Building a resume-to-job matching engine for an HR platform and comparing NeMo-based embeddings against a couple of general-purpose embedding models for semantic matching quality versus cost at scale (processing large applicant pools regularly, not one-off queries).
So far the general-purpose models edge out slightly on raw semantic similarity for well-written resumes, but degrade more on inconsistent formatting (which is most real resumes) — where a NeMo model fine-tuned on domain-specific text has held up better. Cost per embedding at our current volume is comparable either way, so it’s mostly coming down to matching quality on messy real-world input.
For anyone who’s built matching or recommendation systems on messy, semi-structured text specifically (not just clean product/document embeddings): did fine-tuning a domain-specific embedding model end up being worth the setup cost versus just doing more aggressive text normalization upstream of a general-purpose model? Also interested in how people are validating for matching fairness/bias at scale, since that’s turning out to be a harder problem than the embedding choice itself.
Working on this as part of HRMS software at Dev Technosys — this is the newest of our efforts, so still gathering real data rather than firm conclusions.