Embrix: A Node.js Framework for Embedding Vector Generation and Similarity Evaluation (System Description and Initial Benchmarking Results)

Abstract

I present Embrix, a lightweight Node.js framework I built for testing embedding inference: it generates sentence embedding vectors, measures latency/throughput, and checks semantic similarity using cosine distance. In a CPU-only Windows x64 environment (Node.js v25.5.0), I benchmark Embrix's current pipeline using two embedding models, MiniLM and BGE. For MiniLM, cold-start latency (first inference including model load) averages 30.18ms across five trials, with one outlier run at 140.10ms; warm-start inference over 50 iterations averages 2.68ms per embedding. Batching improves throughput and reaches 652 embeddings/s at batch size 100 in an end-to-end test; a dedicated batch-scaling sweep peaks at 696.9 embeddings/s at batch size 20. To validate similarity behavior, I analyze cosine similarity for paraphrases and unrelated sentence pairs. MiniLM achieves mean similarity 0.902 for same-meaning pairs versus 0.0548 for unrelated pairs, yielding a separation margin of 0.217. In a small model comparison, both MiniLM and BGE obtain clustering purity 1.0 on a two-cluster toy dataset; MiniLM is faster (mean 2.56ms) than BGE (mean 4.28ms) and shows a slightly larger separation margin (0.204 vs. 0.182). Overall, this paper documents Embrix's current capabilities and provides baseline measurements to guide future development.