Skip to content
Blog

HNSW Parameters: M, efConstruction and efSearch Tuning

A practical guide to tuning HNSW parameters: what M, efConstruction and efSearch do, how they trade memory, build time, recall and latency.

Published on • November 22, 2026

AI Assistant

Introduction

HNSW gives you three knobs — M, efConstruction, and efSearch — that control the entire memory / build-time / recall / latency budget of your index. Tuning them blindly wastes memory or recall; tuning them deliberately gets you exactly the trade-off your application needs.

The Three Knobs

M — connectivity per node

M is the maximum number of neighbours each node keeps on the base layer (upper layers use M/2 in the original design).

  • Higher M → denser graph → better recall and shorter paths, but more memory and slower builds
  • Typical range: 8–64; 16–48 covers most workloads (16 is FAISS/Qdrant’s common default; 30+ suits high-recall RAG)
  • Memory cost: roughly M × 4–8 bytes × n for the graph alone (plus upper layers) — at M=16 this can rival the vector data itself

efConstruction — build quality

During insertion, efConstruction is the size of the dynamic candidate list used to select the M neighbours.

  • Higher efConstruction → better-quality edges → better recall at any efSearch, but slower index builds
  • Typical range: 100–500 (200 is a common default)
  • Rule of thumb: set efConstruction ≥ efSearch; beyond ~400–800 the gains flatten while build time keeps climbing

efSearch — query-time search depth

At query time, efSearch is the size of the candidate list explored on the base layer. Results are truncated to k from the efSearch best found.

  • Higher efSearch → deeper exploration → higher recall, higher latency
  • Typical range: 10–512; often 64–256 for production RAG
  • Key advantage: it can be changed per query — no rebuild needed. Great for “cheap browse, expensive rerank” pipelines

A Worked Tuning Path

  1. Baseline: M=16, efConstruction=200, efSearch=64. Measure recall@10 against exact ground truth and p95 latency.
  2. Recall short? First raise efSearch (runtime-only, zero rebuild). If still short, raise M and rebuild.
  3. Latency too high with acceptable recall? Lower efSearch; only then consider reducing M.
  4. Build too slow? Lower efConstruction to ~100 and re-measure — recall often barely moves.
  5. Memory pressure? Lower M first (graph size scales linearly), or add quantisation (HNSW + PQ/HNSW+int8).

Rough Defaults by Workload

WorkloadMefConstructionefSearch
Interactive search, low latency8–16100–20032–64
General purpose1620064–128
High-recall RAG / rerank pipelines32–48200–400128–256

Conclusion

M sets the graph’s memory and structural quality, efConstruction sets build quality, and efSearch sets per-query recall vs latency. Tune in that order of cost: adjust efSearch first (free at runtime), then pay for rebuilds with M/efConstruction — and always validate against exact ground truth.

References