Problem
Job Search at LinkedIn picked its first-stage candidates by matching terms. That gave results, but little personalization, no semantic matching and no way to weigh a title match differently from a description match. Teams that wanted embeddings also had to build model training, index building and serving each time, and a retrained embedding model could silently stop matching the item index. The goal was more relevant jobs in front of job seekers.
Example from the article: a recruiter searching "senior burn ward nurse" has a request embedding that ends up geometrically close to several nursing profiles with experience in that area.
Solution in brief
LinkedIn trains a two-tower model on impression logs, indexes job embeddings in an IVFPQ index, and embeds each search request at query time to fetch the nearest jobs. A shared platform handles embedding generation, index versioning and inference graphs. In Job Search it replaced much of the term-based candidate selection.
Questions this case answers
- How to move first-stage candidate selection from keyword matching to vector similarity while keeping latency in check
- How to ship new versions of an embedding model when the query and item embeddings must stay aligned
- How to replace term-based candidate selection in job search with embedding-based retrieval?
- How to train a two-tower retrieval model from impression and engagement logs?
Tools and methods
- Two-tower model — encodes the request and the job into one embedding space, trained with softmax loss.
- In-batch negative sampling — builds training negatives from the impression data.
- IVFPQ index (Zelda) — approximate nearest neighbor search over job embeddings.
- Multi-task, composite models — one model co-trains several embeddings and objectives.
- Ray Serve (Model Cloud) — runs inference graphs without a per-team middle tier.
Google or a one-time email code. No password.