Search and filters

import { retrieveDocuments } from "@anvia/core/vector-store";
import { vectorFilter } from '@anvia/core/vector-store';
const results = await retrieveDocuments({
    store: store,
    model: embeddings,
    query: 'reset a password',
    topK: 5,
    filter: vectorFilter.and(vectorFilter.eq('tenantId', 'acme'), vectorFilter.gt('revision', 3))
});

filterToPgVectorWhere produces parameterized SQL and values. Equality serializes the metadata value for JSONB comparison; gt and lt require numeric metadata and reject other types.

Cosine scores are 1 - distance. L2 and inner-product scores are -distance; for inner product this negates pgvector's negative-inner-product operator result. Treat the result as adapter ranking data and calibrate any thresholds per metric.

Add JSONB or expression indexes for frequent filters. Database filters still require independent authorization.