Migration guide

Move an existing vector workload in three steps.

1. Create an equivalent index

Match your embedding dimensions and select cosine for normalized embeddings or euclidean when vector magnitude matters.

2. Import in idempotent batches

Write up to 500 vectors per request. Preserve the source document identifier as the vector id; it makes re-imports and deletes deterministic.

3. Validate recall before switching traffic

Run representative production queries against both systems. Compare retrieved IDs, apply metadata filters, then move read traffic. The console shows rows written and query transfer so you can observe the cutover.