Provider support¶

Provider support is not one binary capability. Witdem can observe a provider through a dedicated SDK adapter, through a framework such as Haystack, or through standard OpenTelemetry/the generic callable wrapper.

Support matrix¶

Provider	Current path	Evidence	Cost behavior	Status
OpenAI	OpenAI Agents native processor; LangChain/LangGraph/Haystack callbacks; generic wrapper	Unit/integration tests, tutorials, live Product Factory and external Chinook runs	Catalog entries for listed OpenAI models; provider money accepted	Beta native support
Anthropic	Messages client proxy; Claude Agent observer; framework callbacks	Unit/integration tests, tutorials, live Product Factory and external Chinook runs	Catalog entries for listed Claude models; cache tokens supported	Beta native support
DeepSeek	Haystack OpenAI-compatible generator or generic wrapper	Live Product Factory matrix	Catalog entries for `deepseek-v4-flash` and `deepseek-v4-pro`	Experimental provider path
Mistral	Haystack Mistral generator or generic wrapper	Live Product Factory matrix	Catalog entries for `mistral-small-2603` and `mistral-medium-3-5`	Experimental provider path
Azure OpenAI	Generic wrapper or GenAI OTLP; recognized by Haystack	Standalone tutorial and tests	Separate `azure_openai` catalog; deployment aliases supported by override	Experimental provider path
Amazon Bedrock	Generic wrapper or GenAI OTLP; recognized by Haystack	Standalone tutorial and tests	Bundled standard rates for selected Nova, Cohere, and Mistral model IDs	Experimental provider path
Google Vertex AI	Generic wrapper or GenAI OTLP; Google/Gemini recognized by Haystack	Standalone tutorial and tests	Bundled standard rates for selected Gemini IDs	Experimental provider path
Ollama	Generic wrapper or GenAI OTLP	Standalone local tutorial and tests	Normally not measured as API spend; usage can still be recorded	Experimental provider path
Cohere	Recognized from Haystack component identity/metadata; generic wrapper	Structural Haystack adapter tests only; no standalone tutorial	Bundled standard rates for selected Command models	Experimental, not live-validated
OpenRouter	Native OpenAI-compatible proxy or LiteLLM callback	Unit tests plus direct response/stream normalization	Provider-reported charged cost is authoritative; upstream cost retained separately	Beta native enrichment
Hugging Face inference	smolagents OpenInference, LiteLLM, or generic OTLP	Official instrumentor conformance and normalization tests	Provider money when exposed; custom endpoints otherwise remain unmeasured	Beta framework path
Any OTel GenAI provider	Standard OTLP/HTTP attributes	Generic protocol and normalization tests	Provider-reported money or custom catalog	Protocol support, not a native adapter

“Recognized by Haystack” means the integration can associate provider/model/usage metadata with the actual native generator span. It does not mean Witdem bundles that provider's client library or credentials.

Provider aliases are normalized before analytics. For example, azure.openai groups with azure_openai, aws.bedrock groups with amazon_bedrock, google.vertex groups with google, and mistralai groups with mistral. The originally observed value is retained as provenance when it differs from the canonical name.

Evidence required for a measured model call¶

For provider/model comparisons, emit or expose:

gen_ai.provider.name
gen_ai.request.model or gen_ai.response.model
gen_ai.usage.input_tokens
gen_ai.usage.output_tokens

Provider-reported cost can be supplied as gen_ai.cost.usd or through a native/generic SDK operation. Without reported money, the model must match src/witdem/pricing/catalog.yaml. Override that catalog with WITDEM_PRICING_FILE when your deployment uses other models or negotiated rates. See Pricing catalog for dimensional rates, custom meters, and automated refresh.

Generic provider integration¶

Use this for an existing provider function whose result exposes conventional metadata:

from witdem_sdk.integrations.generic import instrument

call = instrument(
    call_provider,
    operation_name="provider.generate",
    provider="provider-name",
    model="model-name",
)
response = call(prompt)

If its response fields differ, supply observe_result=:

call = instrument(
    call_provider,
    operation_name="provider.generate",
    provider="provider-name",
    model="model-name",
    observe_result=lambda response: {
        "response_model": response.model_id,
        "input_tokens": response.usage.input,
        "output_tokens": response.usage.output,
        "cost_usd": response.billed_usd,
        "cost_source": "provider_reported",
    },
)

This maps fields; it does not fabricate calls, token counts, or prices.