ADK LlmAgent Uses gemini-1.5-flash Despite Explicit gemini-2.5-pro Configuration

[BsnizND]

Title: ADK LlmAgent Uses gemini-1.5-flash Despite Explicit gemini-2.5-pro Configuration

Description:

I am encountering an issue where the ADK framework appears to be overriding the explicitly configured Gemini model ID for an LlmAgent. I am trying to use gemini-2.5-pro-exp-03-25 (or gemini-2.5-pro-preview-03-25) for an audio transcription and diarization task, but logs consistently show that gemini-1.5-flash is being used instead. This prevents the agent from performing tasks like speaker diarization effectively, which require the capabilities of the Pro model.

Configuration:

1. Agent Initialization (src/transcription_agent/agent.py):
   The agent is configured using google.adk.models.google_llm.Gemini with the desired Pro model ID:

   from google.adk.agents import LlmAgent
   from google.adk.models.google_llm import Gemini
   from .tools import transcribe_diarize_audio # Custom tool
   
   # Use the Pro model
   AGENT_MODEL_ID = "gemini-2.5-pro-exp-03-25" # Or gemini-2.5-pro-preview-03-25
   
   # ... GCP Project/Location setup ...
   
   def create_transcription_agent() -> LlmAgent:
       try:
           gemini_model = Gemini(
               model_id=AGENT_MODEL_ID,
               project=GCP_PROJECT_ID,
               location=GCP_LOCATION
           )
           agent = LlmAgent(
               name="TranscriptionAgent",
               # ... description, instruction ...
               model=gemini_model,
               tools=[transcribe_diarize_audio]
           )
           logger.info(f"Transcription Agent created successfully using model: {AGENT_MODEL_ID}")
           return agent
       except Exception as e:
           logger.exception(f"Failed to create Transcription Agent: {e}")
           raise
   agent = create_transcription_agent()

   The agent creation log confirms this: INFO - agent.py:68 - Transcription Agent created successfully using model: gemini-2.5-pro-exp-03-25

2. Tool Implementation (src/transcription_agent/tools.py):
   The custom tool (transcribe_diarize_audio) also internally initializes a client using vertexai.generative_models.GenerativeModel targeting the Pro model:

   import vertexai
   from vertexai.generative_models import GenerativeModel, Part
   
   # ... inside the tool function ...
   try:
       # ... client setup ...
       model_id = "gemini-2.5-pro-exp-03-25" # Explicitly set within the tool
       model = GenerativeModel(model_name=model_id)
       # ... call model.generate_content ...
   except Exception as e:
       # ... error handling ...

Expected Behavior:

All API calls made by the ADK framework in the context of this agent (including internal reasoning steps, function calling decisions, and the execution of the tool's own API call) should use the configured gemini-2.5-pro-exp-03-25 model.

Actual Behavior:

Logs consistently show that gemini-1.5-flash is being used:

1. Agent's Internal Call: Immediately after agent creation, the framework logs show a request being prepared for Flash:
   INFO - google_llm.py:83 - Sending out request, model: gemini-1.5-flash, backend: vertex, stream: False
2. Tool's API Call (when invoked via Agent): When the agent invokes the tool, the underlying Vertex AI client logs also show the generateContent call targeting Flash:
   INFO - _client.py:1740 - HTTP Request: POST .../models/gemini-1.5-flash:generateContent "HTTP/1.1 200 OK"

Impact:

Using gemini-1.5-flash leads to failed or incomplete speaker diarization for the audio processing task, defeating the purpose of configuring the agent to use the more capable Pro model.

Steps to Reproduce (Conceptual):

1. Create an LlmAgent and explicitly pass a Gemini model instance configured with model_id="gemini-2.5-pro-exp-03-25" (or another Pro model).
2. Register a custom tool with the agent. This tool should internally make its own call to GenerateContent using the same (or another) Pro model via the standard Vertex AI SDK (vertexai.generative_models.GenerativeModel).
3. Invoke the agent with a task that requires the custom tool to be called (e.g., processing an audio file).
4. Observe the logs from google_llm.py and google.cloud.aiplatform_v1.services.prediction_service.client (or similar) to see which model is actually being targeted by the API requests.

Environment:

• ADK Version: [Please fill in your ADK version if known]
• Python Version: 3.9
• OS: macOS

Question:

Is this behavior expected? Is there a known issue or a specific way the ADK framework overrides model configurations provided via google.adk.models.google_llm.Gemini? How can we ensure that the explicitly configured Pro model is consistently used for all agent-related API calls, including the tool's execution when invoked by the agent?

Thank you for any insights or guidance.

[boyangsvl]

Closing this discussion as it's being discussed at #279