Meituan's search 3.0: LLM semantic embeddings lift long-tail NDCG by 2.21pp across three iterations
美团技术团队 · wechat · 2026-08-20
Meituan's search team details three iterations of injecting LLM semantic embeddings into service-retail search ranking: generating vectors for queries, merchants (POI) and deals, then feeding bucketed cosine similarity into the ranking model to bridge the semantic gap in long-tail categories.
Phase 1 (feasibility): Small open-source base with full fine-tuning, three special tokens as aggregation anchors, information isolation via AttentionMask across three forward passes; dual loss (match classification + CTR) on 30M samples. Cosine similarity bucketed into 10 bins with learnable 12-dim embeddings. Online: overall orders +0.20%, long-tail NDCG@5 +2.21pp, BadCase@1 -2.96pp.
Phase 2 (systematic upgrade): Samples expanded to quintets with same-request hard negatives (27.66M); counter-intuitive findings — concise info-statement prompts beat complex task instructions, and CPV attributes added noise; base switched to a mid-size embedding model with LoRA (r=8), where LoRA beat full fine-tuning on NDCG; extraction redesigned to independent sequences with a learnable vector, adding deals into a Query-POI-Deal representation system.
More from Companies & People
- Elorian Hiring: Building AI Systems for Native Visual Reasoning — dhruv2038 · 2026-08-20
- 200+ Interviews: The 5 Most Useful Questions — yangyi · 2026-08-20
- Beijing robot output up 2.3x in H1; Zhiyuan A3 wins China & EU certifications — 创业邦 · 2026-08-20
- Criticism: Anthropic Accused of Taking Credit for Others' Work — ziv_ravid · 2026-08-20
- Ex-SpaceX CIO Ken Venner joins Senra Systems to automate wire harness manufacturing — FinanceYF5 · 2026-08-20
- Convergent agent architectures: Identity separation and human tiebreakers — __hymn · 2026-08-20