fix: ensure researcher agent uses web search tool instead of generating URLs (#702) (#704)

* fix: ensure researcher agent uses web search tool instead of generating URLs (#702)

- Add enforce_researcher_search configuration option (default: True) to control web search requirement
- Strengthen researcher prompts in both English and Chinese with explicit instructions to use web_search tool
- Implement validate_web_search_usage function to detect if web search tool was used during research
- Add validation logic that warns when researcher doesn't use web search tool
- Enhance logging for web search tools with special markers for easy tracking
- Skip validation during unit tests to avoid test failures
- Update _execute_agent_step to accept config parameter for proper configuration access

This addresses issue #702 where the researcher agent was generating URLs on its own instead of using the web search tool.

* fix: addressed the code review comment

* fix the unit test error and update the code
This commit is contained in:
Willem Jiang
2025-11-24 20:07:28 +08:00
committed by GitHub
parent cc9414f978
commit 478291df07
6 changed files with 84 additions and 11 deletions

View File

@@ -54,6 +54,9 @@ class Configuration:
enforce_web_search: bool = (
False # Enforce at least one web search step in every plan
)
enforce_researcher_search: bool = (
True # Enforce that researcher must use web search tool at least once
)
interrupt_before_tools: list[str] = field(
default_factory=list
) # List of tool names to interrupt before execution

View File

@@ -769,8 +769,51 @@ def research_team_node(state: State):
pass
def validate_web_search_usage(messages: list, agent_name: str = "agent") -> bool:
"""
Validate if the agent has used the web search tool during execution.
Args:
messages: List of messages from the agent execution
agent_name: Name of the agent (for logging purposes)
Returns:
bool: True if web search tool was used, False otherwise
"""
web_search_used = False
for message in messages:
# Check for ToolMessage instances indicating web search was used
if isinstance(message, ToolMessage) and message.name == "web_search":
web_search_used = True
logger.info(f"[VALIDATION] {agent_name} received ToolMessage from web_search tool")
break
# Check for AIMessage content that mentions tool calls
if hasattr(message, 'tool_calls') and message.tool_calls:
for tool_call in message.tool_calls:
if tool_call.get('name') == "web_search":
web_search_used = True
logger.info(f"[VALIDATION] {agent_name} called web_search tool")
break
# break outer loop if web search was used
if web_search_used:
break
# Check for message name attribute
if hasattr(message, 'name') and message.name == "web_search":
web_search_used = True
logger.info(f"[VALIDATION] {agent_name} used web_search tool")
break
if not web_search_used:
logger.warning(f"[VALIDATION] {agent_name} did not use web_search tool")
return web_search_used
async def _execute_agent_step(
state: State, agent, agent_name: str
state: State, agent, agent_name: str, config: RunnableConfig = None
) -> Command[Literal["research_team"]]:
"""Helper function to execute a step using the specified agent."""
logger.debug(f"[_execute_agent_step] Starting execution for agent: {agent_name}")
@@ -918,6 +961,27 @@ async def _execute_agent_step(
logger.debug(f"{agent_name.capitalize()} full response: {response_content}")
# Validate web search usage for researcher agent if enforcement is enabled
web_search_validated = True
should_validate = agent_name == "researcher"
validation_info = ""
if should_validate:
# Check if enforcement is enabled in configuration
configurable = Configuration.from_runnable_config(config) if config else Configuration()
if configurable.enforce_researcher_search:
web_search_validated = validate_web_search_usage(result["messages"], agent_name)
# If web search was not used, add a warning to the response
if not web_search_validated:
logger.warning(f"[VALIDATION] Researcher did not use web_search tool. Adding reminder to response.")
# Add validation information to observations
validation_info = (
"\n\n[WARNING] This research was completed without using the web_search tool. "
"Please verify that the information provided is accurate and up-to-date."
"\n\n[VALIDATION WARNING] Researcher did not use the web_search tool as recommended."
)
# Update the step with the execution result
current_step.execution_res = response_content
logger.info(f"Step '{current_step.title}' execution completed by {agent_name}")
@@ -930,7 +994,7 @@ async def _execute_agent_step(
name=agent_name,
)
],
"observations": observations + [response_content],
"observations": observations + [response_content + validation_info],
**preserve_state_meta_fields(state),
},
goto="research_team",
@@ -1000,7 +1064,7 @@ async def _setup_and_execute_agent_step(
pre_model_hook,
interrupt_before_tools=configurable.interrupt_before_tools,
)
return await _execute_agent_step(state, agent, agent_type)
return await _execute_agent_step(state, agent, agent_type, config)
else:
# Use default tools if no MCP servers are configured
llm_token_limit = get_llm_token_limit_by_type(AGENT_LLM_MAP[agent_type])
@@ -1013,7 +1077,7 @@ async def _setup_and_execute_agent_step(
pre_model_hook,
interrupt_before_tools=configurable.interrupt_before_tools,
)
return await _execute_agent_step(state, agent, agent_type)
return await _execute_agent_step(state, agent, agent_type, config)
async def researcher_node(
@@ -1034,6 +1098,7 @@ async def researcher_node(
logger.info(f"[researcher_node] Researcher tools count: {len(tools)}")
logger.debug(f"[researcher_node] Researcher tools: {[tool.name if hasattr(tool, 'name') else str(tool) for tool in tools]}")
logger.info(f"[researcher_node] enforce_researcher_search is set to: {configurable.enforce_researcher_search}")
return await _setup_and_execute_agent_step(
state,

View File

@@ -37,7 +37,8 @@ You have access to two types of tools:
3. **Plan the Solution**: Determine the best approach to solve the problem using the available tools.
4. **Execute the Solution**:
- Forget your previous knowledge, so you **should leverage the tools** to retrieve the information.
- Use the {% if resources %}**local_search_tool** or{% endif %}**web_search** or other suitable search tool to perform a search with the provided keywords.
- **CRITICAL**: You MUST use the {% if resources %}**local_search_tool** or{% endif %}**web_search** tool to search for information. NEVER generate URLs on your own. All URLs must come from tool results.
- **MANDATORY**: Always perform at least one web search using the **web_search** tool at the beginning of your research. This is not optional.
- When the task includes time range requirements:
- Incorporate appropriate time-based search parameters in your queries (e.g., "after:2020", "before:2023", or specific date ranges)
- Ensure search results respect the specified time constraints.
@@ -71,6 +72,8 @@ You have access to two types of tools:
# Notes
- **CRITICAL**: NEVER generate URLs on your own. All URLs must come from search tool results. This is a mandatory requirement.
- **MANDATORY**: Always start with a web search. Do not rely on your internal knowledge.
- Always verify the relevance and credibility of the information gathered.
- If no URL is provided, focus solely on the search results.
- Never do any math or any file operations.

View File

@@ -37,7 +37,8 @@ CURRENT_TIME: {{ CURRENT_TIME }}
3. **规划解决方案**:确定使用可用工具解决问题的最佳方法。
4. **执行解决方案**
- 忘记你之前的知识,所以你**应该利用工具**来检索信息。
- 使用{% if resources %}**local_search_tool**或{% endif %}**web_search**或其他合适的搜索工具以提供的关键词执行搜索
- **关键要求**:你必须使用{% if resources %}**local_search_tool**或{% endif %}**web_search**工具搜索信息。绝对不能自己生成URL。所有URL必须来自工具结果
- **强制要求**:在研究开始时必须使用**web_search**工具至少执行一次网络搜索。这不是可选项。
- 当任务包括时间范围要求时:
- 在查询中纳入适当的基于时间的搜索参数(如"after:2020"、"before:2023"或特定日期范围)
- 确保搜索结果尊重指定的时间约束。
@@ -66,6 +67,8 @@ CURRENT_TIME: {{ CURRENT_TIME }}
# 注意
- **关键要求**绝对不能自己生成URL。所有URL必须来自搜索工具结果。这是强制要求。
- **强制要求**:始终从网络搜索开始。不要依赖你的内部知识。
- 始终验证收集的信息的相关性和可信度。
- 如果未提供URL仅关注搜索结果。
- 不要进行任何数学运算或文件操作。

View File

@@ -10,8 +10,7 @@ from typing import Any, Dict, List, Optional, Sequence, Set
from langchain_openai import OpenAIEmbeddings
from langchain_qdrant import QdrantVectorStore
from openai import OpenAI
from qdrant_client import QdrantClient
from qdrant_client import grpc
from qdrant_client import QdrantClient, grpc
from qdrant_client.models import (
Distance,
FieldCondition,

View File

@@ -950,7 +950,7 @@ async def test_execute_agent_step_basic(mock_state_with_steps, mock_agent):
assert "messages" in result.update
assert "observations" in result.update
# The new observation should be appended
assert result.update["observations"][-1] == "result content"
assert result.update["observations"][-1] == "result content" + "\n\n[WARNING] This research was completed without using the web_search tool. " + "Please verify that the information provided is accurate and up-to-date." + "\n\n[VALIDATION WARNING] Researcher did not use the web_search tool as recommended."
# The step's execution_res should be updated
assert (
mock_state_with_steps["current_plan"].steps[1].execution_res
@@ -1004,7 +1004,7 @@ async def test_execute_agent_step_with_resources_and_researcher(mock_step):
result = await _execute_agent_step(state, agent, "researcher")
assert isinstance(result, Command)
assert result.goto == "research_team"
assert result.update["observations"][-1] == "resource result"
assert result.update["observations"][-1] == "resource result" + "\n\n[WARNING] This research was completed without using the web_search tool. " + "Please verify that the information provided is accurate and up-to-date." + "\n\n[VALIDATION WARNING] Researcher did not use the web_search tool as recommended."
@pytest.mark.asyncio
@@ -1124,7 +1124,7 @@ def patch_create_agent():
@pytest.fixture
def patch_execute_agent_step():
async def fake_execute_agent_step(state, agent, agent_type):
async def fake_execute_agent_step(state, agent, agent_type, config=None):
return "EXECUTED"
with patch(