

call = instrument(
    existing_deepseek_call,
    operation_name="deepseek.chat",
    provider="deepseek",
    model="deepseek-v4-flash",
    observe_result=lambda response: {
        "response_model": response.model,
        "input_tokens": response.usage.prompt_tokens,
        "output_tokens": response.usage.completion_tokens,
    },
)

This assumes those fields match your existing client response; adjust observe_result explicitly if they do not.

Evidence and cost

Provider aliases and deepseek- model prefixes normalize to deepseek. The bundled catalog contains deepseek-v4-flash and deepseek-v4-pro; the legacy deepseek-chat and deepseek-reasoner IDs explicitly alias to V4 Flash. A different model or snapshot is unmeasured until the catalog is extended or the provider reports cost. Rates participate in the automated, reviewable refresh.

Proof and limitations

Product Factory's mixed-v1 and DeepSeek profiles exercise real DeepSeek calls through Haystack. There is no standalone DeepSeek tutorial or native SDK proxy in this release, so support remains experimental.