Semantic Search

Introduction

Semantic search finds documents by meaning, not keywords. 'Cancel my booking' matches 'How do I refund a reservation?' — even though no words overlap.

Beginner analogy: Google with telepathy — it understands what you meant, not just what you typed.

Understanding the topic

Core concepts to understand:

• Embed query + corpus into the same vector space.
• Find nearest neighbours (cosine similarity).
• Combine with keyword search (BM25) = hybrid search.
• Re-rank top-k with a cross-encoder for best precision.

Syntax reference

Visual workflow / architecture:

Query: "cancel my booking"
       │ embed
       ▼
   [vector]
       │
       ▼ nearest neighbour search
   ┌─────────────────────────────┐
   │ "How to refund reservation" │
   │ "Cancellation policy"       │
   │ "Manage trips"              │
   └─────────────────────────────┘

Real-world use

Powers Notion search, Slack AI search, Algolia AI, Spotify recommendations, and every modern docs site Q&A.

Best practices

• Use hybrid search (vector + keyword) for best recall.
• Re-rank top results with a cross-encoder for precision.
• Always show the snippets you matched on.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. How does semantic search differ from keyword search?
• Q2. What is hybrid search?
• Q3. What is a cross-encoder re-ranker?