Support tool-calling for LLMs with coding skills

[nileshtrivedi]

Even if the LLM (for eg: Gemma3 hosted by Gemini) does not support standard tool-calling spec, it can be accomplished by sending the function signatures in the system prompt and post-processing the LLM response. LiteLLM does this with a flag (add_function_to_prompt) and has documented it here.

This would be great to have in ADK as it will become possible to use models like Gemma3-27b for Agentic tasks.

[selcukgun]

Thank you for sharing this!

kwargs for our LiteLlm constructor are passed as-is to litellm generate method:

adk-python/src/google/adk/models/lite_llm.py

Line 630 in 0c40542

def __init__(self, model: str, **kwargs):

The following should work without any change on the ADK side.

litellm.add_function_to_prompt = True 
... 
root_agent = Agent(
    model=LiteLlm(model='YOUR_MODEL', functions=functions),
...

please give it a try.

[nileshtrivedi]

It's currently broken in LiteLLM which is why I submitted this feature request for ADK to support this directly for Gemma etc without LiteLLM.

[genquan9]

After some experiments, it seems that the support of add_function_to_prompt not working well yet.

btw, there is an example to preprocess prompt and post process function calls in https://github.com/GoogleCloudPlatform/vertex-ai-samples/blob/main/notebooks/community/model_garden/model_garden_integration_with_agent.ipynb.