How Trip.com Group Makes Hotel Search Seamless for Travelers Worldwide with Zilliz Cloud
Smarter, Intent-Driven Search
Delivers accurate, multilingual hotel results that understand traveler intent—not just keywords.
Faster Innovation, Lower Overhead
Eliminates cluster management and scaling tasks, freeing teams to focus on innovation.
<100 ms Latency Optimization
Delivers lightning-fast, real-time hotel search with P99.9 response times under 100 ms.
Seamless Global Scalability
Powers real-time search across hundreds of millions of listings with enterprise reliability and low latency.
About Trip.com Group
Trip.com Group is one of the world’s largest online travel service providers, uniting well-known brands including Trip.com, Ctrip, Skyscanner, Qunar, and Travix. Founded in 1999, the company has grown into a truly global platform, serving travelers across more than 200 countries and regions.
With a portfolio that covers flights, hotels, trains, car rentals, and corporate travel, Trip.com Group helps millions of users seamlessly plan and book every stage of their journey. Its platform offers access to over 1.2 million accommodation choices, advanced search and booking features, and multilingual support, making it a trusted one-stop destination for travelers worldwide.
The Challenge: Keyword Search Falls Short in a Global, Multi-Lingual World
Hotel search is central to Trip.com Group’s platforms, helping millions of travelers find the perfect place to stay from a vast global inventory. Yet, as the company expanded into new markets, its legacy keyword-based search often returned incomplete or irrelevant results, resulting in a fragmented customer experience. What was missing was an intent-driven approach—one that could interpret meaning, not just match words, and surface properties that truly align with traveler needs.
Instead, customers often faced poor results. For example, a traveler searching for “beautiful night view” might never see hotels described with “scenic evening view.” Someone typing “allows dogs” wouldn’t be shown listings marked as “pet-friendly.” Even small variations—“king bed” versus “large bed”—led to mismatches. And across dozens of languages and regional spellings, the problem only grew, leaving many travelers to settle for sub-optimal bookings or abandon the platform altogether.
The challenge went beyond semantics. Hotel search is inherently multi-dimensional: travelers weigh location, amenities, reviews, and price together. A query like “quiet boutique hotel near Central Park with free Wi-Fi and late checkout” should return a handful of perfect matches—but keyword-based systems struggled to reconcile all these signals at once.
These search shortcomings weren’t minor inconveniences—they translated directly into lost bookings and dissatisfied customers. To deliver a truly global, seamless search experience, Trip.com Group needed a smarter and more flexible foundation.
Early Research on Vector Search
To overcome the limits of keyword matching, the Trip.com team launched a proof of concept to explore vector search. They used OpenAI’s text-embedding-ada-002 model to convert both hotel names and traveler queries into dense embeddings, then compared them using cosine similarity. The goal was simple: test whether semantic retrieval could bridge gaps that keyword systems routinely missed.
The results were encouraging. For example, the query “allow pets” clustered closely with variations like “carry pets,” “able to bring pets,” and “pet friendly,” all scoring above 0.94 in similarity. These early signals confirmed that vector search was capable of recognizing intent across different wordings—something traditional keyword search simply could not achieve.
This validation gave the team confidence to move forward. If a lightweight experiment could already capture semantic meaning across small variations, then a production-grade vector database had the potential to unlock richer, multilingual, and context-aware hotel search at scale.
The Journey from Milvus to Zilliz Cloud: Evaluation and Migration
When evaluating the right foundation for hotel discovery, the Trip.com team compared three categories of solutions.
Vector libraries, such as Faiss and SPTag, offered strong algorithms but came with heavy engineering and maintenance requirements, lacking sophisticated features.
Traditional databases with vector add-ons, such as Elasticsearch and PostgreSQL, were easy to adopt but couldn’t scale elastically to meet Trip.com’s global traffic demands.
Purpose-built vector databases appeared promising, but their maturity varied—some lacked enterprise-grade features, while others had limited community support.
After a thorough evaluation, Milvus emerged as the best fit. It aligned naturally with Trip.com’s cloud-native stack, offering separation of compute and storage for flexible scaling and seamless integration with Kubernetes and distributed storage. Features such as string primary keys and a broad range of index types gave the team confidence that Milvus could support diverse hotel search scenarios.
To validate performance, they benchmarked two indexes: HNSW, which provided the highest recall accuracy but required more memory (131 GB per replica), and IVF-SQ8, which delivered comparable latency while consuming far less memory (49 GB). Both sustained real-time response times in the 26–36 ms range across millions of listings, giving the team a balanced toolkit to trade off precision and efficiency depending on the workload.
| Index Type | Memory Usage (per replica) | Minimum Deployment | Average Response Time |
|---|---|---|---|
| HNSW | 131 GB | 6 × 4C32G + 1 × 2C4G | TOP5 – 26msTOP10 – 28msTOP20 – 30msTOP50 – 36ms |
| IVF-SQ8 | 49 GB | 1 × 4C64G | TOP5 – 28msTOP10 – 29msTOP20 – 31msTOP50 – 35ms |
Dataset: Full hotel search entities; Data size: 17 million+; Vector dimension: 1024
Migrating to Zilliz Cloud
While Milvus delivered excellent performance, running a distributed database in-house created operational overhead. Cluster management, upgrades, monitoring, and scaling consumed valuable engineering time—resources the team wanted to dedicate to improving the traveler experience. This drove the decision to migrate to Zilliz Cloud, the fully managed service built by the creators of Milvus. Zilliz Cloud addressed the team’s priorities directly:
Operational efficiency by offloading DevOps and performance tuning.
Cost optimization with resource-aware scaling and quantization as data grew into hundreds of millions of hotel images.
Scalability and reliability through built-in high availability, cross-region disaster recovery, and effortless horizontal scaling.
Performance guarantees that consistently met strict SLA requirements of P99.9 latencies under 100 ms—even under peak concurrency, outperforming Elasticsearch.
The migration was seamless. Trip.com gradually shifted workloads from self-managed Milvus clusters to Zilliz Cloud with zero downtime. The result was a fully managed, enterprise-ready vector search foundation that combined operational simplicity with world-class performance at global scale.
The Solution: Semantic Search with Zilliz Cloud and Multilingual-E5
To break past the limits of keyword search, Trip.com built a new semantic search engine with Zilliz Cloud as its vector database. The architecture is multi-layered. Traveler queries are 1) first refined through preprocessing, 2) then vectorized into embeddings, 3) and finally matched against millions of hotel vectors stored in the Zilliz Cloud; 4) Results from semantic and keyword retrieval are merged, ensuring that travelers see the most relevant hotels in real time.
Another major enabler was the multilingual-e5 model, selected for its ability to deliver consistent accuracy across dozens of languages. Hotel descriptions, tags, and facilities were vectorized offline, while traveler queries were processed online at runtime. Zilliz Cloud’s high-performance similarity search identified the nearest matches in under 100 ms, surfacing intent-driven results at scale. Queries such as “beautiful night view” and “nice evening skyline” mapped to the same intent, removing the language and phrasing barriers that had previously caused missed opportunities.
By combining multilingual embeddings with Zilliz Cloud’s fully managed infrastructure, Trip.com achieved the best of both worlds: semantic intelligence that captured traveler intent, lightning-fast retrieval across millions of listings, and enterprise-grade reliability without the burden of operating clusters in-house. The result is a smarter, scalable hotel search system that consistently serves travelers around the world—regardless of the language they use or how they phrase their requests.
Multiple Vector Search Use Cases within Trip.com Group
The new vector engine quickly proved valuable across multiple areas of Trip.com Group’s business.
Hotel Search
For travelers, Zilliz Cloud solved one of the biggest pain points—missed matches caused by phrasing differences. A query for “family-friendly hotel” will now also surface listings described as “great for kids,” while “work-friendly” brings up properties with “business amenities.” By combining semantic search with keyword search, Trip.com achieves better recall of relevant hotels, more precise matches, and greater confidence that travelers are seeing the best options available.
SEO Landing Pages
On the marketing side, vector search transformed how localized landing pages are built. What once required labor-intensive curation is now automated: Zilliz Cloud performs semantic search, retrieves relevant hotels when keyword intent falls short, filters and ranks results, and feeds directly into page generation workflows. This process allows Trip.com to spin up thousands of optimized landing pages across languages and regions, boosting organic traffic while saving significant manual effort.
Image Search
Beyond text, Trip.com extended vector search into the visual domain. With hundreds of millions of hotel images—from room interiors to local attractions—the team enabled travelers to search by image, discovering hotels with similar aesthetics or features. Powered by Zilliz Cloud’s multi-modal capabilities, this feature highlights the platform’s ability to scale beyond text into rich multimedia search experiences.
Together, these use cases demonstrate how vector search is more than a back-end technology—it’s a shared capability that improves front-end traveler experiences and accelerates back-end marketing strategies. For Trip.com, Zilliz Cloud has become a foundation not just for hotel search, but for growth across multiple lines of business.
What Trip.com Gained with Zilliz Cloud
Migrating to Zilliz Cloud gave Trip.com a hotel search platform that is both faster and smarter, with measurable improvements across performance, scalability, and operations:
Latency optimization: Consistently meets P99.9 response times under 100 ms, satisfying strict real-time search requirements.
Scalability at scale: Handles hundreds of millions of text and image vectors without performance degradation.
Higher recall quality: Captures intent-driven matches that keyword systems missed, improving traveler satisfaction and reducing booking drop-offs.
Reduced engineering overhead: Offloads cluster management, upgrades, and scaling to Zilliz, freeing engineers to focus on traveler-facing innovations.
Cost efficiency: Resource optimization and quantization keep costs predictable compared to Elasticsearch or self-managed Milvus.
Enterprise reliability: Built-in high availability and cross-region disaster recovery deliver uninterrupted search for millions of travelers worldwide.
Together, these gains transformed hotel search from a keyword-driven system into an intelligent, scalable, and resilient service—one that delights travelers while giving Trip.com confidence in its long-term growth strategy.
Conclusion
For Trip.com, search is more than a feature—it is the gateway to the travel experience. By adopting Zilliz Cloud, the company reshaped how travelers discover hotels and destinations, while giving internal teams new ways to scale and innovate. Smarter, intent-driven search now powers both better customer journeys and more efficient business growth.
This story highlights a broader shift: managed vector databases are no longer just backend infrastructure—they are a strategic foundation for competitive advantage. With Zilliz Cloud ensuring performance, reliability, and scale, the Trip.com team can focus on the future of travel discovery: seamless, multilingual, and globally consistent search that sets a new industry standard.
- About Trip.com Group
- The Challenge: Keyword Search Falls Short in a Global, Multi-Lingual World
- Early Research on Vector Search
- The Journey from Milvus to Zilliz Cloud: Evaluation and Migration
- The Solution: Semantic Search with Zilliz Cloud and Multilingual-E5
- Multiple Vector Search Use Cases within Trip.com Group
- What Trip.com Gained with Zilliz Cloud
- Conclusion
Content
Industry
Entertainment