Spatial Visualization of Embeddings

This weekend I was pointed toward a github repo for a thing called Project Golem.

Image Card

Original Design

Local Image

Raw image under static directory.

Local Image

Image under assets directory, processed by Hugo.

This looked really interesting because it was visualizing something I’ve been working on understanding. Lately, I’ve been looking into different retrieval methods, reranking, data normalization, chunking, and embedding. I’m a visual learner so this was immediately interesting to me.

The interface allows you to search a pre-built corpus of information, which is a handful of wikipedia entries sorted into broad categories. The categories are assigned colors. Then the documents are chunked, embedded, then their embeddings are converted to be visible in a 3D space.