How does metadata filtering work in Pinecone?

Pinecone stores arbitrary JSON metadata with each vector and supports filtering queries by metadata conditions using operators like equals, in, greater than, and less than. Filters are applied during the vector search so the database does not return items that fail the filter, known as pre-filtering. Effective filtering requires metadata fields to be indexed, which Pinecone handles automatically in serverless indexes. Highly selective filters can hurt recall if the ANN index cannot find enough matching candidates, so cardinality matters in filter design.