Hey Isaac, good question. Most text embedding models are typically designed for representing semantic meaning of longer text passages like sentences, paragraphs, or pages of text. With just that information I'd expect embeddings of those phrases to be very similar, but larger passages with "lawnmower" and "lawnmower charger" as subjects to have more distinct embeddings. here's an experiment i ran to demonstrate this (

pc.inference.embed(...)

): - Embeddings for the two phrases you mentioned from our model offering 'multilingual-e5-large' have a cosine similarity score of 0.9815. - Embeddings for the sentences "I mowed the lawn with my nearly brand new electric lawnmower and covered half an acre of land." and *"I charged by nearly brand new electric lawnmower with the new charger I purchased from Home Depot. It filled the battery to 100% in 1 hour and has high voltage"*(same model) have cosine similarity 0.855. Even a small bit of extra context allows the embedding model to better understanding their meaning resulting more proper sim. scores, and that trend should continue with larger and richer context in text chunks. For distinguishing between those two specific entities, techniques such as using sparse vectors are supported for lexical search (https://www.pinecone.io/learn/learn-pinecone-sparse/)