Possibility to post-process / transform search results with multivector query

[dr-duplo]

Is it possible to run a search with multiple query vectors and only return the minimum distance across the query vectors per dataset vector?

Use case: For a audio snippet, I have a embedding vector for each 100ms frame and a dataset of vectors that can fit to any of those frames. For 1s audio snippet I would want to set the search data to a (10, d) matrix for a (N, d) dataset with N vectors (d being the dimensionality). The result should give me the minimum distance to any of the frames per dataset vector. Basically return a list of N minimum distances (or shorter if limited).

Is that possible with Milvus alone?

[xiaofan-luan]

@liliu-z

this is what are actively working on and should be delivered in the next 2 months.

[xiaofan-luan]

Is it possible to run a search with multiple query vectors and only return the minimum distance across the query vectors per dataset vector?

Use case: For a audio snippet, I have a embedding vector for each 100ms frame and a dataset of vectors that can fit to any of those frames. For 1s audio snippet I would want to set the search data to a (10, d) matrix for a (N, d) dataset with N vectors (d being the dimensionality). The result should give me the minimum distance to any of the frames per dataset vector. Basically return a list of N minimum distances (or shorter if limited).

Is that possible with Milvus alone?

basically what we do:

1. support a data model with a list of embeddings
2. support a two stage search, first stage is an ANN request to find top X candidates for each vector in your query data set, ideally you got N * X candidates
3. rerank the N * X candidates with the N Query vectors (if ncessary)