[Feature]: Vector similarity search by multiple id's

[cf-lam]

Is there an existing issue for this?

• [X] I have searched the existing issues

Is your feature request related to a problem? Please describe.

I'd like to perform vector similarity search by multiple id's. This comes up a lot in my application because I'm storing id's and operating primarily off of id's in my application code.

The current workaround involves 2 steps instead of 1, is a poor developer experience, and also creates unnecessary network and memory load. I have to first get the embeddings for my ids, and then perform similarity search on the embeddings. My list of ids has length 20. I'm using OpenAI's large text embedding model, so each embedding is 11 KB. With the workaround, the first network request returns 20 11 KB = 220 KB over network into memory, before sending 220 KB again over the network to get the final results. So now performing a vector similarity search sends 2 220 KB = 440 KB over the network. I'm performing these searches pretty frequently.

Similarity search by id exists in other solutions like Pinecone.

Describe the solution you'd like.

A new API method in Python that performs vector similarity search given a list of id's. If I have one method that accepts a list of ids, 440 KB data sent over network can be drastically cut down by 1-2 orders of magnitude.

Describe an alternate solution.

No response

Anything else? (Additional Context)

No response

[xiaofan-luan]

Agree this is an valid request.

1. milvus do a query before search
2. milvus do search and get topK for each query.
3. milvus can also cache search result by id

[congguosn]

We have encountered the same requirement in various scenarios and strongly advocate for the ‘search by ID’ feature.

[xiaofan-luan]

This will be in 2.5 release