Embeddings & Search
Turning meaning into geometry, then searching that geometry fast enough to matter.
Prerequisites
- How LLMs Work — transformers, and the difference between an encoder and a generator.
- Tokens & Context Windows — embedding models have token limits too, and silently truncate past them.
Why this section exists
Keyword search matches strings. Users ask in meaning. Someone searching "how much time off do I get" will never match a document titled "PTO accrual policy" on shared words alone.
The fix is an old idea executed well: represent text as a point in high-dimensional space, trained so that things meaning the same thing land near each other. Then "find relevant documents" becomes "find nearby points" — and nearest-neighbour search is a problem computer science has spent decades making fast.
That single move is the engine under RAG, semantic search, recommendations, deduplication, and clustering. Which means this section is load-bearing: nearly every retrieval question eventually bottoms out in "…and why does that work?" — and the answer is here.
The chain you're building:
Each stage in that chain has its own page, in that order.
What's in here
| Page | The question it answers |
|---|---|
| Embeddings | What an embedding is, how contrastive training produces one, how to pick a model |
| Similarity Metrics | Cosine vs dot product vs Euclidean — and when the choice actually changes results |
| ANN Indexes | HNSW and IVF, quantization, and the recall/speed/memory triangle you trade along |
| Hybrid Search | Why BM25 still wins on IDs and rare terms, and how RRF fuses two ranked lists |
| Rerankers | Bi-encoder vs cross-encoder, and why two-stage retrieval is the standard shape |
| Embeddings Beyond RAG | Classification, recommendation, dedup, anomaly detection — the uses that get overlooked |
Where this connects
- Vector Databases — where these vectors live in production, and what operating that store costs you.
- RAG — the flagship consumer of everything in this section.
- Multimodal — the same geometry trick, applied to images and audio.