Offline. In Expert Judge replays, MUSE-powered semantic retrieval delivers the best of both worlds as a single sourcing strategy — replacing an earlier engagement-optimized system that fused a sister organization's embeddings in retrieval and vanilla open-weight LLM embeddings in the L2 ranker as strong baselines. Its liquidity — ~76% average utilization of the retrieval pool at 96% query coverage — is on par with the prior high-liquidity faceted strategies and 2.1x higher than the prior high-quality Boolean strategy. Its quality matches the high-quality strategy too, with pre-L2 HRR ~1% higher in absolute value (and 8% higher than the engagement-optimized EBR). After the MUSE-featured L2 Ranker, the full stack drives pre-LLM Guard HRR up +4% and Facepalm rate down −5%, while semantic retrieval's candidate share rises from 18% to 31% — confirming improved retrieval-ranker consistency.

Online. The online pattern confirms the design intent: the system surfaces a tighter, higher-quality candidate pool rather than optimizing for volume. A two-week A/B test validates this in production (here relative gains): post-LLM Guard HRR +2.7%, InMail Sends per seat +4.1%, Eval Pass Rate trending +3.8%, with candidates sourced per seat down ~4% — fewer but better. InMail Accepts per seat trends +1.8%. These gains represent a step-function improvement in the quality of semantic search results for Hiring Assistant. 

Surprising Findings & Lessons Learned

Building semantic search surfaced several insights we didn't anticipate — some about modeling, some about infrastructure, and some that changed how we think about the problem.

High-confidence LLMs outperform human annotators on knowledge-intensive labeling

When we analyzed disagreements between the early version of MUSE Teacher and human annotators, high-confidence model predictions were correct roughly four out of five times — many apparent "errors" turned out to be human mistakes, especially on technical qualifications where the model had comprehended the entire context. Humans consistently won on common-sense inference (e.g., "MS Word experience" implied by the role) and arithmetic (years-of-experience calculations). Treat human labels as a noisy signal, not a ceiling — and invest in understanding where each source is more reliable.

Contrastive post-training alignment beats model size

For embedding tasks, alignment between the base model's post-training objective and the downstream task matters more than raw parameter count. Models post-trained with contrastive objectives on large-scale text pairs — where the training signal is already "place semantically related texts close together" — adapt to retrieval far more effectively than models that are stronger at generation but lack that alignment. Tokenization plays a role too: models with tokenizers optimized for long, structured texts encode our inputs (resumes, qualification lists) more faithfully.

Approximate retrieval and post-filtering compound each other's losses

ANN search trades recall for speed, and post-filters then discard more candidates after retrieval — the two losses multiply. ABM + EBR spends limited retrieval slots on candidates that get filtered out, while strong matches that an exact search would find are never scored. According to our estimates, exhaustive kNN with pre-filters over the same query understanding would yield ~30% more candidates at the same relevance — a gap that motivates our next infrastructure investment, as a part of broader effort to reshape LinkedIn’s search stack. 

Conclusion

Semantic search transforms recruiter sourcing from keyword matching into genuine understanding of qualification fit — finding candidates who actually match the role, not just those whose profiles contain the right words. By combining LLM-based relevance supervision with embedding models that serve both retrieval and ranking at billion scale, we've resolved the long-standing trade-off between matching quality and coverage. For recruiters, this means spending more time engaging with strong matches and less time on candidates who aren't the right fit. For candidates, it means being surfaced for roles they're genuinely qualified for. As Hiring Assistant continues to evolve, the platform is designed to improve alongside it — as models, supervision, and the definition of "qualified" all grow together.

Acknowledgements

The success of this work stems from extensive collaboration, with technical, product, and leadership support across LinkedIn's Talent Enterprise and partner teams in LLM serving, ML training, search infrastructure, and MLOps. We express our gratitude to:

Prashanthi Padmanabhan, Manav Sharma, Greg Pounds, Daniel Hewlett, Xiaoyang Gu, Peter Rigano, Adi Pruthi, Adriana Meza, Hossein Hamooni, Tony Lee, Xiao Shi, Lucky Wang, Haoxin Li, Ayoub Amil, Eddie Cheung, Carter Chung, Pramod Para, Gurvineet Dhillon, Juan Grosso,  Jady Rodriguez, Victor Hu, Rui Zhang, Shitao Wang, Jaya Bansal, Ketan Thakkar, Manish Baldua, Abhinav Gosavi, Komal Arvind Dhuri.