Hey everyone, I’ve been hitting a wall with something in Arize and I’m hoping someone here has navigated this before. We’re trying to use the dimension reduction visualizations (UMAP/t-SNE) for our embedding vectors, but the charts are either completely broken—just a blob or a few strange clusters—or the process times out entirely. It’s becoming a real blocker for our team's ability to visually validate model drift in our semantic search system.
For context, we’re coming from the email/marketing automation world (my usual haunt!), where data is more tabular. These high-dimensional embedding spaces are a new frontier for us. Here’s what we’ve checked so far:
* **Embedding Dimensions:** Our vectors are 768-dimensions, which I understand is standard but perhaps heavy.
* **Sample Size:** We’ve tried sending down samples from 5k to 50k observations. The smaller samples sometimes render but look nonsensical; larger ones just fail.
* **Normalization:** We’ve ensured our embeddings are L2 normalized before logging, as per the documentation.
* **Payload & Logging:** We’re using the Python SDK, logging the `feature` as a `Vector` type, and there are no errors on the logging side.
My gut tells me it’s either a dimensionality issue (maybe we need to pre-reduce?) or something in the Arize configuration we’re missing. I love the *idea* of these visualizations—they feel like a hidden gem for understanding clusters—but we can’t get them to work.
**My questions for the community:**
1. Has anyone successfully visualized 768-dimension embeddings in Arize? Was there a specific trick?
2. Does Arize perform better with a pre-computed reduction (like PCA to 50 dimensions) before the built-in UMAP?
3. Are there any settings on the board itself (like perplexity for t-SNE) that we’re overlooking that could stabilize this?
I’d be so grateful for any workflow reports or pitfalls you’ve experienced. Even knowing “this just doesn’t work above X dimensions” would save us a lot of time!
—Aurora
don't spam bro