Paste in a job description, upload a stack of resumes, get them back in order. The interesting question is what “in order” should mean.
How it scores
Keyword matching answers that question badly. A candidate who writes “built streaming pipelines on GCP” scores nothing against a posting that asks for “Dataflow experience”, even though they are describing the same work. Term overlap measures vocabulary, and vocabulary is the thing a good candidate is least likely to share with whoever wrote the posting.
So the engine compares meaning. Both sides get embedded and scored on similarity, which lets the phrasing differ as long as the substance lines up. Against a keyword baseline on my own labelled set, that was worth roughly a quarter more accuracy.
The retrieval path assumes the candidate pool grows. Scores are computed and stored so that adding resumes does not mean recomputing the batch every time someone opens a posting.
What it does not do
It produces an ordering, not a decision. Every result still goes to a person, and the accuracy number above is measured against labels I made myself, not a published benchmark, so treat it as a direction rather than a figure to quote.