GLOSSARY

Vector Search

Vector search finds semantically similar items by comparing embeddings instead of exact keywords. It helps shoppers when wording differs from catalog language.

What is Vector Search?

Vector search retrieves items by similarity in embedding space. Queries and documents are encoded as vectors; the engine finds nearest neighbors using metrics like cosine or inner product, typically via ANN indexes.

How It Works (quick)

  • Encode: Turn queries/products into embeddings (domain-tuned).
  • Index: Build an ANN structure (e.g., HNSW/IVF).
  • Retrieve: Get top-k nearest vectors; apply filters (stock/ACL).
  • Rank: Blend vector score with lexical/LTR signals (hybrid).

Why It Matters in E-commerce

  • Bridges synonyms and paraphrases (“joggers” ↔ “sweatpants”).
  • Improves long-tail and cross-language recall.
  • Powers recommendations and similar-items widgets.

Best Practices

  • Use domain-specific embeddings; refresh regularly.
  • Always run under hard filters (stock, region).
  • Combine with BM25 and phrase/brand guards.
  • Monitor drift and p95/p99 latency.

Summary

Vector search adds meaning-level recall. Pair it with lexical precision, enforce filters, and tune embeddings to your catalog.