Vector databases, vector search, and semantic search

🔴 MUST KNOW · 🟠 SHOULD KNOW for hybrid search and reranking.

Overview

• A vector database stores vectors and supports similarity queries with metadata.
• Vector search is the numerical operation; semantic search is the application goal of finding meaningfully related content.
• Existing databases can also add vector indexes, so you do not always need a separate product.

Why this matters

Comparing a query with every document becomes expensive as the corpus grows.

An everyday example

An organized warehouse can find a nearby matching item without checking every shelf individually.

A technical example

Query the index for the 20 closest chunks where tenantId=t1 and documentStatus=active.

How it works

1. Store embeddings
2. Build a search index
3. Encode query
4. Apply supported filters
5. Search candidates
6. Return text IDs and scores.

• Approximate nearest neighbor, or ANN, searches efficiently while accepting that it may miss some true closest results.

Architecture and flow

Query → embedding → filtered vector index → top-k chunks → application.

The insurance assistant example

Retrieve relevant active motor-policy clauses for the signed-in customer's insurer and product.

When to use it

For repeated similarity queries over a corpus too large or too active for simple scans.

When to avoid it

A few hundred records may work with exact in-memory comparison; relational or document queries are better for exact IDs.

Interview answer

“A vector index accelerates similarity search. I choose exact or approximate search based on scale, filtering, recall, and latency requirements.”

Benefits, trade-offs, and production considerations

• HNSW is a graph-like ANN index; IVF partitions vector space into groups.
• Both trade recall, memory, construction cost, and query time.
• Benchmark with real tenant filters: global recall numbers can conceal poor filtered search.