Skip to main content

Meituan Search 3.0: How Large Models Are Rebuilding Local Life Search

When you type "pet spa + bath" into Meituan, the app doesn't just look for those exact words. It understands you're probably after a grooming service, even if the merchant's listing says "Pet Cleaning and Care Package." This kind of semantic understanding is the core of Meituan Search 3.0, a system that's been quietly rebuilt from the ground up using large language models (LLMs).

For years, search engines relied on lexical matching—basically, matching the words in your query to words in a listing. But in the messy, varied world of local life services, that approach falls short. A search for "Spring Festival cleaning" might not match a "Deep Cleaning Service Package," even though that's exactly what you need. The gap between what users mean and what merchants list has always been a headache.

Meituan's technical team recently shared a detailed breakdown of how they've tackled this problem. Their solution? A three-phase evolution that's turning the search experience from a guessing game into a precise, intent-driven process.

Phase One: Testing the Waters

The first step was simple: could LLMs actually help? The team fine-tuned a small, open-source model to generate semantic vectors for queries, merchants, and deals. They then calculated cosine similarity between these vectors and fed that as a feature into their ranking model.

The results were promising. Offline tests showed a significant boost in click-through rates, and online, overall search payment orders and service retail orders saw a notable uptick. Long-tail searches—those obscure, specific queries where traditional matching failed—improved dramatically. This was the proof they needed: LLMs could bridge the semantic gap.

Phase Two: Building a Better Representation

But Phase One had its limitations. The product side lacked semantic representation, full-parameter fine-tuning was expensive, and the optimization objectives weren't complete. So, the team went back to the drawing board.

They introduced a pentuple training dataset—including queries, positive product samples, positive merchant samples, and hard negatives—and switched to contrastive learning with InfoNCE and Triplet losses. This was a shift from binary classification to a more nuanced understanding of relative order among candidates.

They also moved to LoRA for lightweight fine-tuning and adopted Matryoshka Representation Learning for flexible, multi-scale embeddings. The result? Merchant ranking saw significant improvements in click GAUC and conversion efficiency. Effective clicks and result page conversion rates (QV_CTR) rose, showing that better representation leads to not just more exposure, but better matches.

Phase Three: Scaling Up and Down

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

They also combined LLM representation transfer with cross-dimensional modeling, like "personalization × product" and "query intent × product." This dual approach pushed search precision even higher.

What This Means for You

For the average user, this technical overhaul translates to a search experience that just feels smarter. You're less likely to wade through irrelevant results or miss the perfect service because the merchant used different wording. Whether you're looking for a last-minute plumber or a niche hobby class, Meituan's search is getting better at reading your mind—or at least, your intent.

The journey from keyword matching to semantic understanding isn't overnight. It's a steady climb from validating a single feature to rebuilding the entire representation pipeline. But as Meituan's experience shows, the payoff is real: a local life search that's more accurate, more intuitive, and ultimately, more useful.

Key Points

  • Meituan Search 3.0 uses large language models to understand complex user intents, not just keywords.
  • The system evolved in three phases: feasibility testing, systematic representation upgrade, and downstream ranking integration.
  • Contrastive learning and lightweight fine-tuning (LoRA) were key to improving efficiency and accuracy.
  • The result is better search precision, especially for long-tail and non-standardized service queries.
  • This technical shift aims to make local life search more intuitive and effective for everyday users.