With the deep evolution of large language model (LLM) technology, search engines are undergoing a third paradigm shift, moving from traditional keyword-based text matching toward a 3.0 era centered on complex intent understanding and cognitive decision-making. Recently, the Meituan technical team published an in-depth technical analysis, comprehensively revealing three iterations of Meituan Search 3.0 in service retail ranking scenarios, and exposing the implementation path and core value of LLM semantic representation in complex local life scenarios.
I. Breaking Through Service Retail: From Traditional Lexical Matching to LLM Semantic Reconstruction
Service retail differs fundamentally from product retail in business form. Compared to the highly standardized product retail, Meituan's service retail and local lifestyle scenarios feature a wide variety of categories, diverse types of search demands (transactional, informational, and lead-generation), and highly non-standardized supply.
In daily searches, traditional ranking models rely heavily on lexical matching. However, when facing massive long-tail categories and complex user intents, the generalization ability of traditional feature engineering is clearly insufficient. For example, when a user searches for "pet spa + bath," the corresponding merchant or product name might be "Pet Cleaning and Care Package"; when searching for "Spring Festival cleaning," it could be "Deep Cleaning Service Package." In such typical cases, there is little overlap between the query and the supply on the surface, but their underlying semantics are highly related.
To bridge this traditional semantic gap, Meituan's service retail search ranking team has systematically explored the application of LLMs in ranking models over the past year. By generating high-quality semantic vector representations for search queries (Query), merchants (POI), and deals (Deal), they injected semantic matching signals into the ranking model using cosine similarity, achieving a leap from single-point feature validation to systematic reconstruction, and then to cross-scenario migration and reuse.
II. Three Stages of Technical Evolution: From Feasibility Verification to Full-Scale System Construction
1. Phase I: Introducing LLM Semantic Representation into Ranking (Verifying Feasibility)
In the first phase, the team's core goal was simple: to verify whether LLM representations could provide substantial assistance to the ranking model.
- Technical Design: A small-parameter open-source base model was selected for full-parameter fine-tuning, with special tokens introduced into the vocabulary as aggregation anchors. Attention masks were carefully designed to isolate queries from merchant information.
- Feature Injection: After the model generated embeddings, cosine similarity was calculated and discretely binned, then concatenated as learnable embeddings into the ranking model features.
- Online Results: Offline validation showed a significant improvement in click NDCG. 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, with a notable reduction in bad cases. This directly proved the great potential of LLM semantic representation in the ranking scenario.
2. Phase II: Systematic Upgrade of Merchant Ranking Representations
Addressing the pain points exposed in Phase I, such as the lack of semantic representation on the product side, high costs of full-parameter fine-tuning, and incomplete optimization objectives, Phase II involved a systematic reconstruction of the entire representation production process.
- Pentuple and Contrastive Learning: A pentuple training dataset was built, including Query, positive samples of products, positive samples of merchants, and hard negative samples. The traditional binary classification objective was completely abandoned, and InfoNCE Loss and Triplet Loss were fully introduced, perfectly solving the issue of "relative order among multiple candidates" that the ranking model truly cares about.
- Lightweight and Efficient Extraction: Transitioning from full-parameter fine-tuning to LoRA methods, independent sequences and learnable vectors replaced the old approach, and dimensionality reduction was upgraded to MRL-E (Matryoshka Representation Learning), improving training and inference efficiency while supporting flexible multi-scale applications.
- Online Results: The click GAUC and conversion efficiency of merchant ranking improved significantly. Online data showed a significant rise in effective clicks and result page conversion rate (QV_CTR), fully demonstrating that LLM semantic representation not only brings more exposure opportunities but also optimizes overall conversion quality through precise recommendations.
3. Phase III: Introducing Representations and Cross-Features in Downstream Ranking
After completing the systematic reconstruction of merchant ranking, the team further applied the mature solution to downstream ranking (i.e., ranking specific products under a merchant), while simultaneously advancing the governance and supplementation of cross-statistical features across the board.
- Solving Coverage Challenges: Due to the fundamental differences in query distribution between merchant ranking and downstream ranking, the initial migration faced challenges with low query coverage. Through fine-grained alignment and data governance on both ends, the team successfully extended the LLM semantic representation capability seamlessly to the product level.
- Parallel Progress: Combining "LLM representation transfer" with modeling of cross dimensions such as "personalization × product" and "query intent × product," the search results achieved another leap in precision.
III. Summary and Outlook
The evolution of Meituan Search 3.0 shows that the application of LLMs in the local lifestyle field is not an overnight success, but requires a solid path from single-point feature validation, systematic representation reconstruction, to cross-scenario generalization and reuse. By transforming the powerful semantic understanding capability of LLMs into dense continuous representations and discrete binning features, Meituan has successfully broken down the barriers between complex intents and non-standardized supply in service retail scenarios, providing users with a more accurate and intelligent local lifestyle search experience.
