tool_choice is not supported by the Google Realtime API.

[vicky-sai]

pipecat version

0.0.104

Python version

3.12.12

Operating System

mac OS Apple M4 Pro

Issue description

1. set_tool_choice() exists on LLMContext (line 351) and stores the value correctly
2. OpenAI adapter reads it → context.tool_choice at line 67 → passes to OpenAI API → works
3. Gemini adapter ignores it → GeminiLLMInvocationParams has no tool_choice field, get_llm_invocation_params() never reads context.tool_choice
4. Gemini Live _connect() builds LiveConnectConfig from the adapter output → only gets system_instruction and tools → no tool_choice propagated

The set_tool_choice setter method IS real and functional - but the Gemini adapter simply doesn't extract it from context. It's a gap in pipecat's Gemini integration. The OpenAI path uses it; the Gemini path doesn't. Your first_user_turn_done handler-level gate is still the correct workaround for Gemini Live.

Reproduction steps

Just use

Expected behavior

Used 'tool_choice in LLMContext for Gemini Native live audio model

Actual behavior

tool_choice not settable

Logs

[vicky-sai]

class GeminiLLMInvocationParams(TypedDict):
"""Context-based parameters for invoking Gemini LLM."""

system_instruction: Optional[str]
messages: List[Content]
tools: List[Any] | NotGiven

And its get_llm_invocation_params() returns:
return {
"system_instruction": ...,
"messages": ...,
"tools": ...,
# ← NO tool_choice
}

[omChauhanDev]

hey @vicky-sai, thanks for the detailed breakdown, two things going on here:

non-live gemini (GoogleLLMService / GoogleVertexLLMService): real gap, fixed - the adapter now converts context.tool_choice to gemini's ToolConfig/FunctionCallingConfig.

gemini live: google's live API (LiveConnectConfig) straight up doesn't have a tool_config field, the SDK rejects it.
this is an upstream limitation, tracked here: googleapis/python-genai#468. once google adds it, we can wire it through.

workaround (non-live, before the fix lands):

llm = GoogleLLMService(
    api_key="...",
    tool_config={"function_calling_config": {"mode": "ANY"}},
)