Recruiting · Production

LinkedIn builds AI-powered semantic job search using LLMs, embedding-based retrieval, and GPU infrastructure

The problem

Traditional keyword-based job search failed to capture nuanced user intent and required exact keyword matches, while LinkedIn's legacy AI pipeline had grown into a multi-stage system duplicated across many channels that made relevance improvements difficult.

First attempt

LinkedIn's legacy candidate selection and ranking models used fixed taxonomy-based methods and older LLM technology that lacked capacity for deep semantic understanding, and the multi-stage pipeline had become too complex and duplicated to permit reliable relevance improvements.

Workflow diagram · grounded in source
1
Natural language query submitted
trigger
“This new search experience lets members describe the job they want in their own words and receive results that include jobs they might not have considered and that are more closely aligned with their ideal job”
2
Query engine processes intent
ai_action
“The query engine constructs the appropriate retrieval strategy by classifying the user intent, fetching external data such as profile and preferences needed for an effective search, and performing natural entity recognition to tag strict…”
3
Tool Calling on fine-tuned LLM
ai_action
“we leverage the Tool Calling pattern when querying our fine-tuned LLM”
4
Personalized facet suggestions via RAG
ai_action
“stored in a vector database and passed to the query engine LLM via the RAG pattern”
5
GPU-accelerated embedding retrieval
ai_action
“serve the K closest job postings for a given search query in just a few milliseconds per query”
6
Distilled SLM ranking
ai_action
“we have generated a "small" language model (or SLM) that learns from the Teacher while still having close performance”
7
LLM annotation feedback loop
feedback_loop
“we built out an LLM fine-tuned on human annotations to apply learned product policies and grade arbitrary query-member-job records. This approach allows for automated and scalable data annotation (to the tune of millions or tens of milli…”
Reported outcome

LinkedIn delivered a semantic job search system that understands natural language queries, reduces pipeline complexity by an order of magnitude, processes annotation grades at scales well beyond human capacity, and serves retrieval results in a few milliseconds per query, while significantly enhancing developer velocity.

Reported metrics
Model pipeline complexity reductionreducing the model pipeline complexity by an order of magnitude
Automated annotation throughputmillions or tens of millions of grades per day
Retrieval latencya few milliseconds per query
Developer velocitysignificantly enhanced developer velocity
Show all 7 reported metrics
model pipeline complexity reductionreducing the model pipeline complexity by an order of magnitude
automated annotation throughputmillions or tens of millions of grades per day
retrieval latencya few milliseconds per query
developer velocitysignificantly enhanced developer velocity
retrieval performancesignificantly better retrieval performance
legacy pipeline stage countnine
legacy pipeline channel duplicationduplicated over a dozen different job search and recommendation channels
Reported stack
LLMembedding-based retrievalGPUvector databaseRAGTool CallingFSDPDeep and Cross Network architecture
Source
https://www.linkedin.com/blog/engineering/ai/building-the-next-generation-of-job-search-at-linkedin?utm_source=chatgpt.com
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

LinkedIn delivered a semantic job search system that understands natural language queries, reduces pipeline complexity by an order of magnitude, processes annotation grades at scales well beyond human capacity, and se…

What tools did this team use?

LLM, embedding-based retrieval, GPU, vector database, RAG, Tool Calling, FSDP, Deep and Cross Network architecture.

What results were reported?

Model pipeline complexity reduction: reducing the model pipeline complexity by an order of magnitude; Automated annotation throughput: millions or tens of millions of grades per day; Retrieval latency: a few milliseconds per query; Developer velocity: significantly enhanced developer velocity (source-reported, not independently verified).

What failed first in this deployment?

LinkedIn's legacy candidate selection and ranking models used fixed taxonomy-based methods and older LLM technology that lacked capacity for deep semantic understanding, and the multi-stage pipeline had become too com…

How is this recruiting AI workflow structured?

Natural language query submitted → Query engine processes intent → Tool Calling on fine-tuned LLM → Personalized facet suggestions via RAG → GPU-accelerated embedding retrieval → Distilled SLM ranking → LLM annotation feedback loop.

Related recruiting cases