From Text to Vector Space

Embeddings map text into dense vectors where semantic similarity is measurable as geometric distance. "Reset my password" and "I can't log into my account" land close together in vector space despite sharing almost no words - that is the property every semantic retrieval system is built on.

How Similarity Is Measured

# Cosine similarity: 1.0 = identical direction, 0 = unrelated
similarity(v1, v2) = (v1 · v2) / (|v1| � - |v2|)

"reset my password"        ↔ "can't log in"          → 0.87  (same intent)
"reset my password"        ↔ "quarterly revenue up"   → 0.09  (unrelated)

Embedding Models on Bedrock

Engineering Realities

• Same intent, different wording should cluster closely - test this with your own domain phrases, not generic benchmarks.
• Domain vocabulary matters - internal product names and acronyms may embed poorly; consider glossary expansion at ingestion.
• Dimensions trade cost for fidelity - higher dimensions = better separation but larger indexes and slower search.
• You cannot mix embedding models - vectors from different models live in incompatible spaces. Re-embed the whole corpus when you switch, and version your index by embedding model.
• Index maintenance matters as much as model quality - deleted docs must leave the index, updated docs must be re-embedded.