Array-valued metadata filters, bounded batching, and Supabase hardening

VecStore SDK 0.1.3 fixes filter semantics on list fields, bounds how fast a large upsert fans out, and closes the last paths where a write could reach another namespace’s record. Supabase users should re-run sql/supabase.sql.

eq and isIn match list fields on pgvector and Supabase

eq("tags", "x") and isIn("tags", ["x"]) compiled to a jsonb containment that never matches a list, so a record with tags: ["x", "y"] was invisible to both. Both compilers, the TypeScript one and vecstore_filter_sql, now emit a containment per value that the GIN index serves, and ne and notIn are the exact negations. Qdrant notIn now compiles to must_not of match.any, the same filter not(isIn(...)) produces.

Batched calls run four at a time

A 100k-record upsert used to open every chunk in one burst, which met rate limits on Pinecone, Upstash, and Supabase and took every connection of a pg Pool. Batches now go out four at a time. An upsert that repeats an id keeps the last record on every adapter. Pinecone delete({ ids }) is chunked at 1,000, Redis fetch splits JSON.MGET at 500, and every adapter’s query returns invalid_argument for a topK that is not a positive integer.

Namespace isolation on every write

On Upstash in metadata mode, a default-namespace upsert whose id has the {namespace}/{length}/{id} shape of a stored namespaced record now returns invalid_argument. On Qdrant and Vectorize, a raw id that is a version 8 UUID, the version the adapter’s hashed ids use, is now hashed like every other id, so no raw id can equal another record’s stored id.

Supabase enables row level security

vecstore_create_index now enables row level security on the table it creates, so the anon and authenticated roles read and write nothing until you add a policy. The script also revokes vecstore_create_index and vecstore_drop_index from those roles, and vecstore_filter_sql rejects a filter leaf with a missing or non-numeric value instead of dropping the clause, which could widen a delete. Tables created by earlier versions are unchanged: run alter table "docs" enable row level security on them.

Also in this release

• The cached pgvector metric is forgotten when createIndex or deleteIndex touches a table.
• Postgres index names over 49 bytes are rejected on pgvector and Supabase, because the _embedding_idx suffix made long names collide with their own indexes.
• The npm tarball ships the MIT license, and @pinecone-database/pinecone 7 or later is required.