Skip to main content

Assignment: FPT Customer Chatbot - Multi-Agent System

Assignment Metadata​

FieldDescription
Assignment NameFPT Customer Chatbot - Multi-Agent System
CourseLangGraph and Agentic AI
Project Namefpt-chatbot-multi-agent
Estimated Time240 minutes
FrameworkPython 3.10+, LangGraph, LangChain, SQLite/PostgreSQL, Tavily API, OpenAI

Learning Objectives​

By completing this assignment, you will be able to:

  • Design hierarchical multi-agent architectures with Primary Assistant and specialized agents
  • Implement dialog stack management for agent transitions
  • Apply context injection patterns for user information
  • Build CompleteOrEscalate pattern for agent decision making
  • Create production-ready tool schemas with Pydantic validation

Problem Description​

Build a customer service chatbot for FPT with the following capabilities:

  1. FAQ Agent: Answer questions about FPT policies using RAG
  2. Ticket Support Agent: Create, track, update, cancel support tickets
  3. IT Support Agent: Troubleshoot technical issues using Tavily search
  4. Booking Agent: Manage meeting room reservations

The Primary Assistant coordinates routing between specialized agents.

Technical Requirements​

Environment Setup​

  • Python 3.10 or higher
  • Required packages:
    • langgraph >= 0.2.0
    • langchain >= 0.1.0
    • sqlalchemy >= 2.0.0
    • pydantic >= 2.0.0
    • tavily-python >= 0.3.0

Database Schema​

Implement the following tables:

  • Ticket: ticket_id, content, description, customer_name, customer_phone, email (optional), status, time
  • Booking: booking_id, reason, time, customer_name, customer_phone, email (optional), note, status

Tasks​

Task 1: State & Context Management (20 points)​

  1. Define AgenticState with:

    • messages: Conversation history with add_messages
    • dialog_state: Stack tracking agent hierarchy
    • user_id, email (optional): Context injection fields
    • conversation_id: Session tracking

  2. Implement dialog stack functions:

    • update_dialog_stack: Push/pop agent transitions
    • pop_dialog_state: Return to Primary Assistant

  3. Create context injection that auto-populates user info into tool calls

Task 2: Ticket Support Agent (25 points)​

  1. Define Pydantic schemas for:

    • CreateTicket: content (required), description, customer_name, phone, email (optional)
    • TrackTicket: ticket_id
    • UpdateTicket: ticket_id (required), optional fields for updates

  2. Implement tools:

    • create_ticket: Creates ticket with "Pending" status
    • track_ticket: Returns ticket information
    • update_ticket: Updates provided fields only
    • cancel_ticket: Sets status to "Canceled"

  3. Add CompleteOrEscalate tool for returning to Primary Assistant

Task 3: Booking Agent (20 points)​

  1. Define Pydantic schemas with validation for:

    • BookRoom: reason (required), time (required, must be future)
    • TrackBooking, UpdateBooking, CancelBooking

  2. Implement booking tools with status management:

    • Status flow: Scheduled → Finished (or Canceled)

Task 4: IT Support & FAQ Agents (15 points)​

  1. IT Support Agent:

    • Uses Tavily search for troubleshooting information
    • Returns practical guides from reliable sources

  2. FAQ Agent:

    • Implements RAG for FPT policy questions
    • Returns answers with source references

Task 5: Primary Assistant & Routing (20 points)​

  1. Define routing tools for Primary Assistant:

    • ToTicketAssistant, ToITAssistant, ToBookingAssistant
    • Include user context injection in tool calls

  2. Implement entry nodes for agent transitions:

    • Create create_entry_node factory function
    • Push new agent to dialog_state stack

  3. Build complete graph with:

    • All agent nodes and their tool nodes
    • Conditional routing from Primary Assistant
    • Tool fallback handling for errors

  4. Create tool_node_with_fallback for graceful error handling

Submission Requirements​

Required Deliverables​

  • Source code with all agents implemented
  • Database schema (SQLite or PostgreSQL)
  • README.md with setup and usage instructions
  • Test conversations demonstrating each agent flow

Submission Checklist​

  • All 4 specialized agents working correctly
  • Context injection auto-populates email when available
  • CompleteOrEscalate enables return to Primary Assistant
  • Tool schemas validate inputs properly
  • Dialog stack correctly tracks agent hierarchy
  • Code runs without errors

Evaluation Criteria​

CriteriaPoints
State & context management20
Ticket Support Agent25
Booking Agent20
IT Support & FAQ Agents15
Primary Assistant & routing15
Code quality and documentation5
Total100

Hints​

tip
  • Use ToolNode(tools).with_fallbacks([...]) for error handling
  • The dialog_state stack enables proper agent hierarchy tracking
  • Optional email fields should use Optional[EmailStr] in Pydantic schemas
  • Reference the LangGraph customer support tutorial for architecture patterns

References​