Web Search Tool
Tavily · Real-time Networking · Information Retrieval Agent

1Why do you need a web search? Fundamental Limitations of LLM Knowledge Cutoff

All LLMs have a fundamental limitation:Training data has a deadline (Knowledge Cutoff). GPT-4, DeepSeek, Claude - they only know the knowledge in the training set and cannot know new events and new data that occur after training.

• "Who is the principal of Birralee Primary School?" → Management subject to change
• "Who is the current mayor of San Francisco?" → Possible turnover of political offices
• "What's Apple's stock price today?" → Changing every minute
• "What is the latest GPT version?" → The AI ​​industry is developing rapidly

For this type of question, relying solely on LLM will yield answers that may be out of date or wrong.Web search toolsSolve this problem - it allows the Agent to "go out and look it up" during reasoning and bring back the latest information.

2Tavily: Search API designed for AI Agents

Tavily is a search service optimized for LLM/Agent scenarios. Compared with crawling web pages directly, it has several key advantages:

3Define web_search tool: @tool + TavilyClient

from langchain.tools import tool
from typing import Dict, Any
from tavily import TavilyClient

tavily_client = TavilyClient()  # 自动读取 TAVILY_API_KEY 环境变量

@tool("web_search", description="Search web")
def web_search(query: str) -> Dict[str, any]:
    return tavily_client.search(query, search_depth="basic")

Key parameters of this tool:

• query: str: Search keywords. The Agent will automatically generate appropriate query terms based on the user's questions.
• search_depth="basic": Quick search (suitable for most scenarios); change to"advanced"Will go deeper but slower
• return value: Tavily's JSON results, includingresults、query、answeretc fields

4Without search tools: Demonstration of the limitations of LLM

Build one firstno toolsA common Agent to see how it answers real-time questions:

agent_no_tools = create_agent(model=deepseek_model())

question = {
    "role": "user",
    "content": "Who is the principal of Birralee Primary School?"
}

response = agent_no_tools.invoke({"messages": question})

The model honestly admits not knowing – which is the best case scenario. A worse case scenario is that the modelConfidently give a wrong answer(hallucination).

5After binding the search tool: answer real questions in real time

Scenario 1: Find the school principal

# 给 Agent 绑定 web_search 工具
agent = create_agent(
    deepseek_model(),
    tools=[web_search]
)

question = {
    "role": "user",
    "content": "Who is the principal of Birralee Primary School?"
}

response = agent.invoke({"messages": question})
print(response['messages'][-1].content)

The model is called this timeweb_search("Birralee Primary School principal"), I found the answer from real web pages instead of making it up out of thin air.

Scenario 2: Find mayor information

agent = create_agent(
    model=deepseek_model(),
    tools=[web_search]
)

question = HumanMessage(content="Who is the current mayor of San Francisco?")
response = agent.invoke({"messages": [question]})
print(response['messages'][-1].content)

6Parsing search results: JSON data in ToolMessage

Tavily returns a JSON string, encapsulated asToolMessage.content. If you need to extract specific information from the search results, you can parse this JSON:

import json

# messages[2] 是 ToolMessage（工具执行结果）
content = response['messages'][2].content
data = json.loads(content)   # 字符串 → dict

from pprint import pprint
pprint(data)

Key fields for Tavily search results

7Advanced usage and precautions for search tools

Test the tool independently first, then bind the Agent

# 先独立测试 web_search，确认可以正常工作
web_results = web_search.invoke("Who is the current mayor of San Francisco?")
print(web_results['results'][0]['content'])  # 第一条结果的内容

Complete tool list example

# 可以同时绑定多个工具
agent = create_agent(
    model=deepseek_model(),
    tools=[web_search, square_root, another_tool],
    system_prompt="You have access to web search and math tools. Use them when needed."
)