Skip to main content

Assignment: Tool Calling & Tavily Search Integration

Assignment Metadata​

FieldDescription
Assignment NameTool Calling & Tavily Search Integration
CourseLangGraph and Agentic AI
Project Nametavily-research-agent
Estimated Time90 minutes
FrameworkPython 3.10+, LangGraph, LangChain, Tavily API, OpenAI API

Learning Objectives​

By completing this assignment, you will be able to:

  • Understand Tool/Function Calling mechanics in LLMs
  • Integrate Tavily Search API for real-time web information
  • Build agents with multiple tools working in parallel
  • Implement error handling and retry logic for tool execution
  • Apply caching and rate limiting for production readiness

Problem Description​

Building on the Multi-Expert Research Agent from Assignment 02, you will add a Web Search tool using Tavily API. This enables the agent to:

  1. Access real-time information beyond LLM knowledge cutoff
  2. Combine expert analysis with current web data
  3. Execute multiple tools in parallel when needed

Technical Requirements​

Environment Setup​

  • Python 3.10 or higher
  • Required packages:
    • langgraph >= 0.2.0
    • langchain >= 0.1.0
    • langchain-community >= 0.1.0
    • tavily-python >= 0.3.0

API Requirements​

  • OpenAI API key (OPENAI_API_KEY)
  • Tavily API key (TAVILY_API_KEY) - Get free at tavily.com

Tasks​

Task 1: Tavily Search Tool Setup (25 points)​

  1. Configure TavilySearchResults tool with:

    • max_results: 5
    • search_depth: "advanced"
    • include_answer: True (for AI-generated summary)

  2. Create a wrapper tool with proper description:

    • Clear description of when to use (current events, real-time data)
    • Specify input format expectations
    • Handle search errors gracefully

Task 2: Multi-Tool Agent Integration (35 points)​

  1. Extend existing Research Agent to include:

    • AI Research Expert (from Assignment 02)
    • Financial Analyst (from Assignment 02)
    • Web Search Tool (new Tavily integration)

  2. Update coordinator prompt to:

    • Advise when to use web search vs expert LLMs
    • Enable parallel tool calling for independent queries
    • Guide synthesis of web results with expert analysis

  3. Ensure parallel execution capability:

    • Coordinator can call multiple tools in single response
    • ToolNode executes all tool_calls and returns all ToolMessages

Task 3: Error Handling & Optimization (25 points)​

  1. Implement retry logic for Tavily API:

    • Use tenacity for exponential backoff
    • Handle rate limits and timeouts gracefully

  2. Add basic caching for search results:

    • Cache identical queries for 1 hour
    • Use hash-based cache keys

  3. Implement rate limiting:

    • Max 10 searches per minute
    • Queue overflow handling

Task 4: Testing & Validation (15 points)​

  1. Test queries requiring web search:

    • "What are the latest AI announcements this week?"
    • "Current NVIDIA stock performance and recent news"

  2. Test combined expert + web search:

    • "Analyze the impact of recent AI regulations on tech companies"

  3. Verify parallel tool execution:

    • Show both expert and search tools called simultaneously
    • Document execution time comparison (parallel vs sequential)

Submission Requirements​

Required Deliverables​

  • Source code in Python script or Jupyter notebook
  • README.md with setup instructions and API key configuration
  • Test output showing web search integration
  • Comparison of response quality with/without web search

Submission Checklist​

  • Tavily tool configured with proper parameters
  • All three tools work together (2 experts + web search)
  • Parallel tool execution demonstrated
  • Error handling implemented for API failures
  • Code runs without errors

Evaluation Criteria​

CriteriaPoints
Tavily Search tool setup25
Multi-tool integration35
Error handling & optimization25
Testing & validation10
Code quality and documentation5
Total100

Hints​

tip
  • Use TavilySearchResults from langchain_community.tools.tavily_search
  • For parallel calls, the coordinator returns multiple tool_calls in one AIMessage
  • Store Tavily API key in environment variable, never hardcode
  • Use @retry(stop=stop_after_attempt(3)) from tenacity for retries

References​