Performance Tuning

Vector search performance is defined by the recall-latency-memory triangle — you can optimize any two at the expense of the third.

---

1. The Recall-Latency Tradeoff

Configuration	Recall	Latency (1M vectors)	Memory
Flat (exact)	100%	10 ms	100%
HNSW (ef=64)	98%	0.5 ms	120% (graph overhead)
HNSW (ef=16)	92%	0.2 ms	120%
IVF (nprobe=32)	96%	0.8 ms	100%
IVF + PQ	90%	0.3 ms	10% (compressed!)

Product Quantization (PQ): Compress vectors from 768 floats (3 KB) to 96 bytes — 32× memory reduction. Enables billion-scale in-memory search.

---

2. Key Tuning Parameters

Parameter	Higher value →	Lower value →
HNSW ef_search	Better recall, slower search	Faster search, lower recall
HNSW M (connections per node)	Better recall, more memory	Less memory, lower recall
IVF nprobe	Better recall, slower search	Faster search, lower recall
IVF nlist (partitions)	Faster per-partition, need more nprobe	Fewer partitions, each larger
PQ num_subvectors	Better compression, lower recall	More precise, more memory

Tuning Protocol

# Sweep ef_search to find optimal recall-latency point
for ef in [16, 32, 64, 128, 256]:
    index.set_ef(ef)
    recall = measure_recall(index, queries, ground_truth)
    latency = measure_latency(index, queries)
    print(f"ef={ef}: recall={recall:.3f}, latency={latency:.2f}ms")
    
# Target: minimum ef that achieves recall > 0.95

---

3. Batch Query Optimization

For RAG with many concurrent queries:

# Single query: 0.5ms × 100 queries = 50ms total (sequential)
# Batch query: 100 queries in one call = 5ms total (parallel!)

results = index.search(query_batch, k=10)  # batch of 100 queries
# 10× speedup from batching

---

4. Production Performance Targets

Use case	Latency target	Recall target	Scale
RAG (real-time)	< 10 ms P99	> 95%	1M-100M
Search (web)	< 50 ms P99	> 90%	100M-1B
Recommendation	< 20 ms P99	> 85%	10M-100M
Deduplication (batch)	Throughput-focused	> 99%	1B+

For operational concerns (index build, updates), see Operational Concerns.

---

Self-Check

• Explain the recall-latency-memory triangle
• Tune HNSW ef_search for a target recall of 95%
• Explain Product Quantization and its 32× memory reduction
• Calculate batch query speedup for a given query count

---

References

• ANN Benchmarks
• Faiss Performance Guide