Building a Real-Time Web Search Agent with LangChain and Tavily

Large Language Models (LLMs) are powerful, but they have a fundamental limitation: they lack real-time awareness. Their responses are constrained by training data cutoffs, which makes them unsuitable for use cases that require up-to-date information.

To address this, this project demonstrates a LangChain-based web search agent capable of deciding when to search the internet, retrieving fresh data, and synthesizing accurate answers with explicit source citations using the Tavily Search API.

👉 GitHub repository langchain-tavily-search-agent

🚀 What This Project Does

This project implements an AI-powered search agent that:

• Understands a user’s question using a large language model
• Determines whether external web search is required
• Retrieves up-to-date information from the internet
• Synthesizes a coherent final answer
• Displays the source URLs used to generate the response

Unlike simple “search wrappers”, this solution follows an agentic architecture, combining reasoning and tools, which is a foundational pattern in modern AI systems.

🧠 Core Features

🤖 Agent-based reasoning using LangChain

🔍 Real-time web search via Tavily (optimized for LLMs and RAG)

📚 Automatic source citations

🛠️ Custom LangChain tools using the @tool decorator

🎯 Clean, structured output

⚙️ Simple and extensible architecture

🧩 How It Works (High Level)

• A user submits a question
• The LangChain agent analyzes the intent
• The agent decides whether to invoke the search tool
• Tavily performs a web search (up to 5 results)
• Results are returned with title, URL, and content snippets
• The LLM synthesizes a final response grounded in those sources

This pattern is well-suited for RAG pipelines, research agents, and AI assistants that require current information.

🛠️ Key Implementation Details

Custom Search Tool

The Tavily integration is exposed as a LangChain tool:

@tool
def search(query: str) -> str:
    response = tavily.search(query, max_results=5)
    ...

This enables the agent to autonomously decide when a web search is necessary rather than forcing every query through a search step.

Language Model Configuration

llm = ChatOpenAI(
    model=os.getenv("OPENAI_MODEL"),
    temperature=0.2
)

• Low temperature ensures focused, deterministic responses
• The model is configurable via environment variables
• Compatible with OpenAI-compatible endpoints

Agent Creation

agent = create_agent(model=llm, tools=[search])

LangChain manages:

• Tool invocation
• Message history
• Reasoning steps
• Final response synthesis

▶️ Running the Agent

python main.py

Example output includes:

📚 Sources Used (URLs)

🤖 Final Answer, synthesized from multiple sources

📦 Tech Stack

• Python 3.13
• LangChain >= 1.2.0
• OpenAI (via langchain-openai)
• Tavily Search API
• uv for fast dependency management