Latent Space Explorer: Navigating Paper Embeddings in 3D
REPORT_ABSTRACT
Latent Space Explorer embeds arXiv abstracts offline with a MiniLM sentence transformer, projects the 384-dimensional vectors down to 3D through truncated SVD and UMAP, and renders the result as a navigable point cloud in a single WebGL draw call. All expensive work happens once in a reproducible three-stage pipeline, so what ships is a purely static site: under a megabyte of binary buffers. Free-text queries are embedded in the browser with transformers.js and ranked against int8-quantized vectors inside a Web Worker.
01. CORPUS EMBEDDING STAGE
Every abstract is embedded offline with a MiniLM sentence transformer (all-MiniLM-L6-v2), producing 384-dimensional, L2-normalized vectors. Embedding at build time rather than at runtime is what keeps the shipped site static — the visitor never needs a GPU, a network fetch, or a vector database to explore the corpus.
02. DIMENSIONALITY REDUCTION
The 384-D space is far too high to render directly. A truncated SVD is computed once to capture the dominant linear structure, then UMAP embeds into 3D for spatial layout. A critical design detail: the same SVD basis is dual-use — it also produces 128-dimensional search codes, so runtime queries land in exactly the same space as the offline corpus.
03. SINGLE-DRAW RENDERING
The entire cloud renders as one three.js draw call. A custom GLSL shader handles per-point colour, filtering, and selection state, which means frame cost tracks rasterized pixels rather than point count — panning and orbiting stay smooth even with hundreds of thousands of points. Binary buffers keep the geometry under a megabyte.
04. IN-BROWSER SEMANTIC SEARCH
Free-text queries are embedded on the client with transformers.js using the same MiniLM model used offline, then a Web Worker brute-force scans the int8-quantized codes and returns nearest neighbours in milliseconds. Quantization to int8 cuts memory and bandwidth fourfold while preserving ranking quality — a direct trade, with no external API in the loop.
05. INTERACTION LAYER
The UI shifts between desktop and mobile: hover-to-inspect tooltips on desktop become tap-and-sheet on mobile. A legend allows per-archive filtering of the cloud, and selecting a point reveals its paper metadata plus nearest-neighbour links, closing the loop between spatial exploration and retrieval.