🎯 What You'll Learn Today

LangGraph Tutorial: Advanced Message Classification with State Management - Unit 1.3 Exercise 1

This tutorial is also available in Google Colab here or for download here

Joint Initiative: This tutorial is part of a collaboration between AI Product Engineer and the Nebius Academy.

This tutorial demonstrates how to implement sophisticated message classification in LangGraph using TypedDict and proper annotations. We'll create a state structure that combines message handling with classification metadata for enhanced conversation understanding.

Key Concepts Covered

1. Classification-Aware State Management
2. Confidence Score Tracking
3. Type-Safe Message Handling
4. Message Classification Metadata

from typing import Annotated, TypedDict
!pip install langchain-core
!pip install langgraph
from langchain_core.messages import BaseMessage, HumanMessage
from langgraph.graph.message import add_messages

Step 1: Enhanced State Definition with Classification

We define a state structure that combines message handling with classification metadata for richer conversation understanding.

class State(TypedDict):
    """Advanced state container with classification capabilities.

    This implementation demonstrates three key concepts:
    1. Message storage with proper LangGraph annotations
    2. Classification category tracking
    3. Confidence score management

    Attributes:
        messages: List of conversation messages with proper message
                 handling annotation
        classification: Category label for message content
        confidence: Numerical confidence score for classification

    Note:
        The combination of classification and confidence enables
        more nuanced conversation handling and decision making.
    """

    messages: Annotated[list[BaseMessage], add_messages]
    classification: str
    confidence: float

Why This Matters

Classification-aware state management is crucial because:

1. It enables intent-based conversation routing
2. Supports confidence-based decision making
3. Enables more sophisticated agent behaviors
4. Facilitates conversation analytics

Step 2: State Usage and Type Safety

Demonstrate proper state initialization and usage patterns.

Initialize state with classification metadata
initial_state: State = {
    "messages": [HumanMessage(content="hello there")],
    "classification": "greeting",
    "confidence": 0.9,
}

Debug Tips

1. Classification Issues:

• Verify classification string formatting
• Check confidence score range (0.0 to 1.0)
• Validate message content types

2. State Management:

• Monitor classification updates
• Track confidence score changes
• Verify message handling

3. Common Errors:

• TypeError: Check field type consistency
• KeyError: Verify all required fields
• ValueError: Validate confidence scores

Key Takeaways

1. State Design:

• Include both content and metadata
• Use proper type annotations
• Maintain field consistency

2. Classification Management:

• Track message categories
• Maintain confidence scores
• Enable informed decision making

Common Pitfalls

1. Missing type annotations
2. Improper confidence score range
3. Inconsistent classification labels
4. Incomplete state initialization

Next Steps

1. Add classification validation
2. Implement confidence thresholds
3. Add classification history
4. Enable multi-label classification

Example Usage

example_state: State = {
    "messages": [HumanMessage(content="hello there")],
    "classification": "greeting",
    "confidence": 0.9,
}
print("Message:", example_state["messages"][0].content)
print("Classification:", example_state["classification"])
print("Confidence:", example_state["confidence"])

Expected Output

Message: hello there Classification: greeting Confidence: 0.9

Variations and Extensions

1. Enhanced Classification:

• Multi-label classification support
• Hierarchical classification schemes Example use case: Complex intent recognition

2. Advanced Confidence:

• Confidence history tracking
• Threshold-based routing Scenario: High-stakes decision making

Rod Rivera

🇬🇧 Chapter