GoogleLLMService throws 400 error when using gemini 3 flash and executing multiple tool calls in one LLM turn

[markbackman]

pipecat version

0.0.98

Python version

3.12.11

Operating System

macOS 15

Issue description

When the LLM executes multiple tool calls in a single turn, the following error is thrown:

GoogleLLMService#0 exception (~/src/pipecat/.venv/lib/python3.12/site-packages/google/genai/errors.py:210): Unknown error occurred: 400 Bad Request. {'message': '{\n  "error": {\n    "code": 400,\n    "message": "Function call is missing a thought_signature in functionCall parts. This is required for tools to work correctly, and missing thought_signature may lead to degraded model performance. Additional data, function call `default_api:get_restaurant_recommendation` , position 7. Please refer to https://ai.google.dev/gemini-api/docs/thought-signatures for more details.",\n    "status": "INVALID_ARGUMENT"\n  }\n}\n', 'status': 'Bad Request'}

Reproduction steps

1. Run https://github.com/pipecat-ai/pipecat/blob/main/examples/foundational/14e-function-calling-google.py
2. Say, "Hi. I'm in Atlanta. Can you give me the weather and a restaurant recommendation?"
3. Bot says "Let me check on that", then thee 400 error.

Expected behavior

Bot functions are called.

Actual behavior

400 error

Logs

[omChauhanDev]

i did some digging into this & found :

Issue: Gemini 3 is way stricter than 2.5 about thought_signature validation. Specifically, when it returns parallel function calls in one response, it only puts the signature on the first function call part. But when we reconstruct the context for the next LLM turn, our aggregator was splitting those parallel calls into separate messages. Gemini then sees them as separate "steps" and fails validation because each step's first function call needs a signature.

Breaking in our example bcz, when Gemini returns Content(parts=[Part(function_call=get_weather, sig='...'), Part(function_call=get_restaurant)]), we were creating two separate messages. The second message (get_restaurant) ends up as a new "step" without a signature, which triggers that 400 INVALID_ARGUMENT.

Fix: We need to keep parallel function calls together in a single message, matching how Gemini returns them. This also aligns with the standard format expected by all LLM adapters (OpenAI, Anthropic, Bedrock).

I've verified this approach locally & this made our foundational example working as expected. Since @kompfner is already working on this, I'm happy to either share the logic or open a quick PR, just let me know what works best!

[markbackman]

@omChauhanDev thanks for looking into this 🙇

@kompfner hasn't started work on this. I've flagged it for him after he comes back from vacation. So, if you'd like to submit a PR, we'd be happy to have it.