Evaluating and Tuning Vector Search

Metric Definitions

Score your retrieval before tuning it.

• Recall@k - Of the truly relevant items, what fraction appears in the top k?
• Precision@k - Of the top k results, what fraction is relevant?
• Latency - Time per query at your intended load.
• Memory - Index size relative to raw vectors.

Build a Test Set

Create a few dozen real questions with known answers from your corpus. Run every retrieval change against that set so improvements are measurable, not guessed.

Tuning HNSW

# larger efSearch improves recall at higher cost
hnsw.hnsw.efSearch = 128
hnsw.hnsw.efConstruction = 200

Raising M and efConstruction at build time improves graph quality; raising efSearch improves recall at query time.

Trading Recall for Speed

• Raise nprobe for IVF indexes.
• Add quantization only when memory is the bottleneck.
• Prefer recall@5 over recall@1 targets.
• Benchmark at your real data size; small tests lie.

Key Points

• Measure recall and latency before and after each change.
• A fixed test set turns tuning into science.
• efSearch and efConstruction control HNSW behaviour.
• Only add quantization when memory forces you to.