Separating Semantic Expansion from Linear Geometry for PubMed-Scale Vector Search

TL;DR

The paper tackles PubMed-scale retrieval by separating semantic interpretation from geometric similarity. It uses a deterministic LLM-based expansion to transform queries into concise biomedical phrases, then operates in a fixed, mean-free embedding space produced by a Johnson–Lindenstrauss projection, enabling exact cosine search over a large index without trained encoders. Geometry-focused evaluations show strong alignment with semantic intent, with head cosine around 0.68 and centroid closure around 0.81, on a corpus of about 40 million records and an index size around 12.8 GiB. The work demonstrates that, once semantic expansion is reliable, a fixed linear embedding can achieve coherent, reproducible retrieval at PubMed scale without training, offering favorable cost–quality characteristics for large-scale biomedical search.

Abstract

We describe a PubMed scale retrieval framework that separates semantic interpretation from metric geometry. A large language model expands a natural language query into concise biomedical phrases; retrieval then operates in a fixed, mean free, approximately isotropic embedding space. Each document and query vector is formed as a weighted mean of token embeddings, projected onto the complement of nuisance axes and compressed by a Johnson Lindenstrauss transform. No parameters are trained. The system retrieves coherent biomedical clusters across the full MEDLINE corpus (about 40 million records) using exact cosine search on 256 dimensional int8 vectors. Evaluation is purely geometric: head cosine, compactness, centroid closure, and isotropy are compared with random vector baselines. Recall is not defined, since the language-model expansion specifies the effective target set.