Vector Index Build Configuration

Data artifactKnowledge & DataKnowledge & DataVariation point (abstract)arc:VectorIndexConfig

A build-time configuration fixing a vector collection's index type, graph connectivity (M), construction candidate-list size (efConstruction) and distance metric; changing it requires full re-indexing.

Responsibility. Specifies how the ANN index is constructed and which similarity metric it uses.

Also known as: index_params, HNSW index configuration, Index type selection, Index parameters, Vector Index Build Config

Direct neighbourhood (hover for relationship types)

Variants

Variant	When to choose
Flat Index Configuration	Choose only for small collections (under about 100K vectors) where exact results are needed.
HNSW Index Configuration	Choose for large collections (beyond ~10 million vectors) where search latency outweighs higher memory usage.
IVF Index Configuration	Choose as a balanced default for mid-size collections needing good recall with manageable memory.

Relationships

configures structural

is evaluated by assurance

Retrieval Quality Evaluator Ch6.2A

Design guidance

SHOULD start from M=16, efConstruction=200 and a dot-product metric for normalized embeddings (cosine otherwise), then tune from measured recall and latency.
SHOULD keep M within 16-32; raise M to 24-32 if recall falls below 90%, lower to 12 if memory is constrained.
SHOULD invest in higher efConstruction (300-600) for batch-built indexes; continuous-ingestion pipelines MAY start at 200 to keep pace with the stream.
SHOULD choose the distance metric by embedding source and validate empirically with Recall@10/NDCG@10 on labeled queries.

Quantitative guidance

As stated by the sources; verify before use.

M=16 yields 95%+ Recall@10 at sub-100 ms; M=32 doubles memory, ~30% slower for +2-3% accuracy; M=64 adds <1% (Ch6.2A).
Index build ~10 min at efConstruction=200 vs ~30 min at 400 (Ch6.2A).
HNSW graph overhead ~200-300 bytes/vector at M=24 (Ch6.2A).
Brute-force 100K x 1024-dim: dot product 12.4 ms, cosine 18.7 ms, L2 23.1 ms; dot product ~30-50% faster than cosine (Ch6.2A).
Production tuning M=24 (maxConnections), efConstruction=300, ef=100, cosine: 95%+ recall with a few seconds of ingestion delay per thousand documents (Ch6.2B).

Classification

Patterns: HNSW M / maxConnectionsefConstructionCosine for models trained on cosine (e.g., NV-Embed)Dot product for pre-normalized or MRL embeddingsL2 for image/magnitude-bearing embeddings
Quality attributes: Functional suitability: correctness and validity (ISO/IEC 25010 | NIST AI RMF: valid)Performance efficiency (ISO/IEC 25010)

Sources

Ch6.2A: T. Nguyen, "Vector Database Selection," in Mastering Agentic AI Systems: Guide for the NVIDIA NCP-AAI Exam, 1st ed. 2026, ch. 6.2A. ISBN: 9798244538229.
Ch6.2B: T. Nguyen, "Production Vector Database Deployment," in Mastering Agentic AI Systems: Guide for the NVIDIA NCP-AAI Exam, 1st ed. 2026, ch. 6.2B. ISBN: 9798244538229.
Ch6.3B: T. Nguyen, "ETL Worked Example - Load Phase & Pipeline Integration," in Mastering Agentic AI Systems: Guide for the NVIDIA NCP-AAI Exam, 1st ed. 2026, ch. 6.3B. ISBN: 9798244538229.