Feedback on Tool Usage Issue with Ollama Provider

[Shedulab]

Hi,

I encountered an issue where automatic tool usage (tools= together with max_turns=) didn’t work for locally downloaded OSS models when using the Ollama Provider. I had GPT-5 fix the code, and it somehow works now — though I haven’t had the time to inspect exactly how GPT-5 fixed it. At least it works. 🙂

I’m sharing the code below for reference, in case it might be useful for an official fix.
ollama_provider.py

By the way, thank you very much for the incredible work you’ve done creating this amazing library.

Best regards,
Huseyin

[tushardhole]

@Shedulab

I debugged this to understood why your script works and whats issue in, existing OllamaProvider.

in OllamaProvider,

        normalized_response.choices[0].message.content = response_data["message"][
            "content"
        ]

and then inside message we try to find tool_calls

            tool_calls = (
                getattr(response.choices[0].message, "tool_calls", None)
                if hasattr(response, "choices")
                else None
            )

But actual response from ollama is,

{
  "model": "qwen3:4b",
  "created_at": "2025-12-14T22:21:29.23928Z",
  "message": {
    "role": "assistant",
    "content": "",
    "thinking": "Okay, let's see. The user is asking if it will rain in San Francisco at 2pm today so they can plan an outdoor picnic. I need to check the weather using the provided function.\n\nFirst, the function name is will_it_rain. The required parameters are location and time_of_day. The user mentioned San Francisco as the location and 2pm as the time. Wait, the time_of_day parameter is a string. So I should format the time as \"2pm\" or maybe \"14:00\"? But the user said 2pm, so probably \"2pm\" is acceptable here.\n\nI need to make sure the parameters are correctly named. The function requires location (string) and time_of_day (string). So I'll set location to \"San Francisco\" and time_of: \"2pm\".\n\nWait, the user says \"at 2pm\", so the time_of_day should be \"2pm\". Let me check if the function expects a specific format, but since the user provided \"2pm\", I'll use that. Maybe the function handles it as is.\n\nSo the tool call should be will_it_rain with location \"San Francisco\" and time_of_day \"2pm\".\n\nI need to output that in the tool_call XML tags as a JSON object.\n\nLet me structure the JSON: {\"name\": \"will_it_rain\", \"arguments\": {\"location\": \"San Francisco\", \"time_of_day\": \"2pm\"}}\n\nYes, that should do it.\n",
    "tool_calls": [
      {
        "id": "call_bdpxbxqr",
        "function": {
          "index": 0,
          "name": "will_it_rain",
          "arguments": {
            "location": "San Francisco",
            "time_of_day": "2pm"
          }
        }
      }
    ]
  },
  "done": true,
  "done_reason": "stop",
  "total_duration": 18166467750,
  "load_duration": 161415541,
  "prompt_eval_count": 180,
  "prompt_eval_duration": 97708042,
  "eval_count": 341,
  "eval_duration": 17515174174
}

Notice content is blank/empty and tool_calls is sibling to content (not child).

The script you fixed, explicitly finds that tool_calls and sets it

raw_tool_calls = msg.get("tool_calls") or []

[themerius]

@Shedulab Your fix works! Did you open a pull request so that it can be merged into main?

[Shedulab]

Hi @themerius, thank you!
For now, I didn’t open a pull request, since @tushardhole put significant effort into debugging both my fix and the original problematic Ollama provider to pinpoint where it actually breaks.

His suggestion was to migrate the Ollama provider to an OpenAI-compatible API, as newer Ollama versions support OpenAI API standards. For now, I’ll stick with that approach as well.