How LinkedIn search works: dual-tower embedding recall feeding multi-stage ranking
_jaydeepkarale · x · 2026-11-06
A technical walkthrough of LinkedIn's content search retrieval layer: a token-based retriever fetches posts matching query keywords, while an embedding-based retriever uses a two-tower architecture (text embedding model + MLP per tower) to encode queries and posts, ranking by cosine similarity. Post embeddings are pre-computed offline in batch and stored. Retrieved candidates then feed a multi-stage ranking layer.
More from Research
- ~150 async RL runs: score centering suffices as long as trainer stays close to sampler — brandondamos · 2026-11-06
- Inaugural Human-AI Interaction Conference HAIC 2027 launches, CFP opens — tongshuangwu · 2026-11-06
- Amplify Bio Bets on AI×Bio, Welcomes Scientist-Turned-Founder Luke Riggan — sarahcat21 · 2026-11-06
- Vault paper from the Kundaje lab is now up on bioRxiv — anshulkundaje · 2026-11-06
- New alignment research: LLMs are score-seeking, not reward-seeking — LessWrong 精选 · 2026-11-06
- Critical voices question AI companies' 'publications' lacking peer review — anshulkundaje · 2026-11-06