xb

Qdrant Optimization Summary (English)

English adaptation of xb/doc/QDRANT_OPTIMIZATION_SUMMARY.md. It captures tuning tips for latency, recall, and diversity.

---

Performance levers

• HNSW params – expose HnswEf, ScoreThreshold, Exact flags via QdrantCustom.
• Payload projection – include only necessary fields to reduce payload size.
• Vector normalization – pre-normalize to keep scoring predictable.
• Batching – send bulk inserts/updates instead of single record calls.

---

Recall vs latency

Goal	Recommended settings
Low latency	Lower limit, moderate HnswEf, enable payload trimming
High recall	Increase HnswEf, limit, disable payload trimming
Balanced	Default NewQdrantBuilder().Build() plus WithHashDiversity

---

Observability

• Record TraceID, TenantID, and adapter config in logs.
• Track P95 latency per API (Search, Recommend, Discover, Scroll).
• Monitor score distributions to catch vector drift.

---

Related docs

• doc/en/VECTOR_GUIDE.md
• doc/en/QDRANT_ADVANCED_API.md
• doc/en/AI_APPLICATION.md

This site is open source. Improve this page.