GoogleLLMService: apply model-aware default thinking config for Gemini 3+

[yydrift-code]

pipecat version

0.0.104

Python version

3.12

Operating System

macOS

Issue description

GoogleLLMService currently applies a default thinking override only for gemini-2.5-flash* via _maybe_unset_thinking_budget() in pipecat/services/google/llm.py.

That works for Gemini 2.5 Flash, but Gemini 3+ uses a different thinking surface (thinking_level, not thinking_budget). If a caller uses GoogleLLMService or GoogleVertexLLMService with a Gemini 3 Flash model and does not pass an explicit params.thinking, Pipecat leaves thinking_config unset and the model falls back to provider defaults.

For latency-sensitive voice use cases this is surprising because:

• gemini-2.5-flash* gets an automatic low-latency default (thinking_budget=0)
• gemini-3* gets no automatic default at all, even though the model family requires a different control surface

This is especially easy to hit now that newer Gemini 3 / 3.1 Flash models are becoming default choices.

Related Google docs:

• https://ai.google.dev/gemini-api/docs/thinking

Reproduction steps

1. Initialize GoogleLLMService or GoogleVertexLLMService with a Gemini 3 Flash model, for example gemini-3.1-flash-lite-preview.
2. Do not pass params.thinking.
3. Inspect the generation params or run a normal call.
4. Observe that Pipecat does not populate thinking_config at all for Gemini 3 models.

Minimal example:

from pipecat.services.google.llm_vertex import GoogleVertexLLMService

llm = GoogleVertexLLMService(
    credentials_path="...",
    project_id="...",
    location="europe-west2",
    model="gemini-3.1-flash-lite-preview",
    system_instruction="You are a voice assistant.",
)

params = llm._build_generation_params(system_instruction="x", tools=[], tool_config=None)
llm._maybe_unset_thinking_budget(params)
print(params)
# No thinking_config is added for Gemini 3 models.

Expected behavior

If the caller did not specify params.thinking, Pipecat should apply model-aware defaults:

• gemini-2.5-flash* -> thinking_budget=0, include_thoughts=False
• gemini-3-flash* / gemini-3.1-flash* -> thinking_level="minimal", include_thoughts=False
• do not override explicit user-provided thinking config
• probably leave Pro-family models untouched unless there is a clear safe default

Actual behavior

Only the Gemini 2.5 Flash family gets a default low-thinking config. Gemini 3+ models get provider-default thinking behavior unless every application remembers to configure it explicitly.

Suggested fix

Replace _maybe_unset_thinking_budget() with a model-aware default-thinking helper, for example:

• detect the model family
• apply the appropriate control surface for that family
• skip entirely when thinking_config is already present

This would make latency behavior much more predictable across Gemini model upgrades.