vector-index-tuning

When to Use

Trigger this skill when building RAG systems, optimizing vector search performance, or choosing vector index solutions.

Core Decisions

Index Selection

Method	Recall	Latency	Memory	Use Case
HNSW	High	Low	High	<1M vectors, low-latency priority
IVF-PQ	Medium	Medium	Low	Large scale, memory-constrained
Flat	Highest	High	Highest	<100K, precision priority

HNSW Tuning

• M (connections): 16-64, higher recall but more memory
• efConstruction (build search width): 128-256, affects index build time
• efSearch (query search width): top-k to 2*M, tunable at runtime
• Rule of thumb: M=32, efConstruction=200 for most scenarios

Quantization

• PQ (Product Quantization): 8-32x compression, 2-5% recall loss
• SQ (Scalar Quantization): 4x compression, <1% recall loss
• Combination: IVF + PQ for ultra-large scale

RAG Tuning

• Fetch top-k of 20-50, rerank to 5-10
• Use same model for embedding and query
• Chunking: 512-1024 tokens, 10-20% overlap