--- title: Profile Relevance: Boosting Feed Relevance with LLMs at... description: LinkedIn built Profile Relevance, an unsupervised retrieval system using LLM-generated embeddings to match members with posts outside their network. It... canonical: https://daily.dev/posts/profile-relevance-boosting-feed-relevance-with-llms-at-linkedin-b6ds0bn0d twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Profile Relevance: Boosting Feed Relevance with LLMs at LinkedIn | daily.dev og:description: LinkedIn built Profile Relevance, an unsupervised retrieval system using LLM-generated embeddings to match members with posts outside their network. It... og:url: https://daily.dev/posts/profile-relevance-boosting-feed-relevance-with-llms-at-linkedin-b6ds0bn0d og:image: https://api.daily.dev/og/posts/b6DS0Bn0d.png og:image:alt: Profile Relevance: Boosting Feed Relevance with LLMs at LinkedIn og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Profile Relevance: Boosting Feed Relevance with LLMs at LinkedIn **[LinkedIn Engineering](https://daily.dev/sources/linkedin)** · 11 min read · 0 upvotes · 0 comments ## Summary LinkedIn built Profile Relevance, an unsupervised retrieval system using LLM-generated embeddings to match members with posts outside their network. It compares Rewrite (condensed text, cheaper) versus Non-Rewrite (raw text, richer context) embedding approaches, and appends country names to recover geographic signal lost in rewriting. An AutoEval LLM-judging pipeline scaled evaluation from 20 to 1,000 human-labeled records per model. Incremental profile updates cut compute to 24% of a full-refresh baseline, and horizontal GPU scaling beat vertical scaling. Online tests showed a +0.54% lift in Feed Updates Viewed, +0.94% viral rate, -18.8% hide rate, and +17.7% viral engagement gain for cold-start members. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How does LinkedIn use LLM embeddings to recommend posts from outside your network? LinkedIn built Profile Relevance, an unsupervised retrieval system that generates LLM embeddings from a member's public profile (headline, skills, work and education history) and matches them against post embeddings. It compares a Rewrite approach, where a large generative model extracts topics before embedding, against a Non-Rewrite approach using a retrieval-optimized embedding model directly on raw text, which better preserves subtle context like geography. _Anyone tuning recommendation embeddings can follow daily.dev for more feed-ranking engineering writeups._ ### What is the performance difference between vertical and horizontal GPU scaling for LLM inference pipelines? Horizontal scaling, running many smaller pods in parallel, proved more efficient than vertical scaling for LLM inference at LinkedIn. Vertical scaling (packing more work onto a single large pod) often left GPUs idle waiting on data loading and preprocessing, while horizontal scaling kept GPUs consistently fed and absorbed workload bursts and stragglers more gracefully, improving runtime efficiency and cluster utilization. _Teams sizing GPU inference clusters can track real-world scaling tradeoffs like this on daily.dev._ ### How much can incremental embedding updates reduce compute cost compared to recomputing all user profiles? LinkedIn's incremental update framework, which refreshes embeddings only for members whose profile prompts changed, reduced compute load to just 24% of a naive approach that recomputes inference for every member. This keeps recommendations aligned with the latest profile edits without the cost of full re-inference across the entire user base. _Engineers designing incremental refresh pipelines can follow daily.dev for similar cost-optimization case studies._ ## Similar posts on daily.dev - [How LinkedIn Feed Uses LLMs to Serve 1.3 Billion Users](https://daily.dev/posts/how-linkedin-feed-uses-llms-to-serve-1-3-billion-users-wymxiezwf) · ByteByteGo · 20 upvotes · 1 comments - [Rebuilding LinkedIn’s Follows Recommendations with LLM-Based Semantic Retrieval and Ranking](https://daily.dev/posts/rebuilding-linkedin-s-follows-recommendations-with-llm-based-semantic-retrieval-and-ranking-gkrfwikie) · LinkedIn Engineering · 0 upvotes · 0 comments - [Using LLMs to build content embeddings for DoorDash search and recommendations](https://daily.dev/posts/using-llms-to-build-content-embeddings-for-doordash-search-and-recommendations-fuy4zkqul) · Doordash · 0 upvotes · 0 comments - [Reimagining LinkedIn’s search tech stack](https://daily.dev/posts/reimagining-linkedin-s-search-tech-stack-mwajgbtsf) · LinkedIn Engineering · 1 upvotes · 0 comments - [LLM-Powered Relevance Assessment for Pinterest Search](https://daily.dev/posts/llm-powered-relevance-assessment-for-pinterest-search-zx5akjrjt) · Pinterest Engineering · 1 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#linkedin](https://daily.dev/tags/linkedin), [#embeddings](https://daily.dev/tags/embeddings), [#recommendation-systems](https://daily.dev/tags/recommendation-systems) [View this post on daily.dev](https://daily.dev/posts/profile-relevance-boosting-feed-relevance-with-llms-at-linkedin-b6ds0bn0d) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Profile Relevance: Boosting Feed Relevance with LLMs at LinkedIn","url":"https://daily.dev/posts/profile-relevance-boosting-feed-relevance-with-llms-at-linkedin-b6ds0bn0d","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/profile-relevance-boosting-feed-relevance-with-llms-at-linkedin-b6ds0bn0d"},"datePublished":"2026-11-09T06:31:08.893Z","dateModified":"2026-11-09T06:31:39.780Z","description":"LinkedIn built Profile Relevance, an unsupervised retrieval system using LLM-generated embeddings to match members with posts outside their network. It...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/1ab8d10314c6191e2c1d4b580138601e?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/1ab8d10314c6191e2c1d4b580138601e?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"LinkedIn Engineering","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"LinkedIn Engineering","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/le","url":"https://daily.dev/sources/linkedin"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/profile-relevance-boosting-feed-relevance-with-llms-at-linkedin-b6ds0bn0d","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,linkedin,embeddings,recommendation-systems","timeRequired":"PT11M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"LinkedIn Engineering","item":"https://daily.dev/sources/linkedin"},{"@type":"ListItem","position":3,"name":"Profile Relevance: Boosting Feed Relevance with LLMs at LinkedIn"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/profile-relevance-boosting-feed-relevance-with-llms-at-linkedin-b6ds0bn0d#faq","mainEntity":[{"@type":"Question","name":"How does LinkedIn use LLM embeddings to recommend posts from outside your network?","acceptedAnswer":{"@type":"Answer","text":"LinkedIn built Profile Relevance, an unsupervised retrieval system that generates LLM embeddings from a member's public profile (headline, skills, work and education history) and matches them against post embeddings. It compares a Rewrite approach, where a large generative model extracts topics before embedding, against a Non-Rewrite approach using a retrieval-optimized embedding model directly on raw text, which better preserves subtle context like geography. Anyone tuning recommendation embeddings can follow daily.dev for more feed-ranking engineering writeups."}},{"@type":"Question","name":"What is the performance difference between vertical and horizontal GPU scaling for LLM inference pipelines?","acceptedAnswer":{"@type":"Answer","text":"Horizontal scaling, running many smaller pods in parallel, proved more efficient than vertical scaling for LLM inference at LinkedIn. Vertical scaling (packing more work onto a single large pod) often left GPUs idle waiting on data loading and preprocessing, while horizontal scaling kept GPUs consistently fed and absorbed workload bursts and stragglers more gracefully, improving runtime efficiency and cluster utilization. Teams sizing GPU inference clusters can track real-world scaling tradeoffs like this on daily.dev."}},{"@type":"Question","name":"How much can incremental embedding updates reduce compute cost compared to recomputing all user profiles?","acceptedAnswer":{"@type":"Answer","text":"LinkedIn's incremental update framework, which refreshes embeddings only for members whose profile prompts changed, reduced compute load to just 24% of a naive approach that recomputes inference for every member. This keeps recommendations aligned with the latest profile edits without the cost of full re-inference across the entire user base. Engineers designing incremental refresh pipelines can follow daily.dev for similar cost-optimization case studies."}}]} ```