db-specialized-fit (M20) — vector · time-series · graph · search

M20 covers purpose-built engines whose correctness depends on parameters a generic relational audit ignores. Sub-modules carry a letter (M20a..M20d); the scorer maps them to the parent M20 in the paradigm profile (Vector / Time-series / Graph categories in references/scoring-model.md).

M20a — Vector (pgvector, Qdrant, Pinecone, Weaviate)

Feeds: design (Métrica & dimensión, Modelo-version, Metadata/filtro) + performance (Índice & params, Búsqueda filtrada, Recall-vs-latencia).

• Dimension match — column/collection dim equals the embedding model's output dim (e.g. vector(1536) for text-embedding-3-small). A mismatch is a hard bug (design, sev 5).
• Distance metric match — the index metric (cosine / L2 / inner-product) matches how the model was trained; a mismatch silently wrecks recall. Sev-5 only when the model is declared in-repo; else directional and do not cap (per scoring-model honesty rule).
• Index present & tuned — HNSW (m, ef_construction, ef_search) or IVFFlat (lists/probes) declared, not a brute-force seq scan on a large table (performance).
• Model version captured — embeddings are tied to a model version so a re-embed is possible (design).
• Filtered search — metadata used for pre/post-filtering is itself indexed (performance).