Building the next generation of job search at LinkedIn

Caleb Johnson29 May 2025

Problem

Keyword job search forces people to guess the exact title in a posting, so a query like "mostly remote, not from sourcing companies, above median pay" cannot be expressed. LinkedIn wanted natural-language queries, which need semantic understanding matched against millions of postings tens of thousands of times per second.

The old pipeline had grown to nine independently optimized stages, duplicated across a dozen channels. Nobody could tell where a relevance problem came from.

Example from the article: a member asks for "jobs in New York Metro Area where I have a connection already", and the system must resolve the region to a geo ID and the connections to company IDs.

Solution in brief

Train one large teacher model that grades a query-job pair, then distill it into a fast retrieval model and a smaller cross-encoder ranker, so both agree with the same teacher. An LLM query engine turns each query into strict filters plus a soft embedding. An LLM tuned on human grades supplies labels at scale.

Questions this case answers
  • How to replace keyword matching with natural-language semantic search over a fast-changing catalog of short-lived items at high query volume
  • How to keep a cheap retrieval model and a ranking model consistent with one expensive teacher model without maintaining a chain of independently tuned stages
  • How to produce relevance training labels at the scale of millions per day when human grading is the bottleneck
  • How to let users search for jobs in natural language when keyword search cannot express their intent?

Tools and methods

  • Teacher model with multi-stage distillation — defines relevance for both students
  • LLM grader trained on human annotations — applies a 5-point grading policy at scale
  • Query engine (fine-tuned LLM, tool calling) — classifies intent, tags entities for filters
  • Exhaustive GPU nearest-neighbor retrieval — dense matrix multiplication over job embeddings
  • Cross-encoder small language model — ranks query-job pairs after logit distillation
Continue reading — it's free

Google or a one-time email code. No password.