Skip to content

Standardize Golden Query Set Generation & Recall Calculation #5

@brian-ogrady

Description

@brian-ogrady

Golden Query Generation

Given an embedding dataset on HuggingFace, we need to be able to evaluate the quality of the ANN results against true KNN. This requires generating representative queries (not random). We can do this in a few ways:

Techniques

  1. Manually craft representative queries (not scalable)
  2. Use an LLM to generate queries based on the dataset description (requires scripting)
  3. Leave out a set of embeddings from the given dataset as a holdout "validation" set (probably shittiest but easiest)
    • Could just delete points randomly from collection, but this has limitations

Recall Calculation

Given set of golden queries, execute the brute force queries. This will require GPUs.

Given a framework for calculating brute force queries using GPUs.

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions