---

name: index-selection-and-search description: "Guides dense-vector metric selection, index construction, approximate search tuning, and CPU search validation in Faiss." disable-model-invocation: true metadata: disco-role: operating license: MIT

Index selection and search

Use this route when the task is to choose or operate a CPU dense-vector index, compare exact and approximate nearest-neighbor search, or diagnose distance semantics, result contracts, search tuning, or OpenMP behavior. It covers float32 dense vectors, L2, inner product, cosine-as-normalized-inner-product, Flat, IVF-Flat, and graph indexes.

Route the workflow

1. Normalize the data to a contiguous float32 matrix of shape (n, d) and keep database and query dimensions identical. Read the API contract before calling an index.
2. Choose the smallest index that meets the latency, recall, and memory target. Use the selection and tuning workflows for the decision, including the 64D memory-conscious factory case.
3. Build with the lifecycle in the workflow reference: create, train when is_trained is false, add, then search. Validate against an exact Flat baseline before increasing approximation.
4. Use the bundled deterministic helper for a local CPU smoke check: python path/to/skills/disco/faiss/sub-skills/index-selection-and-search/scripts/smoke_search.py --help. It has no downloads and is safe to run from any working directory.
5. For malformed arrays, unavailable methods, empty result ranges, or poor recall, follow troubleshooting.

Boundaries

• Route codecs, binary indexes, PQ/SQ, and compressed-memory choices to training-and-compression.
• Route IDs, selectors, composition, refinement, shards, and replicas to composition-and-filtering.
• Route persistence, I/O, clustering, and evaluation tooling to persistence-and-evaluation.
• Route GPU, cuVS, ROCm, Metal, SVS, and C/C++ interop to accelerated-and-interoperable.
• Use the Faiss root route for package-wide installation, provenance, and cross-workflow routing.