How Alibaba Cloud’s AI-Powered OpenSearch Boosts Search Accuracy and Cuts Costs

Based on the 2023 Cloud Expo conference, Alibaba Cloud senior technical expert Guo Ruijie presented the intelligent upgrade of Alibaba Cloud Search products.

Conversational Search with Enterprise‑Specific Large Models

The OpenSearch LLM Q&A edition offers a SaaS‑style conversational search solution that integrates Alibaba’s Tongyi Qianwen and third‑party open‑source large language models, allowing users to build enterprise‑specific models on business data. It includes paragraph splitting, vectorization, a vector engine, and retrieval‑enhanced LLM capabilities, enabling minute‑level PoC and hour‑level production launch. Compared with open‑source models, it improves baseline answer accuracy by about 20% and reduces hallucinations by 40%. Inference optimizations increase token generation speed 2–3× and cut GPU usage by 50%.

Example: the Shilin platform for pharmaceutical compliance uses OpenSearch to provide natural‑language Q&A over regulations, delivering answers with clear references without repeated keyword searches.

Alibaba Cloud also released SmartArxiv, an AI‑enhanced academic paper assistant built on the OpenSearch Q&A edition, supporting research, rapid paper reading, method comparison, and literature reviews.

Vector Search Performance Improves Over Open‑Source Engines

The OpenSearch vector search edition has transitioned from PaaS to Serverless, improving usability. Its core engine upgraded to VectorStore delivers millisecond‑level responses for billions of vectors and second‑level real‑time updates. Compared with mainstream open‑source vector engines, retrieval performance is more than three times faster and memory usage is reduced by 50%. It also supports tag‑plus‑vector hybrid search and end‑to‑end image‑to‑vector search scenarios.

VectorStore, built on the Havenask engine, offers high‑performance vector algorithms, data compression, and non‑full‑memory loading, reducing costs. It is widely used in Alibaba’s e‑commerce personalization, recommendation, multimodal search, and large‑model applications, and will be open‑sourced in Havenask.

Cost‑Effective Elasticsearch Serverless Edition

Alibaba Cloud’s Elasticsearch Serverless product provides a compatible, on‑demand service that automatically scales resources based on traffic, achieving second‑level elasticity and pay‑as‑you‑go pricing, cutting idle resource costs. Optimizations across hardware selection, cluster architecture, and kernel performance boost write performance by 150% and lower per‑unit storage cost by 70%.

The service integrates Elasticsearch 8.9 and the Elasticsearch Relevance Engine (ESRE) to offer AI‑enhanced capabilities such as RRF hybrid ranking and third‑party model integration.