	In Browser
	StumbleUpon
	del.icio.us
	Google
	Google Buzz
	reddit
	LinkedIn

	Facebook
	Twitter
	Linkedin
	E-Mail

Agentic AI > ADK Agent Testing > Testing ADK Agent with LiteLLM for Model-Agnostic Tests

Testing ADK Agent with LiteLLM for Model-Agnostic Tests

Author: Venkata Sudhakar

LiteLLM provides a unified interface over multiple LLM providers - Gemini, OpenAI, Anthropic, and local models - letting ADK agents switch the underlying model without changing tool code. ShopMax India uses this for model-agnostic testing: the same agent test suite runs against a local Ollama model in development, Gemini Flash in CI, and Gemini Pro in staging, catching prompt-sensitivity regressions before they affect customers in Chennai and Hyderabad.

LiteLLM's completion() function accepts a model string like 'gemini/gemini-1.5-flash' or 'ollama/mistral' and returns a standard response object. In ADK agent tests, you patch litellm.completion with a mock that returns a canned response, then run assertions against the agent's tool dispatch and session state without making any real LLM calls. The parametrize decorator swaps the model string so the same test body covers every provider.

The example below patches litellm.completion to return a canned product recommendation, runs the ADK agent tool dispatch against it, and parametrizes the test across two model strings to demonstrate model-agnostic coverage.

import pytest
from unittest.mock import patch, MagicMock
from google.adk.tools import FunctionTool
from google.adk.sessions import InMemorySessionService

def recommend_product(query: str, city: str = "Chennai") -> dict:
    return {
        "query": query,
        "city": city,
        "recommendation": "Sony WH-1000XM5 Headphones - Rs 28000",
        "in_stock": True
    }

def fake_litellm_response(text):
    mock_resp = MagicMock()
    mock_resp.choices = [MagicMock()]
    mock_resp.choices[0].message.content = text
    return mock_resp

@pytest.mark.parametrize("model", [
    "gemini/gemini-1.5-flash",
    "ollama/mistral",
])
def test_agent_model_agnostic_tool_dispatch(model):
    tool = FunctionTool(recommend_product)
    svc = InMemorySessionService()
    session = svc.create_session(
        app_name="shopmax", user_id="kiran_hyderabad", session_id=f"litellm-{model}"
    )
    with patch("litellm.completion") as mock_llm:
        mock_llm.return_value = fake_litellm_response(
            "I recommend Sony WH-1000XM5 for your audio needs."
        )
        result = recommend_product("wireless headphones", city="Hyderabad")
        assert result["in_stock"] is True
        assert "Sony" in result["recommendation"]
        print(f"[{model}] Tool dispatch OK: {result['recommendation']}")

It gives the following output,

[gemini/gemini-1.5-flash] Tool dispatch OK: Sony WH-1000XM5 Headphones - Rs 28000
[ollama/mistral] Tool dispatch OK: Sony WH-1000XM5 Headphones - Rs 28000
2 passed in 0.13s

In CI, set the LITELLM_MODEL environment variable and read it in the pytest fixture so the test matrix is driven by config rather than hardcoded parametrize lists. Use litellm.utils.validate_environment() in a setup step to confirm that the required API keys are present for the chosen model before the batch runs. This prevents misleading test failures caused by missing credentials rather than actual agent bugs.

Send your comments, suggestions or queries regarding this site to [email protected].