Knowledge Retrieval

Why Retrieval Matters

Classification tells you WHAT the customer needs. Retrieval finds the answer. This is the step that determines whether the AI gives a helpful, specific response or a generic "please check our help center."

A great retrieval system finds "How to Reset Your Password" (KB-001, 342 helpful votes) within milliseconds when someone asks "I can't log in after changing my password." A bad one returns irrelevant results or nothing at all.

Key Concepts

The Two-Path Architecture

Support retrieval uses two paths, and always tries the fast one first:

Path 1: FAQ Fast Path (< 1ms)

80% of support queries are variations of 15-20 common questions. Title and tag matching identifies these instantly without running expensive vector search. "How do I reset my password?" matches "How to Reset Your Password" via direct title overlap.

Path 2: Full Hybrid Search (50-200ms)

For queries that don't match a known FAQ, combine keyword search with semantic search:

Source Type Boosting

Not all sources are equally authoritative for answering questions:

This means a KB article about password resets always outranks a 6-month-old resolved ticket about the same topic.

Popularity Signals

For KB articles, helpful votes provide a crowd-sourced relevance signal. An article with 342 helpful votes is more likely to be the right answer than one with 5 votes. This is the same principle as PageRank — popular content is more likely to be relevant.

The popularity boost is capped at 0.1 to prevent a popular-but-outdated article from outranking a new-but-accurate one.

Snippet Extraction

KB articles are long (500-2,000 words), but the response should cite a specific, relevant passage. Snippet extraction:

The snippet includes character offsets for UI highlighting and a relevance score for confidence calculation.

Architecture Pattern

Customer Query
    │
    ├──→ FAQ Matcher ──→ Found? ──→ Return article + snippet
    │                       │
    │                       No
    │                       │
    └──→ Hybrid Search ──→ Score & Rank ──→ Top-K results
              │
              ├── Keyword match (title + content)
              ├── Source type boost
              ├── Popularity boost
              └── Pre-filter by topic/category

Category-Aware Search

When the classifier says the topic is "authentication", the retriever can limit search to authentication-related documents before scoring. This narrows the search space and improves precision.

Query Expansion

"Can't log in" and "unable to access my account" mean the same thing but share no words. Query expansion adds synonyms from a support terminology map:

What You'll Build

Glossary