Skip to main content

Meituan Search 3.0: How Large Models Rebuild Local Life Search

When you search for "pet spa + bath" on Meituan, you're not just typing words—you're expressing a need. But traditional search engines often struggle to connect that query with a merchant offering a "Pet Cleaning and Care Package." The words don't match, even though the intent is clear. This is the challenge Meituan's Search 3.0 aims to solve.

In a recent technical deep-dive, Meituan's team revealed how they're using large language models (LLMs) to rebuild their search infrastructure from the ground up. The goal? To move beyond surface-level keyword matching and truly understand what users mean, even when they don't use the exact words a business does.

The Service Retail Puzzle

Unlike product retail, where items are standardized and easily categorized, service retail is messy. Think about the variety: from haircuts to home repairs, from pet grooming to deep cleaning. The search queries are just as diverse—transactional ("book a haircut"), informational ("how much does a deep clean cost?"), and lead-generation ("find a plumber near me").

Traditional ranking models rely heavily on lexical matching—they look for the same words in the query and the merchant's description. But this approach falls short with long-tail categories and complex intents. For instance, searching "Spring Festival cleaning" might not match a merchant offering "Deep Cleaning Service Package." The words don't overlap, but the meaning is spot on.

Three Steps to Smarter Search

Meituan's team didn't jump straight to a full overhaul. Instead, they took a phased approach, each step building on the last.

Phase 1: Proving It Works

The first step was simple: could LLM representations actually help ranking? The team fine-tuned a small open-source model, adding special tokens to act as anchors. They carefully designed attention masks to keep queries and merchant info separate. Then, they calculated cosine similarity between the query and merchant embeddings, binned the results, and fed them into the ranking model as features.

The offline results were promising—click NDCG improved significantly. Online, search payment orders and service retail orders saw a notable boost, especially in long-tail scenarios where traditional matching was weakest. This was the proof they needed.

Phase 2: Building a Better Representation

Phase 1 exposed some pain points: the product side lacked semantic representation, full-parameter fine-tuning was expensive, and the optimization objectives were incomplete. So, the team rebuilt the entire representation pipeline.

They introduced a pentuple training dataset, including queries, positive product samples, positive merchant samples, and hard negatives. They ditched the binary classification objective in favor of InfoNCE Loss and Triplet Loss, which better capture the relative order among multiple candidates—exactly what ranking models care about.

They also switched to LoRA for lightweight fine-tuning and adopted MRL-E (Matryoshka Representation Learning) for flexible, multi-scale embeddings. The result? Merchant ranking saw significant improvements in click GAUC and conversion efficiency. Online, effective clicks and result page conversion rate (QV_CTR) rose, showing that LLM representations not only bring more exposure but also better conversion quality.

Phase 3: Going Deeper

With merchant ranking sorted, the team turned to downstream ranking—ranking specific products under a merchant. This wasn't a simple transfer. The query distribution was different, and initial coverage was low. Through fine-grained alignment and data governance, they extended the semantic representation capability to the product level.

They also combined LLM representation transfer with cross-dimensional modeling, like "personalization × product" and "query intent × product." This double-pronged approach led to another leap in search precision.

The Road Ahead

Meituan's journey with Search 3.0 shows that applying LLMs to local life search isn't an overnight success. It requires a solid path: from single-point feature validation, to systematic representation reconstruction, to cross-scenario generalization. By turning LLMs' powerful semantic understanding into dense continuous representations and discrete binning features, Meituan has broken down the barriers between complex intents and non-standardized supply. The result is a more accurate, intuitive search experience that understands you—even when you don't say the exact words.

Image

Key Points

  • Semantic Gap: Traditional keyword matching fails when queries and merchant descriptions don't share words, even if they're semantically related.
  • Phased Approach: Meituan validated LLM representations in ranking, then systematically rebuilt the representation pipeline, and finally extended it to downstream ranking.
  • Technical Innovations: Use of pentuple training data, InfoNCE/Triplet Loss, LoRA fine-tuning, and MRL-E for efficient, flexible embeddings.
  • Real-World Impact: Significant improvements in click NDCG, conversion rates, and long-tail search performance.
  • Future Outlook: The success paves the way for more intelligent local life search, bridging the gap between user intent and service supply.