Support with long runtime collection query when filtering metadata

[roy-armis]

Hey all,

First time here, so I apologise if this isn't the place for support.

My data contains ~270K documents, embedded with OpenAI's Embeddings, and 3 metadatas: is_valid, is_purchased and num_words. The first two are ints that act as booleans - meaning 1 is True 0 is False. The latter is int.

When quering the collection only passing the query document, the query runtime is ~2sec, but when adding metadata filtering the query takes much longer:

• ~8s for including all 3 metadatas
• ~17s for including 2 of the metadata fields
• ~35s for including one of the metadata fields

I would really appreciate any assistance on how to improve performace here, as Ideally the requirement for my app is for this process to be no longer than 3-4 seconds.

Thanks you all for building a great product!

[HammadB]

The best solution here is a WIP #1125

[itaismith]

Hi @roy-armis! This issue was opened some time ago, and Chroma has made some major improvements since, including adding a metadata index that speeds up query time significantly. I just tested locally a similar collection to the one you described and query time was only a second or so. Feel free to reach out or re-open the issue if you're still having performance issues!