Anton Troynikov

> It would appear that the issue exists on hnswlib 0.7.3 too (Windows 10, AMD Ryzen 5) - https://discord.com/channels/1073293645303795742/1265778818422145149 @tazarov can you please post a summary of this long conversation...

This is also likely related to ef_search being too low by default, as per https://github.com/chroma-core/chroma/issues/1737 @MissJingRongLi could you try creating a new collection and setting the hnsw:search_ef collection metadata key...

* **#2611** * **#2601** * **#2495** 👈 * `main` This stack of pull requests is managed by Graphite. Learn more about stacking. Join @atroyn and the rest of your teammates...

* **#2610** 👈 * `main` This stack of pull requests is managed by Graphite. Learn more about stacking. Join @atroyn and the rest of your teammates on Graphite

Hi @LumenYoung, thanks for pointing this out. Our initial implementation of multi-modal embeddings does indeed assume a single datatype at a time. While the proposed change might work for the...

I think we would want to carefully consider any approach here since it would touch a lot of the API surface. Could you expand on your proposal with some example...

This is probably happening because of the way we do dynamic imports. I agree it's a bad user experience. @tazarov if you can take some time to see how this...

I think we can eat the increased DB size for faster filtering. We also wouldn't expect indices to help where the returned set is large, but we should probably think...

could we call `upsert` instead so that the client doesn't need to know about any of its db internals? Maybe we want to change `upsert` to return some indication of...

No I mean underneath `batch_add` we call `upsert` ourselves. This will be idempotent for any existing IDs, so a restart won't cause duplicates.