| name | 0.2.4-understand-idx_parameter_sweep |
| description | [0.2.4] Systematic exploration of parameter space to find optimal recall speed trade off for given constraints |
understand-idx_parameter_sweep
CALL NUMBER: deep_nearest_neighbor_sea.idx_parameter_sweep : vector_databases_and_embeddings(8)
DEFINITION: Systematic exploration of parameter space to find optimal recall speed trade off for given constraints
Invoke this skill to understand idx_parameter_sweep down to its primitives. The RELATIVE ROOT below is the least-fixed-point closure of everything it bundles from — the full import cone, grouped by the lib each prim comes from. Projected from a prover-typed KB (MAP/SWI-Prolog consistency gate): every reference below resolves.
THE RELATIVE ROOT (the import cone, by lib)
from deep_nearest_neighbor_sea
- idx_query_latency_budget (d1): Maximum acceptable query response time constraining algorithm selection and pruning aggressiveness
- idx_recall_target (d1): Desired fraction of true nearest neighbors returned governing aggressiveness of search pruning and parameter tuning
from vector_databases_and_embeddings
- approximate_nearest_neighbor (d2): An ANN algorithm that trades exact recall for sub-linear query time by probabilistically organizing the embedding space into clusters or graphs.
- faiss (d3): Meta's open-source C++/Python library for dense vector similarity search; implements IVF, PQ, HNSW, ONNG, and GPU-accelerated kernels with optional 8-bit quantization.
- hnsw_index (d3): Hierarchical Navigable Small World graph; layered skip-list + NN graph enabling logarithmic-layer traversal then constant-time local search.
- ivf_index (d3): Inverted File Index; clusters vectors into Voronoi cells via k-means; search prunes to nearest centroids then scans partitions.
- product_quantization (d4): A vector quantization technique that splits a high-dimensional vector into subvectors, independently quantizes each subvector, and concatenates codes to form a compact representation.
- scalar_quantization (d4): Floating-point values mapped to low-precision (e.g. int8); reduces memory 4x with modest accuracy loss.
- indexing_strategy (d4): Choice and configuration of ANN algorithm (HNSW, IVF, ANNOY, LSH) and its parameters (ef, nlist, nprobe) governing recall/speed trade-off.
- memory_footprint (d5): RAM or VRAM required to hold index and vectors; quantization (PQ, int8) is the primary compression lever.
CONSUMERS (what needs this)
idx_workload_profiling
Projected from the vector databases and embeddings KB (262 concepts / 150 relations) — consistency-typed by MAP; the facet list after the colon IS the cross-lib dependency web.
(leaf — this is an actual skill.)
