How Semantic Search Transforms Hotel Booking: From Entity Recognition to Vector Retrieval

Background

In the fast‑moving AI era, traditional keyword search can no longer satisfy complex user intents such as “2 adults 1 child” or “family‑friendly hotels in Jiang‑Zhe”. Ctrip has built self‑developed intelligent products (e.g., "Wenda", "TripGen") that rely on semantic search to improve recall and user experience in the hotel domain.

Our Goal: Reduce Search Friction

By interpreting natural language queries like “Tokyo 2 adults 1 child 2 rooms next May 5 for 7 days family hotel”, the system automatically extracts location, dates, guest count, and other factors, matches them to filter criteria, and directly presents a relevant hotel list, eliminating manual filter steps.

Core AI Capabilities

Entity Recognition

Deep‑learning models identify key entities (e.g., place, date, room type) in user queries, handling ambiguity and multi‑dimensional intent. Ctrip adopted the QianWen‑7B large language model, whose high‑dimensional embeddings and attention mechanisms enable precise extraction of complex semantic components.

Model inference is accelerated with TensorRT, reducing latency from ~3000 ms to ~300 ms.

Entity Linking

Recognized entities are linked to Ctrip’s OTA knowledge graph via vector embeddings. The pipeline converts entities into dense vectors, then uses a vector engine to retrieve matching hotel records.

For embedding models, BGE‑M3 was chosen for its superior performance in Chinese, balancing recall and precision, while Elasticsearch with HNSW indexing provides fast approximate nearest‑neighbor search.

Vector Retrieval Mechanism

After evaluating vector database options, Ctrip selected Elasticsearch as the core component, integrating HNSW for efficient high‑dimensional search. This solution aligns with the existing tech stack, offering stability, scalability, and low operational cost.

Conclusion and Outlook

Semantic understanding turns search engines from passive information providers into interactive, user‑centric platforms that predict and fulfill complex needs with natural language. Future search systems will continue to embed AI deeper, delivering faster decisions and richer experiences.