03 · vector

Four metrics. Seven index variants. One write.

Every vector lives inside the same atomic write as the row it describes. HNSW, IVF, IVF-PQ, binary and PQ quantization, sparse, plus predicate-aware over-fetch with `min_score` thresholds that admit empty as a real result.

distance metrics

Euclidean (L2)

Default. Geometric distance.

Inner Product

For pre-normalized embeddings.

Cosine

Direction only. Magnitude ignored.

Manhattan (L1)

Robust to outliers. Categorical embeddings.

feature overview

index variants

HNSW

IVF

IVF-PQ

Binary quant

PQ quant

Sparse

0.96

recall@10

64x

less memory

< 30 ms

p99 kNN

topk(embedding, :q, 10)
  WHERE in_stock = true
  AND   region  = 'EU'
  MIN_SCORE 0.65

sku-8842 0.84

sku-1207 0.71

sku-0042 0.58 · below min_score

Below the threshold is dropped - an empty result is a real result.

what you can do

RAG pipelines

Ground LLM answers in your own data - retrieve, filter and rank in one call.

Recommendations

Blend vector similarity with graph signals and SQL filters for relevance.

Semantic search

Meaning-based search over documents, products or logs, not just keywords.

Agent memory

Durable, queryable long-term memory for agents - embeddings beside the facts.

measured

4 x 7

metrics x index variants

< 30 ms

p99 kNN @ k=10, n=1M dense

64x

memory savings on IVF-PQ at D=128

embedding-vs-row consistency lag

Read the vector docs, then try it.

Vector docs → Architecture Try it