Thanks Kishore. That's a good way of handling complete re-indexes. With 2 vCPU I can indeed read another index just fine while generating embeddings on another. I'm still worried about the huge search latencies while inserting/updating even a few records with text embeddings. I experimented with setting

{batch_size: 2}

on the import. I expected that would make the import slower overall, but reduce the search latency because Typesense would service the read queue much more often. I don't get that result though - search latencies are still at least several seconds, with about half hitting the 5s timeout. Are you sure that

batch_size

param is actually working? It seems to have no effect on the latency or on the Pending Write Batches shown on the dashboard. I'd have expected that to spike much higher. I'm no C++ expert but I had a look through the source and can't see it doing anything useful. Thanks for persisting with me!