Meituan Search 3.0 Rebuilds Local-Life Search With Large Language Models

Meituan's engineering team has released a technical analysis describing three successive upgrades to its search infrastructure, branded Meituan Search 3.0, which replaces conventional lexical matching with large-language-model (LLM) semantic representations in service-retail and local-lifestyle ranking scenarios, according to reporting published on 23 September 2026.

The Problem: Keyword Matching Falls Short in Service Retail

According to the source, service retail differs structurally from product retail because it involves "diverse categories, varied search demands (transactional, informational, and lead-generation), and highly non-standardized supply." The team identified a core gap in traditional systems: a user searching for "pet spa + bath" might need a result labelled "Pet Cleaning and Care Package," while a search for "Spring Festival cleaning" corresponds to a "Deep Cleaning Service Package." In both cases, the source notes, "there is little overlap between the query and the supply on the surface, but their underlying semantics are highly related."

Phase I: Proving LLMs Can Aid Ranking

In the first phase, the team selected "a small-parameter open-source base model" for full-parameter fine-tuning, introducing special tokens as aggregation anchors and designing attention masks to isolate queries from merchant information. Cosine similarity scores were discretely binned and injected as learnable embeddings into the ranking model. The source reports that offline validation showed a significant improvement in click NDCG, and after deployment, "the overall search payment orders and service retail orders saw a significant increase, especially in long-tail scenarios where traditional lexical matching was weakest."

Phase II: Systematic Reconstruction of Merchant Ranking

The second phase addressed limitations exposed in Phase I, including the absence of product-side semantic representations, the high cost of full-parameter fine-tuning, and incomplete optimisation objectives. The team built a "pentuple training dataset" containing queries, positive product samples, positive merchant samples, and hard negative samples, and replaced a binary classification loss with InfoNCE Loss and Triplet Loss to handle "relative order among multiple candidates." Training efficiency was improved by moving from full-parameter fine-tuning to LoRA methods and by upgrading dimensionality reduction to MRL-E (Matryoshka Representation Learning). The source states that online results showed a "significant rise in effective clicks and result page conversion rate (QV_CTR)."

Phase III: Extending Representations to Downstream Product Ranking

In the third phase, the mature merchant-ranking solution was carried over to downstream ranking of specific products within a merchant. Initial migration faced "challenges with low query coverage" due to differences in query distribution between merchant and product ranking, which the team resolved through "fine-grained alignment and data governance on both ends." Simultaneously, the team modelled cross-dimensional features combining "personalization × product" and "query intent × product," which the source says produced "another leap in precision."

Broader Context

The source characterises these developments as part of a wider "third paradigm shift" in search, describing a transition "from traditional keyword-based text matching to a 3.0 era focused on complex intent understanding and cognitive decision-making." Meituan's team frames the work as evidence that LLM adoption in local-lifestyle applications "requires a solid path from single-point feature validation, systematic representation reconstruction, to cross-scenario generalization and reuse" rather than being an immediate transformation.