Intent & Classification

Why Classification Matters

Every support interaction starts with three questions:

Get these wrong, and you route a security incident to the billing team. Get them right, and you resolve 80% of tickets without human intervention.

Classification is what separates a chatbot from a support agent. Without it, the AI gives generic responses. With it, the AI knows exactly which KB article to retrieve and whether to escalate.

Key Concepts

Intent Detection

The primary classifier. Maps a customer query to one of 10+ intent categories:

The baseline classifier uses keyword pattern matching with confidence scoring. Keywords like "password", "login", "locked out" trigger the password_reset intent. Multiple matching keywords increase confidence.

The Confidence Problem

Confidence is the most important output of classification. It's not just "how sure are we?" — it's a routing signal:

Low confidence isn't a failure — it's a safety mechanism. The worst outcome in support is a confident wrong answer. A confused escalation is always better.

Priority Scoring

Priority combines three signal types:

The score (0-10) maps to priority levels: urgent (8+), high (6-7), medium (4-5), low (1-3).

Entity Extraction

Structured data embedded in natural language:

Entity extraction makes responses feel personalized: "I can see your account ACC-4521..." instead of "Please provide your account ID."

Topic Classification

Maps queries to product areas (authentication, billing, API, mobile, workspace, data, infrastructure). Topics drive two things:

Architecture Pattern

Customer Query
    │
    ├──→ Intent Detector ──→ { primary, confidence, secondary }
    │
    ├──→ Priority Scorer ──→ { priority, score, reasons }
    │
    ├──→ Topic Classifier ──→ { topic, confidence }
    │
    └──→ Entity Extractor ──→ { type, value, start, end }[]

All four classifiers run in parallel on the same query. Together they produce a complete classification profile in <5ms.

What You'll Build

Glossary