Known Provider Limitations

[rwjdk]

This issue tracks the known provider limitations that are not bugs in AgentFrameworkToolkit itself, but underlying implementation issues from providers/nuget packages:

Anthropic

• Do not have Embedding Models, so do not have an EmbeddingFactory

Mistral

• The latest mistral-small-2603, mistral-medium-2603, and mistral-large-2603 are incapable of producing correct Structured Output JSON... The -2506 models work.

OpenRouter

• Combining OpenAI Responses API and Tool Calling results in a 400 Bad Request in the tool call part

XAI

• Combining OpenAI Responses API and Tool Calling results in a 400 Bad Request in the tool call part

Cohere

• Despite being OpenAI-based, the ResponsesAPI is not supported

Cerebras

• Despite being OpenAI-based, the ResponsesAPI is not supported

Microsoft Foundry

• The system can only do Chat (ChatClient and Responses API). The rest of OpenAI features, like embeddings, do not work (use AzureOpenAI instead)

---

I disagree that some of the above are not provider/nuget limitations but an issue with AgentFrameworkToolkit!

If so, please create a new issue with raw nuget usage of provider-nuget/Microsoft Agent Framework that shows that the scenario is supported

[arturcic]

@rwjdk can you update the info for Google?

[rwjdk]

@arturcic Done (thank you for the reminder)

[rwjdk]

Anthropic - Structured Output will be supported in the next release (post rc3)

[rwjdk]

Updated with new Mistral limitations (Their newest models (...-2603) are messed up and can't generate correct structured output anymore 🤦)

[rwjdk]

Removed Google: Do not support Raw HTML Call Inspection as Google 1.6.2 now supports it, so will be part of next release

[rwjdk]

Removed GitHub limitations as the Service is being retired and has been removed from AgentFrameworkToolkit

[rwjdk]

Added MicrosoftFoundry limitations