This article is compiled based on the transcript of the speech at the 2023 Yunqi Conference. The speech information is as follows:
Speaker : Guo Ruijie | Senior technical expert and head of search at Alibaba Cloud
Speech topic : Alibaba Cloud search product intelligent upgrade released
Recently at the 2023 Yunqi Conference, Alibaba Cloud Search Director Guo Ruijie released an intelligent upgrade of Alibaba Cloud search products, providing intelligent product capabilities in all-scenario search segments, including large model-based Open Search (OpenSearch) LLM Intelligent question and answer version and vector retrieval version, as well as Elasticsearch Serverless version for log and text analysis scenarios, etc.
Conversational search based on enterprise-specific large models
OpenSearch LLM Intelligent Question and Answer Edition is a one-stop SaaS conversational search solution with built-in Tongyi Qianwen and third-party open source large language models. Users can build enterprise-specific models based on business data. The product integrates full-link conversational search capabilities such as paragraph segmentation, vectorization model, vector engine, and retrieval enhancement large model, and can achieve minute-level PoC and hour-level official launch. Compared with the open source large model, it has a higher accuracy baseline, the answer accuracy is increased by about 20%, and the hallucinations are reduced by 40%. Open Search has done a lot of inference optimization, the token generation speed has been increased by 2-3 times, and the GPU resource usage has been reduced by 50%.
Shilin is a platform focused on collecting and organizing knowledge in the field of pharmaceutical compliance, providing pharmaceutical companies with legal and regulatory retrieval and consulting services throughout the life cycle from drug development, production to changes. In the first half of this year, Shilin built an intelligent question and answer system with a large language model based on OpenSearch. Now, users no longer need to search repeatedly through keywords or obtain answers by browsing the original texts of laws and regulations. Instead, they ask questions in a natural language. Shilin The platform understands the intent of the question, quickly reaches the final answer through the large model, and clearly displays the reference basis.
At the same time, Alibaba Cloud launched SmartArxiv, an intelligent academic paper assistant built on the open search question and answer board, which supports academic research, speed reading papers, method comparison, literature review and other functions to help solve the problem of "information overload" in academia. It is currently online on the ModelScope official website and can be used by users for free.
https://modelscope.cn/studios/smart_arxiv_opensearch/SmartArXiv
Compared with open source engines, vector retrieval performance is improved by 3 times
In this new upgrade, the open search vector retrieval version has evolved from the original PaaS form to a serverless form, and the ease of use has been significantly improved. The core engine is upgraded to the new engine VectorStore, which significantly improves performance. It supports millisecond response to hundreds of billions of data and real-time data updates in seconds. Compared with the industry's mainstream open source vector engines, the retrieval performance is improved by more than 3 times, and the memory usage is reduced to 50%. It also supports tag + vector hybrid retrieval capabilities, providing an end-to-end image search scenario solution from image vectorization to vector storage query.
VectorStore's differentiated feature is high performance. It has done a lot of performance optimization of vector retrieval algorithms. It can respond to queries on hundreds of billions of vector data in milliseconds, and vector updates can be seen in seconds. A lot of data compression optimization has been done to support non-full memory loading and lower costs. As Alibaba's self-developed vector engine, it is widely used in Taobao's personalized search, recommendations, multi-modal search, and various recent large model applications. The execution framework of VectorStore is based on the Havenask engine and will be fully open sourced in Havenask later.
Elasticsearch Serverless version with 70% cost reduction
Alibaba Cloud's retrieval and analysis service Elasticsearch has officially launched a serverless product that is compatible with open source and can be used on demand. Users no longer need to manually allocate resources. The platform will automatically schedule and determine the increase or decrease of resources according to fluctuations in business traffic, enabling second-level elastic expansion and contraction, truly realizing Pay-as-you-go pay-as-you-go dynamically matches loads and resources, reducing redundant costs for idle resources. Based on the Indexing Service index construction service and the OpenStore storage service, Alibaba Cloud has comprehensively optimized hardware selection, cluster architecture, and kernel performance, achieving a 150% increase in write performance and a 70% reduction in unit data storage costs. At the same time, Alibaba Cloud Elasticsearch has integrated the latest version 8.9 and provides AI enhancement capabilities based on the ESRE component (Elasticsearch Relevance Engine), including RRF hybrid sorting algorithm, third-party model integration, deployment and management, etc.
