TechPrufer
Get started
LinkedIn
LinkedIn logoLinkedIn · Feature trail

Semantic Search Stack

4 posts · Jan 2026 – Aug 2026

What this trail covers

This post details the engineering effort behind LinkedIn Premium All-in-One, a unified product for small businesses. It focuses on leveraging AI and LLMs to create unified workflows for sales, marketing, and hiring. Key technical aspects include an LLM-based approach for inferring Target Audiences (TA) using profile and company information, with privacy guardrails. The TA definition is further refined through activity-based signals and implicit signals (buyer-side and seller-side). The system in this post builds upon this foundation by detailing the semantic search system (MUSE) for AI agents, specifically for the Hiring Assistant. It describes the MUSE platform which learns from expert judgments to encode understanding into a model capable of scoring over a billion member profiles in real time. The system uses a dual-tower Siamese architecture with Matryoshka embeddings for efficient retrieval and ranking. It also details the MUSE Teacher, an LLM-as-a-judge system for scalable supervision, and the MUSE Embeddings architecture, including the query and profile towers and the use of 2048-dimensional truncations for ANN search.

All LinkedIn trails
  1. Building LinkedIn Premium All‑in‑One: Unified AI‑Powered Workflows for Small BusinessesThis post details the engineering effort behind LinkedIn Premium All-in-One, a unified product for small businesses. It focuses on leveraging AI and LLMs to create unified workflows for sales, marketing, and hiring. Key technical aspects include an LLM-based approach for inferring Target Audiences (TA) using profile and company information, with privacy guardrails. The TA definition is further refined through activity-based signals and implicit signals (buyer-side and seller-side). The system in this iteration focuses on unifying these capabilities into a single product experience, optimizing for rapid iteration and early user feedback, and enabling experimentation at speed. It also details features like daily prospect suggestions, trending posts within a target audience, post boosting, a profile writing assistant, and recommended actions, all powered by AI and LLMs.
  2. Rebuilding LinkedIn’s Follows Recommendations with LLM-Based Semantic Retrieval and RankingThis post details the rebuilding of LinkedIn's Follows Recommendation system using LLM-based semantic retrieval and ranking. Key contributions include: transforming profile data into natural language prompts for LLMs, fine-tuning a bi-encoder LLM for follow prediction using supervised contrastive learning (with PEFT, mixed precision, and gradient checkpointing), developing both large-scale offline (Ray, FAISS) and low-latency online (Proxima, hosted vector search) Embedding-Based Retrieval (EBR) pipelines, and integrating LLM embeddings into downstream ranking models via task-aware supervised projection for dimensionality reduction.
  3. Semantic Search for AI Agents at Scale: Retrieval and Ranking for LinkedIn’s Hiring AssistantThis post details the engineering of the MUSE (Member Understanding Semantic Embeddings) platform for LinkedIn's Hiring Assistant. It describes the MUSE Teacher, an LLM-as-a-judge system for scalable supervision, and the MUSE Embeddings architecture, a dual-tower Siamese model using Matryoshka embeddings. The post highlights the use of 2048-dimensional embeddings for ANN search and the challenges of productionizing billion-scale semantic search, including the query understanding, embedding-based retrieval (EBR), and blending/ranking (L2) layers. It also discusses the evaluation and supervision foundation, including the "Expert Judge" LLM and the MUSE Teacher's prompt engineering for improved accuracy.
  4. Reimagining LinkedIn’s search tech stackThis post details the engineering effort behind LinkedIn's semantic search stack, which leverages LLMs for a more intuitive and personalized search experience. It describes the high-level infrastructure involving query understanding, embedding-based retrieval (EBR) on GPUs, and a Cross-Encoder SLM for ranking. Key optimizations include score caching and a ranking-depth controller. The post also elaborates on the process of measuring relevance quality using LLM judges, including defining product policies, training LLM judges via distillation, and continuous quality measurement. It provides technical details on EBR relevance modeling, including the dual-tower architecture, contrastive InfoNCE loss, margin-based ranking loss, and the use of hard positives and negatives for training.