mirror of
https://github.com/encounter/adk-python.git
synced 2026-07-09 18:19:28 -07:00
feat: Support both output_schema and tools at the same time in LlmAgent
1. Allow developers to specify output schema and tools together. 2. If both are specified, do the following: 2.1 Do not set output schema on the model config 2.2 Add a special tool called set_model_response(result) 2.3 `result` has the same schema as the requested output_schema 2.4 Instruct the model to use set_model_response() to output its final result, rather than output text directly. 2.5 When the set_model_response() is called, ADK will extract its content and put it in a text part, so the client would treat it as the model response. PiperOrigin-RevId: 792686011
This commit is contained in:
Xiang (Sean) Zhou
committed by
Copybara-Service
Copybara-Service
parent
b4ce3b12d1
commit
af635674b5
@@ -0,0 +1,145 @@
|
||||
# Copyright 2025 Google LLC
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
"""Tests for basic LLM request processor."""
|
||||
|
||||
from google.adk.agents.invocation_context import InvocationContext
|
||||
from google.adk.agents.llm_agent import LlmAgent
|
||||
from google.adk.agents.run_config import RunConfig
|
||||
from google.adk.flows.llm_flows.basic import _BasicLlmRequestProcessor
|
||||
from google.adk.models.llm_request import LlmRequest
|
||||
from google.adk.sessions.in_memory_session_service import InMemorySessionService
|
||||
from google.adk.tools.function_tool import FunctionTool
|
||||
from pydantic import BaseModel
|
||||
from pydantic import Field
|
||||
import pytest
|
||||
|
||||
|
||||
class OutputSchema(BaseModel):
|
||||
"""Test schema for output."""
|
||||
|
||||
name: str = Field(description='A name')
|
||||
value: int = Field(description='A value')
|
||||
|
||||
|
||||
def dummy_tool(query: str) -> str:
|
||||
"""A dummy tool for testing."""
|
||||
return f'Result: {query}'
|
||||
|
||||
|
||||
async def _create_invocation_context(agent: LlmAgent) -> InvocationContext:
|
||||
"""Helper to create InvocationContext for testing."""
|
||||
session_service = InMemorySessionService()
|
||||
session = await session_service.create_session(
|
||||
app_name='test_app', user_id='test_user'
|
||||
)
|
||||
return InvocationContext(
|
||||
invocation_id='test-id',
|
||||
agent=agent,
|
||||
session=session,
|
||||
session_service=session_service,
|
||||
run_config=RunConfig(),
|
||||
)
|
||||
|
||||
|
||||
class TestBasicLlmRequestProcessor:
|
||||
"""Test class for _BasicLlmRequestProcessor."""
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_sets_output_schema_when_no_tools(self):
|
||||
"""Test that processor sets output_schema when agent has no tools."""
|
||||
agent = LlmAgent(
|
||||
name='test_agent',
|
||||
model='gemini-1.5-flash',
|
||||
output_schema=OutputSchema,
|
||||
tools=[], # No tools
|
||||
)
|
||||
|
||||
invocation_context = await _create_invocation_context(agent)
|
||||
llm_request = LlmRequest()
|
||||
processor = _BasicLlmRequestProcessor()
|
||||
|
||||
# Process the request
|
||||
events = []
|
||||
async for event in processor.run_async(invocation_context, llm_request):
|
||||
events.append(event)
|
||||
|
||||
# Should have set response_schema since agent has no tools
|
||||
assert llm_request.config.response_schema == OutputSchema
|
||||
assert llm_request.config.response_mime_type == 'application/json'
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_skips_output_schema_when_tools_present(self):
|
||||
"""Test that processor skips output_schema when agent has tools."""
|
||||
agent = LlmAgent(
|
||||
name='test_agent',
|
||||
model='gemini-1.5-flash',
|
||||
output_schema=OutputSchema,
|
||||
tools=[FunctionTool(func=dummy_tool)], # Has tools
|
||||
)
|
||||
|
||||
invocation_context = await _create_invocation_context(agent)
|
||||
llm_request = LlmRequest()
|
||||
processor = _BasicLlmRequestProcessor()
|
||||
|
||||
# Process the request
|
||||
events = []
|
||||
async for event in processor.run_async(invocation_context, llm_request):
|
||||
events.append(event)
|
||||
|
||||
# Should NOT have set response_schema since agent has tools
|
||||
assert llm_request.config.response_schema is None
|
||||
assert llm_request.config.response_mime_type != 'application/json'
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_no_output_schema_no_tools(self):
|
||||
"""Test that processor works normally when agent has no output_schema or tools."""
|
||||
agent = LlmAgent(
|
||||
name='test_agent',
|
||||
model='gemini-1.5-flash',
|
||||
# No output_schema, no tools
|
||||
)
|
||||
|
||||
invocation_context = await _create_invocation_context(agent)
|
||||
llm_request = LlmRequest()
|
||||
processor = _BasicLlmRequestProcessor()
|
||||
|
||||
# Process the request
|
||||
events = []
|
||||
async for event in processor.run_async(invocation_context, llm_request):
|
||||
events.append(event)
|
||||
|
||||
# Should not have set anything
|
||||
assert llm_request.config.response_schema is None
|
||||
assert llm_request.config.response_mime_type != 'application/json'
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_sets_model_name(self):
|
||||
"""Test that processor sets the model name correctly."""
|
||||
agent = LlmAgent(
|
||||
name='test_agent',
|
||||
model='gemini-1.5-flash',
|
||||
)
|
||||
|
||||
invocation_context = await _create_invocation_context(agent)
|
||||
llm_request = LlmRequest()
|
||||
processor = _BasicLlmRequestProcessor()
|
||||
|
||||
# Process the request
|
||||
events = []
|
||||
async for event in processor.run_async(invocation_context, llm_request):
|
||||
events.append(event)
|
||||
|
||||
# Should have set the model name
|
||||
assert llm_request.model == 'gemini-1.5-flash'
|
||||
@@ -0,0 +1,409 @@
|
||||
# Copyright 2025 Google LLC
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
"""Tests for output schema processor functionality."""
|
||||
|
||||
import json
|
||||
|
||||
from google.adk.agents.invocation_context import InvocationContext
|
||||
from google.adk.agents.llm_agent import LlmAgent
|
||||
from google.adk.agents.run_config import RunConfig
|
||||
from google.adk.flows.llm_flows.single_flow import SingleFlow
|
||||
from google.adk.models.llm_request import LlmRequest
|
||||
from google.adk.models.llm_response import LlmResponse
|
||||
from google.adk.sessions.in_memory_session_service import InMemorySessionService
|
||||
from google.adk.tools.function_tool import FunctionTool
|
||||
from pydantic import BaseModel
|
||||
from pydantic import Field
|
||||
import pytest
|
||||
|
||||
|
||||
class PersonSchema(BaseModel):
|
||||
"""Test schema for structured output."""
|
||||
|
||||
name: str = Field(description="A person's name")
|
||||
age: int = Field(description="A person's age")
|
||||
city: str = Field(description='The city they live in')
|
||||
|
||||
|
||||
def dummy_tool(query: str) -> str:
|
||||
"""A dummy tool for testing."""
|
||||
return f'Searched for: {query}'
|
||||
|
||||
|
||||
async def _create_invocation_context(agent: LlmAgent) -> InvocationContext:
|
||||
"""Helper to create InvocationContext for testing."""
|
||||
session_service = InMemorySessionService()
|
||||
session = await session_service.create_session(
|
||||
app_name='test_app', user_id='test_user'
|
||||
)
|
||||
return InvocationContext(
|
||||
invocation_id='test-id',
|
||||
agent=agent,
|
||||
session=session,
|
||||
session_service=session_service,
|
||||
run_config=RunConfig(),
|
||||
)
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_output_schema_with_tools_validation_removed():
|
||||
"""Test that LlmAgent now allows output_schema with tools."""
|
||||
# This should not raise an error anymore
|
||||
agent = LlmAgent(
|
||||
name='test_agent',
|
||||
model='gemini-1.5-flash',
|
||||
output_schema=PersonSchema,
|
||||
tools=[FunctionTool(func=dummy_tool)],
|
||||
)
|
||||
|
||||
assert agent.output_schema == PersonSchema
|
||||
assert len(agent.tools) == 1
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_basic_processor_skips_output_schema_with_tools():
|
||||
"""Test that basic processor doesn't set output_schema when tools are present."""
|
||||
from google.adk.flows.llm_flows.basic import _BasicLlmRequestProcessor
|
||||
|
||||
agent = LlmAgent(
|
||||
name='test_agent',
|
||||
model='gemini-1.5-flash',
|
||||
output_schema=PersonSchema,
|
||||
tools=[FunctionTool(func=dummy_tool)],
|
||||
)
|
||||
|
||||
invocation_context = await _create_invocation_context(agent)
|
||||
|
||||
llm_request = LlmRequest()
|
||||
processor = _BasicLlmRequestProcessor()
|
||||
|
||||
# Process the request
|
||||
events = []
|
||||
async for event in processor.run_async(invocation_context, llm_request):
|
||||
events.append(event)
|
||||
|
||||
# Should not have set response_schema since agent has tools
|
||||
assert llm_request.config.response_schema is None
|
||||
assert llm_request.config.response_mime_type != 'application/json'
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_basic_processor_sets_output_schema_without_tools():
|
||||
"""Test that basic processor still sets output_schema when no tools are present."""
|
||||
from google.adk.flows.llm_flows.basic import _BasicLlmRequestProcessor
|
||||
|
||||
agent = LlmAgent(
|
||||
name='test_agent',
|
||||
model='gemini-1.5-flash',
|
||||
output_schema=PersonSchema,
|
||||
tools=[], # No tools
|
||||
)
|
||||
|
||||
invocation_context = await _create_invocation_context(agent)
|
||||
|
||||
llm_request = LlmRequest()
|
||||
processor = _BasicLlmRequestProcessor()
|
||||
|
||||
# Process the request
|
||||
events = []
|
||||
async for event in processor.run_async(invocation_context, llm_request):
|
||||
events.append(event)
|
||||
|
||||
# Should have set response_schema since agent has no tools
|
||||
assert llm_request.config.response_schema == PersonSchema
|
||||
assert llm_request.config.response_mime_type == 'application/json'
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_output_schema_request_processor():
|
||||
"""Test that output schema processor adds set_model_response tool."""
|
||||
from google.adk.flows.llm_flows._output_schema_processor import _OutputSchemaRequestProcessor
|
||||
|
||||
agent = LlmAgent(
|
||||
name='test_agent',
|
||||
model='gemini-1.5-flash',
|
||||
output_schema=PersonSchema,
|
||||
tools=[FunctionTool(func=dummy_tool)],
|
||||
)
|
||||
|
||||
invocation_context = await _create_invocation_context(agent)
|
||||
|
||||
llm_request = LlmRequest()
|
||||
processor = _OutputSchemaRequestProcessor()
|
||||
|
||||
# Process the request
|
||||
events = []
|
||||
async for event in processor.run_async(invocation_context, llm_request):
|
||||
events.append(event)
|
||||
|
||||
# Should have added set_model_response tool
|
||||
assert 'set_model_response' in llm_request.tools_dict
|
||||
|
||||
# Should have added instruction about using set_model_response
|
||||
assert 'set_model_response' in llm_request.config.system_instruction
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_set_model_response_tool():
|
||||
"""Test the set_model_response tool functionality."""
|
||||
from google.adk.tools.set_model_response_tool import MODEL_JSON_RESPONSE_KEY
|
||||
from google.adk.tools.set_model_response_tool import SetModelResponseTool
|
||||
from google.adk.tools.tool_context import ToolContext
|
||||
|
||||
tool = SetModelResponseTool(PersonSchema)
|
||||
|
||||
agent = LlmAgent(name='test_agent', model='gemini-1.5-flash')
|
||||
invocation_context = await _create_invocation_context(agent)
|
||||
tool_context = ToolContext(invocation_context)
|
||||
|
||||
# Call the tool with valid data
|
||||
result = await tool.run_async(
|
||||
args={'name': 'John Doe', 'age': 30, 'city': 'New York'},
|
||||
tool_context=tool_context,
|
||||
)
|
||||
|
||||
# Verify the tool now returns dict directly
|
||||
assert result is not None
|
||||
assert result['name'] == 'John Doe'
|
||||
assert result['age'] == 30
|
||||
assert result['city'] == 'New York'
|
||||
|
||||
# Check that the response is no longer stored in session state
|
||||
stored_response = invocation_context.session.state.get(
|
||||
MODEL_JSON_RESPONSE_KEY
|
||||
)
|
||||
assert stored_response is None
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_output_schema_helper_functions():
|
||||
"""Test the helper functions for handling set_model_response."""
|
||||
from google.adk.events.event import Event
|
||||
from google.adk.flows.llm_flows._output_schema_processor import create_final_model_response_event
|
||||
from google.adk.flows.llm_flows._output_schema_processor import get_structured_model_response
|
||||
from google.genai import types
|
||||
|
||||
agent = LlmAgent(
|
||||
name='test_agent',
|
||||
model='gemini-1.5-flash',
|
||||
output_schema=PersonSchema,
|
||||
tools=[FunctionTool(func=dummy_tool)],
|
||||
)
|
||||
|
||||
invocation_context = await _create_invocation_context(agent)
|
||||
|
||||
# Test get_structured_model_response with a function response event
|
||||
test_dict = {'name': 'Jane Smith', 'age': 25, 'city': 'Los Angeles'}
|
||||
test_json = '{"name": "Jane Smith", "age": 25, "city": "Los Angeles"}'
|
||||
|
||||
# Create a function response event with set_model_response
|
||||
function_response_event = Event(
|
||||
author='test_agent',
|
||||
content=types.Content(
|
||||
role='user',
|
||||
parts=[
|
||||
types.Part(
|
||||
function_response=types.FunctionResponse(
|
||||
name='set_model_response', response=test_dict
|
||||
)
|
||||
)
|
||||
],
|
||||
),
|
||||
)
|
||||
|
||||
# Test get_structured_model_response function
|
||||
extracted_json = get_structured_model_response(function_response_event)
|
||||
assert extracted_json == test_json
|
||||
|
||||
# Test create_final_model_response_event function
|
||||
final_event = create_final_model_response_event(invocation_context, test_json)
|
||||
assert final_event.author == 'test_agent'
|
||||
assert final_event.content.role == 'model'
|
||||
assert final_event.content.parts[0].text == test_json
|
||||
|
||||
# Test get_structured_model_response with non-set_model_response function
|
||||
other_function_response_event = Event(
|
||||
author='test_agent',
|
||||
content=types.Content(
|
||||
role='user',
|
||||
parts=[
|
||||
types.Part(
|
||||
function_response=types.FunctionResponse(
|
||||
name='other_tool', response={'result': 'other response'}
|
||||
)
|
||||
)
|
||||
],
|
||||
),
|
||||
)
|
||||
|
||||
extracted_json = get_structured_model_response(other_function_response_event)
|
||||
assert extracted_json is None
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_end_to_end_integration():
|
||||
"""Test the complete output schema with tools integration."""
|
||||
agent = LlmAgent(
|
||||
name='test_agent',
|
||||
model='gemini-1.5-flash',
|
||||
output_schema=PersonSchema,
|
||||
tools=[FunctionTool(func=dummy_tool)],
|
||||
)
|
||||
|
||||
invocation_context = await _create_invocation_context(agent)
|
||||
|
||||
# Create a flow and test the processors
|
||||
flow = SingleFlow()
|
||||
llm_request = LlmRequest()
|
||||
|
||||
# Run all request processors
|
||||
async for event in flow._preprocess_async(invocation_context, llm_request):
|
||||
pass
|
||||
|
||||
# Verify set_model_response tool was added
|
||||
assert 'set_model_response' in llm_request.tools_dict
|
||||
|
||||
# Verify instruction was added
|
||||
assert 'set_model_response' in llm_request.config.system_instruction
|
||||
|
||||
# Verify output_schema was NOT set on the model config
|
||||
assert llm_request.config.response_schema is None
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_flow_yields_both_events_for_set_model_response():
|
||||
"""Test that the flow yields both function response and final model response events."""
|
||||
from google.adk.events.event import Event
|
||||
from google.adk.flows.llm_flows.base_llm_flow import BaseLlmFlow
|
||||
from google.adk.tools.set_model_response_tool import SetModelResponseTool
|
||||
from google.genai import types
|
||||
|
||||
agent = LlmAgent(
|
||||
name='test_agent',
|
||||
model='gemini-1.5-flash',
|
||||
output_schema=PersonSchema,
|
||||
tools=[],
|
||||
)
|
||||
|
||||
invocation_context = await _create_invocation_context(agent)
|
||||
flow = BaseLlmFlow()
|
||||
|
||||
# Create a set_model_response tool and add it to the tools dict
|
||||
set_response_tool = SetModelResponseTool(PersonSchema)
|
||||
llm_request = LlmRequest()
|
||||
llm_request.tools_dict['set_model_response'] = set_response_tool
|
||||
|
||||
# Create a function call event (model calling the function)
|
||||
function_call_event = Event(
|
||||
author='test_agent',
|
||||
content=types.Content(
|
||||
role='model',
|
||||
parts=[
|
||||
types.Part(
|
||||
function_call=types.FunctionCall(
|
||||
name='set_model_response',
|
||||
args={
|
||||
'name': 'Test User',
|
||||
'age': 30,
|
||||
'city': 'Test City',
|
||||
},
|
||||
)
|
||||
)
|
||||
],
|
||||
),
|
||||
)
|
||||
|
||||
# Test the postprocess function handling
|
||||
events = []
|
||||
async for event in flow._postprocess_handle_function_calls_async(
|
||||
invocation_context, function_call_event, llm_request
|
||||
):
|
||||
events.append(event)
|
||||
|
||||
# Should yield exactly 2 events: function response + final model response
|
||||
assert len(events) == 2
|
||||
|
||||
# First event should be the function response
|
||||
first_event = events[0]
|
||||
assert first_event.get_function_responses()[0].name == 'set_model_response'
|
||||
# The response should be the dict returned by the tool
|
||||
assert first_event.get_function_responses()[0].response == {
|
||||
'name': 'Test User',
|
||||
'age': 30,
|
||||
'city': 'Test City',
|
||||
}
|
||||
|
||||
# Second event should be the final model response with JSON
|
||||
second_event = events[1]
|
||||
assert second_event.author == 'test_agent'
|
||||
assert second_event.content.role == 'model'
|
||||
assert (
|
||||
second_event.content.parts[0].text
|
||||
== '{"name": "Test User", "age": 30, "city": "Test City"}'
|
||||
)
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_flow_yields_only_function_response_for_normal_tools():
|
||||
"""Test that the flow yields only function response event for non-set_model_response tools."""
|
||||
from google.adk.events.event import Event
|
||||
from google.adk.flows.llm_flows.base_llm_flow import BaseLlmFlow
|
||||
from google.genai import types
|
||||
|
||||
agent = LlmAgent(
|
||||
name='test_agent',
|
||||
model='gemini-1.5-flash',
|
||||
tools=[FunctionTool(func=dummy_tool)],
|
||||
)
|
||||
|
||||
invocation_context = await _create_invocation_context(agent)
|
||||
flow = BaseLlmFlow()
|
||||
|
||||
# Create a dummy tool and add it to the tools dict
|
||||
dummy_function_tool = FunctionTool(func=dummy_tool)
|
||||
llm_request = LlmRequest()
|
||||
llm_request.tools_dict['dummy_tool'] = dummy_function_tool
|
||||
|
||||
# Create a function call event (model calling the dummy tool)
|
||||
function_call_event = Event(
|
||||
author='test_agent',
|
||||
content=types.Content(
|
||||
role='model',
|
||||
parts=[
|
||||
types.Part(
|
||||
function_call=types.FunctionCall(
|
||||
name='dummy_tool', args={'query': 'test query'}
|
||||
)
|
||||
)
|
||||
],
|
||||
),
|
||||
)
|
||||
|
||||
# Test the postprocess function handling
|
||||
events = []
|
||||
async for event in flow._postprocess_handle_function_calls_async(
|
||||
invocation_context, function_call_event, llm_request
|
||||
):
|
||||
events.append(event)
|
||||
|
||||
# Should yield exactly 1 event: just the function response
|
||||
assert len(events) == 1
|
||||
|
||||
# Should be the function response from dummy_tool
|
||||
first_event = events[0]
|
||||
assert first_event.get_function_responses()[0].name == 'dummy_tool'
|
||||
assert first_event.get_function_responses()[0].response == {
|
||||
'result': 'Searched for: test query'
|
||||
}
|
||||
Reference in New Issue
Block a user