Web Search & Live Data Tools

pip install langchain langchain-openai langchain-community tavily-python

Now, let's just see what Tavily does. We won't even use the LLM yet. We will just ask Python to search the web.

import os
from langchain_community.tools.tavily_search import TavilySearchResults

# 1. Set up keys
os.environ["OPENAI_API_KEY"] = "your_openai_key_here"
os.environ["TAVILY_API_KEY"] = "your_tavily_key_here"

# 2. Initialize the search tool
# k=3 means we want the top 3 results
search_tool = TavilySearchResults(k=3)

# 3. Run a search manually to see what happens
print("Searching for recent news...")
results = search_tool.invoke("Winner of the 2024 Super Bowl")

# 4. Inspect the raw data
for result in results:
    print(f"\nSource: {result['url']}")
    print(f"Content snippet: {result['content'][:200]}...") # Printing first 200 chars

Step 2: Binding the Tool to the LLM

We cannot just paste the search results into the prompt manually every time. We want the LLM to decide when to search.

We use a concept called binding. We tell the LLM: "Here is a list of functions you can call if you get stuck."

from langchain_openai import ChatOpenAI
from langchain_community.tools.tavily_search import TavilySearchResults

# Initialize Tool
search_tool = TavilySearchResults(k=3)
tools = [search_tool]

# Initialize LLM
llm = ChatOpenAI(model="gpt-4o")

# Bind the tool to the LLM
llm_with_tools = llm.bind_tools(tools)

# Ask a question that REQUIRES the tool
query = "What is the current weather in Tokyo?"
response = llm_with_tools.invoke(query)

print(f"Content: {response.content}")
print(f"Tool Calls: {response.tool_calls}")

The LLM is saying: "I don't know the answer, but I see you gave me a search tool. Please run this search for me."

Step 3: Creating the Agent Executor

Handling that back-and-forth (LLM asks for search -> We run search -> We give results back -> LLM answers) is tedious to code manually.

LangChain provides an Agent Executor to handle this loop automatically. It acts as the manager that listens to the LLM's requests and executes the tools.

from langchain_openai import ChatOpenAI
from langchain_community.tools.tavily_search import TavilySearchResults
from langchain.agents import AgentExecutor, create_tool_calling_agent
from langchain_core.prompts import ChatPromptTemplate

# 1. Setup Tools
search_tool = TavilySearchResults(k=5) # Get 5 results for better summary
tools = [search_tool]

# 2. Setup LLM
llm = ChatOpenAI(model="gpt-4o", temperature=0) # Temp 0 for factual accuracy

# 3. Create the Prompt
# We instruct the AI to be a news assistant
prompt = ChatPromptTemplate.from_messages([
    ("system", "You are a helpful news assistant. You verify facts using the search tool. Always cite your sources with URLs."),
    ("human", "{input}"),
    ("placeholder", "{agent_scratchpad}"), # Required: This is where the agent's "thought process" goes
])

# 4. Construct the Agent
agent = create_tool_calling_agent(llm, tools, prompt)

# 5. Create the Executor (The runtime)
agent_executor = AgentExecutor(agent=agent, tools=tools, verbose=True)

# 6. Run it!
print("--- Agent Running ---")
response = agent_executor.invoke({
    "input": "What are the latest updates on the release of GPT-5? Summarize the rumors."
})

print("\n--- Final Response ---")
print(response["output"])

Step 4: Formatting and Citing Sources

The previous step works, but sometimes the AI gets lazy with citations. Let's refine the prompt to force a specific structure. This is "Prompt Engineering" applied to Agents.

# Modified Prompt for strict formatting
prompt = ChatPromptTemplate.from_messages([
    ("system", """You are a rigorous research assistant. 
    1. Search for the latest information on the user's topic.
    2. Summarize the key points in bullet points.
    3. You MUST provide a 'Sources' section at the bottom with clickable URLs.
    4. If you cannot find information, admit it. Do not make things up."""),
    ("human", "{input}"),
    ("placeholder", "{agent_scratchpad}"),
])

agent = create_tool_calling_agent(llm, tools, prompt)
agent_executor = AgentExecutor(agent=agent, tools=tools, verbose=False) # verbose=False for clean output

topic = "Current mission status of Voyager 1"
print(f"Researching: {topic}...")

response = agent_executor.invoke({"input": topic})

print(response["output"])

You should now see a clean, professional summary followed by a list of URLs. This is a functional application you could actually use to catch up on news.

Now You Try

You have a working web-search agent. Now, extend its capabilities.

Change the prompt to make the agent a "Shopping Assistant." Ask it to find the current price of a specific gadget (e.g., "Sony WH-1000XM5 headphones") at 3 different major retailers and output a table comparing prices. Hint: You might need to ask it specifically to search for 'Best Buy price', 'Amazon price', etc.

Feed the agent a statement that might be false (e.g., "The Eiffel Tower was demolished in 2024"). Configure the system prompt to be a "Debunker." It should search for the claim and output: "TRUE", "FALSE", or "UNVERIFIED" along with evidence.

Challenge Project: The "Deep Dive" Research Assistant

Your goal is to build a robust CLI (Command Line Interface) tool that takes a user's question and produces a mini-report.

* Ask the user if they want a "Quick Summary" (1 search) or a "Deep Dive" (force the agent to perform at least 2 distinct searches on different aspects of the topic).

* Print the output nicely formatted.

* Crucial Requirement: The output must verify the date. If the user asks for "news," check that the articles are actually from the current year/month.

> Topic: "The impact of AI on healthcare"

> Mode: Deep Dive

> Research Report: AI in Healthcare

>

> Key Trends:

> * Diagnostic imaging improvements...

> * Drug discovery acceleration...

>

> Recent Developments (2024-2025):

> * Google's Med-PaLM updates...

>

> Sources:

> 1. [Nature Medicine Article](http://...)

> 2. [TechCrunch Report](http://...)

* For the "Deep Dive," you might need a loop. Or, you can prompt the agent: "Search for the general topic first, then perform a second search specifically for 'criticisms' or 'risks' of this topic."