What is the recommended way to batch-insert millions of embedding rows efficiently?
- A Individual INSERT statements in autocommit
- B COPY (or the binary COPY protocol) with the vector index created afterwards (correct answer)
- C INSERT ... ON CONFLICT per row
- D Using a stored procedure loop
Shortcut: COPY is dramatically faster than row-by-row inserts, and deferring index creation until after the load avoids the cost of incrementally maintaining the graph during ingestion.