Search and filters

import { retrieveDocuments, vectorFilter } from "@anvia/core/vector-store";
import { RedisVectorClient } from '@anvia/redis';
const store = new RedisVectorClient({}).vectorStore({
    indexName: 'support_docs',
    dimensions: 1536,
    metadataSchema: { tenantId: 'tag', revision: 'numeric' }
});
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))
});

filterToRedisQuery translates string equality to RediSearch TAG syntax @key:{escaped}; stored tag values carry typed prefixes (s: strings, d: numbers, b:1/b:0 booleans, n: null), so eq('tenantId', 'acme') becomes @tenantId:{s\:acme}. Numeric equality becomes an exact range and gt/lt become open numeric ranges. Compounds translate to RediSearch intersection (and) or union (or) syntax.

Filter fields must be declared in the store's metadataSchema. Automatic index creation defines only the reserved __anvia_document_id TAG field plus the schema-declared fields, and filtering an undeclared key throws a TypeError naming vectorStore({ metadataSchema }). Tag fields support equality; numeric fields support equality and ranges, and numeric filters require finite numbers.

Search uses KNN over the configured vector field. Filters narrow candidates but do not authorize documents.