Using Open Deep Research example with Groq deepseek-r1-distill-llama-70b

[tulas75]

Anyone tried to use the open deep research example with this Groq resoaning model?
It seems to work for reasoning but it fails calling function (websearch).
From the groq website, the model should support fuction calling.

best,
.t

[medecau]

found this discussion while looking for comments on using reasoning models with smolagents

i've been trying to use o3-mini as manager for a 4o-mini based agent with access to some api endpoints – using the managed_agents argument. what i have observed is that the manager model will rarely call into the expert for information and will rather come up with an answer on its own where information from the expert would be preferred.

i think there might be a problem with the default prompt and the react method being used on reasoning models.

[tulas75]

Hi @medecau

Thank you for answering.
Analyzing Groq's logs, I don't think this is the case. The reasoning model tries to call functions, but something wrong happens.
I saw something similar using microsoft autogen calling tools in groq.

This is the error code 400:

{
"request_id": "req_01jn8fq8e9fsf9nbvc3cxc65qb",
"created_at": "2025-03-01T09:13:43.113Z",
"error": {
"message": "Failed to call a function. Please adjust your prompt. See 'failed_generation' for more details.",
"type": "invalid_request_error",
"code": "tool_use_failed"
}
}

[medecau]

okay, if you have the same issue across two different frameworks than it could be that their api changed and these frameworks are not yet up-to-date. you might get away with simply refreshing your dependencies, that would look like uv sync -U or similar.