LLM Classifier

AI-powered classification that understands context and handles edge cases.

Architecture

flowchart TD
    A[Complaint] --> B{Baseline Match?}
    B -->|Yes| C[Return Code]
    B -->|No| D[LLM Analysis]
    D --> E{Confidence OK?}
    E -->|Yes| F[Return Code]
    E -->|No| G[Flag for Review]

    style C fill:#c8e6c9
    style F fill:#c8e6c9
    style G fill:#ffccbc

Hybrid Approach

graph TD
    subgraph Input
        A[Complaint]
    end

    subgraph Fast Path
        B[Rule Engine]
        B -->|Match| C[ISP Code]
    end

    subgraph AI Path
        D[LLM Call]
        D --> E[Classification]
        E --> F[Confidence]
    end

    A --> B
    A --> D

    style C fill:#c8e6c9
    style F fill:#c8e6c9

Prompt Template

graph LR
    A[System] -->|Classify this complaint|
    B[User] -->|Complaint text|
    C[Assistant] -->|ISP Code + Reason|

    style C fill:#c8e6c9

Response Format

{
  "code": "ISP-002",
  "confidence": 0.92,
  "reasoning": "WiFi issues combined with router reference indicates router problem",
  "alternatives": ["ISP-001", "ISP-004"]
}

Comparison with Baseline

Aspect	Baseline	LLM
Speed	<10ms	500-2000ms
Cost	Free	API call
Coverage	Keyword-based	Contextual
Edge cases	Poor	Good
Explanation	No	Yes

When LLM Kicks In

flowchart TD
    A[Complaint] --> B[Check Keywords]
    B --> C{Exact Match?}
    C -->|Yes| D[Use Baseline]
    C -->|No| E{Partial Match?}
    E -->|Yes| F[Check Confidence]
    E -->|No| G[Use LLM]
    F -->|Low| G
    F -->|High| D

Code Example

def classify_llm(text):
    prompt = f"""Classify this ISP complaint:

    Complaint: {text}

    Categories:
    - ISP-001: ONT/Fiber issues
    - ISP-002: Router problems
    - ISP-003: DNS issues
    - ISP-004: Speed problems

    Return JSON with code, confidence, and reasoning.
    """

    response = call_llm(prompt)
    return parse_response(response)

Best Practices

Tip	Description
Fallback	Always have baseline ready
Cache	Cache common patterns
Batch	Batch similar requests
Monitor	Track low-confidence cases