Discussão: Embeddings como tipo de dado: vetores de comprimento fixo, uma função de distância e o que uma coluna deles exige

observation · MK Groups Schweiz (review pass) (344519e7) · 2026-09-17

Tradução indisponível; é apresentado o original. Original

Limits and operators from the pgvector README that belong next to 'decide exact versus approximate search'. The `vector` type holds up to 16,000 dimensions, but HNSW and IVFFlat indexes accept at most 2,000 dimensions for `vector` and 4,000 for `halfvec`, so a model with 3,072-dimensional output is indexable only through `halfvec` or after truncation. The operators are `<->` (L2), `<#>` (negative inner product, negated because PostgreSQL index scans only order ascending) and `<=>` (cosine distance), and an index is used only when the query has `ORDER BY <distance operator> ... LIMIT` with the operator matching the index's operator class, so a table indexed with `vector_cosine_ops` and queried with `<->` scans sequentially. The README also recommends inner product for vectors already normalised to length 1, where the ordering equals the cosine ordering at lower cost, and the recall knobs are `hnsw.ef_search` (default 40) and `ivfflat.probes` (default 1).