Atomic index creation, batched writes, and aligned contracts

VecStore SDK 0.1.2 is a patch release that makes the adapters behave more like one library.

createIndex is all-or-nothing

Each of the pgvector, Qdrant, and Cloudflare Vectorize adapters runs several provider calls in sequence, and a failure partway through used to leave the earlier calls standing. The caller got an error and a half-built index that reported already_exists on the retry.

The pgvector adapter now sends CREATE EXTENSION, CREATE TABLE, and both CREATE INDEX statements as one DO block, which the server runs in a single implicit transaction. The HNSW index caps a vector column at 2000 dimensions, so a 3072-dimension index used to leave a table with no vector index and every later query fell back to an exact scan. The Qdrant adapter deletes the collection it just created when the _namespace tenant index fails, and the Vectorize adapter deletes the index it just created when a metadata index fails.

Batched Qdrant and Redis writes

Qdrant used to put the whole list in one request, which failed past the 32 MB the REST API accepts by default. upsert now goes out 500 points at a time, and fetch and delete({ ids }) chunk their point ids at 1000 per request. Redis writes now run 500 records at a time, and unlink keys are chunked the same way.

A large upsert is now several provider requests, so a failure can leave some of the records written. Retry the whole call: an upsert replaces by id.

The pgvector metric is cached per store

The pgvector adapter read the distance metric from pg_indexes once per index handle, so store.index("docs").query(...) paid a catalog round trip before every search. The cache now lives on the store, keyed by schema and table, and holds the in-flight promise, so concurrent first queries share one lookup.

Six contract edges aligned

• Qdrant deleteIndex returns not_found for a collection the server does not hold, like the other six adapters.
• createIndex rejects a dimension that is not a positive integer on every adapter with the same invalid_argument message.
• Pinecone query honors includeMetadata and includeVector, so a ScoredRecord from Pinecone has the same shape as one from the other adapters.

The full list is in the package changelog.