Closed bobtoalice closed 1 month ago
@bobtoalice Thank you for bringing this issue to our attention. We appreciate you taking the time to identify this limitation in the GPU Index project.
You're right that the current GPU-Index only supports IP
and L2
metrics, but the Cohere
/ OpenAI
datasets (provided by VectorDBBench) rely on Cosine
similarity. Performing an additional normalization
step to handle this discrepancy is a sensible approach.
I will try to fix this issue as soon as possible, and let you know once it's done.
@bobtoalice The latest code has resolved it. We will publish the updated version to PyPI soon. In the meantime, you can test the changes using the following steps.
git clone https://github.com/zilliztech/VectorDBBench.git
cd VectorDBBench
pip install -e ".[test]"
init_bench
Hello,
I’m attempting to run load tests for a GPU-enabled Milvus cluster and I’ve been using IP as a metric. However, the tests always result in an error that tells me L2/IP are either nonexistent or not supported. The error message indicates that Cosine is being used. Would it be possible to specify the metric type from the UI?
This is the error in full:
Thanks!