Vector Database

01What is the core query — what goes in, what comes out?

Given a query vector, return the k nearest neighbors by the collection’s distance metric — ids, similarity scores, and stored payloads, with k set per query.

02Can a query say "nearest neighbors WHERE tenant = X"?

Yes: queries combine a metadata filter with vector similarity and the engine still returns k results — filtered search is designed into the index, not bolted on after.

03Is this a build-once index, or do writes land on a live one?

Live: upserts become searchable within seconds without a full rebuild, and a deleted vector stops appearing in results immediately, even if the index frees its memory later.

04Which distance metric — and who picks it?

Cosine, dot product, or L2, fixed per collection at creation — the metric must match how the embedding model was trained, or result quality silently degrades.

05Is accuracy fixed, or a knob the caller can turn?

A knob: each collection declares a recall target, and each query may trade accuracy for speed (efSearch or nprobe) — approximate search is an explicit contract, not an accident.

06One giant index, or many isolated collections?

Many collections, each with its own dimension, metric, index type, and quantization — one tenant’s hot collection cannot starve another’s.

01How fast, at what scale?

p99 query latency around 50 ms at a billion vectors and thousands of QPS — answers come from RAM-resident index structures, never disk scans.

02Approximate — how wrong are we allowed to be?

Recall@10 of at least 0.95, measured against brute-force ground truth on a sampled query set — approximate means measurably wrong sometimes, so recall is a monitored SLO.

03The corpus doubles next year — then what?

Capacity scales horizontally: vectors shard across nodes and queries scatter-gather, so no single node ever holds the whole corpus.

04Does heavy churn quietly rot the index?

Sustained churn does not rot the index: compaction keeps tombstone ratios bounded, and recall decay triggers rebuilds before users notice.

05A node dies — did we just lose vectors?

A node loss never loses vectors: writes are WAL (write-ahead log) durable and raw vectors replicated — the ANN index is a rebuildable acceleration structure, not the source of truth.