In the past year, the capabilities of large language models have been growing as the technology for large language models has skyrocketed. These have driven the rapid evolution of our search technology to the next generation, the AI search technology. The rapid development of large language models has not only reshaped the foundation of search technology, but also provided strong support for the digital transformation of various industries.
1） AI search technology trend
AI search technology has the following significant features:
Reconstruction: One is the reconstruction of the technical aspects of AI search. the AI search technology is currently undergoing a comprehensive reconstruction based on the large language model, through which the whole chain of search capabilities have been reorganized, including the redefinition of the capabilities of text parsing, slicing and vectorization. Another is the reconstruction of the way of information access and the shape of the product. While traditional search relies on keyword matching, it now uses more natural language Q&A interactions, which brings new business scenarios, such as virtual digital people, Q&A for enterprise knowledge bases and intelligent customer service for e-commerce platforms.
AI Infrastructure: AI search technology has become an important part of AI native applications, including search vector retrieval, semantic search and retrieval generation technologies, which constitute the infrastructure for many AI applications. This not only improves the efficiency of data processing, but also enhances the user interaction experience with the system, helping enterprises realize more efficient information management and services.
Effectiveness enhancement: Currently, the focus on effectiveness has reached an unprecedented level in both academia and industry. The addition of the large language model makes the search effect a qualitative leap compared with the traditional search. Users are able to obtain relevant information more quickly, improving the efficiency and accuracy of decision-making, especially in complex queries and deep information retrieval scenarios.
However, the introduction of large language models also brings some troubles, especially the problem of phantom rate. In some scenarios that require very high answer accuracy, it becomes difficult to ensure the accuracy of the big model answer, which imposes constraints on enterprises and developers in the landing of AI search services.
2）The efforts and evolution of Alibaba Cloud AI search
With the evolution of the AI era, Alibaba Cloud has made corresponding efforts on the cloud. Alibaba Cloud's search products have historically been divided into two engines: one is the open source engine Elasticsearch, which cooperates with Elastic, and the other is Havenask, a self-developed search engine based on years of experience.
2.1 The evolution of the open source ecosystem:
2017: Alibaba Cloud has a strategic partnership with Elastic and a joint product release. Elasticsearch on Alibaba Cloud is what we are currently using.
2019: We continue to target this fully managed ES product on the cloud, continuously improve the operation and maintenance management capabilities, enhance the intelligent elastic expansion and contraction and monitoring and alarm services, to ensure that customers can obtain stable and efficient services in the process of use, and reduce the cost of development. In addition, Alibaba Cloud has developed some new features based on the ES kernel. For example: support for index construction Indexing service to achieve write acceleration. Through self-developed storage engine Openstore, to help customers in the case of massive data, reduce our storage costs.
2022: Serverless service release, to help small and medium-sized customers and large customers in the case of large amounts of data to reduce costs by 50%, optimize resource allocation.
2023: Alibaba Cloud's products fully enter the field of AI search, starting with version 8.X. The vector search capability has been continuously enhanced. The ability characteristics have also been continuously enhanced.
2.2 Self-research program evolution:
2008: Alibaba Cloud began to self-research an internal open source engine Havenask, is also the history of our support for Taobao, Tmall, including internal double eleven a lot of this more high concurrency, more extreme scenarios of a search engine.
2014: Alibaba Cloud went to explore its commercialization in the cloud one after another. Scenario as the core to help courseware construction, to help customers build scenarios of intelligent search, to provide some industry templates and personalized programs. As well as in e-commerce, content education, games and other industries to do some personalized programs and effect enhancement.
2023: After entering the era of large language models, Havenask is also the first one-stop intelligent Q&A RAG product in China, as well as multimodal RAG products. Until today, Alibaba Cloud has been continuously optimizing the capabilities of RAG and AI search based on the accumulation of internal depth.
2.3 The core focus of search products:
From the past to the present, and even into the future, the core concerns of search products still revolve around three main areas of optimization:
Cost:This may also be a very personal concern of many customers. Because in the era of large language models, we have a lot of this resource is mainly based on the GPU to complete the GPU itself is relatively expensive, so how to help customers reduce costs, is the subsequent evolution of the entire product is an important direction.
Performance:This is mainly two points, one is the massive AI data writing and processing speed. One is the response speed of online queries. These two speeds should realize the performance of good experience of such a requirement, but also Alibaba Cloud products to help you solve a problem.
Effect:Alibaba Cloud and many customers have done online or offline exchanges. Customers have some scenarios where they want the interactive ability of AI, but also want to ensure that the results are 100 percent accurate, and they can't introduce any illusion of a large model, and they can't let the large model do any fabrication. We are introducing the search link, and under the guarantee of accuracy, we are able to apply to the AI's ability to carry out a complete interaction. As for the specific effect optimization is as follows:
3）Alibaba Cloud AI Search Product Introduction
Alibaba Cloud AI Search Open Platform provides five scenario-based products, including LLM Intelligent Q&A Edition, Log Retrieval Serverless Edition, Industry Algorithm Edition, Vector Retrieval Edition, and Graph Retrieval Edition. The bottom layer is based on a very large number of open source search engines, engines including the enterprise version of Elastcsearch, Alibaba's own engine Havenask and Milvus, etc., to facilitate support and docking. Through these products, users are able to realize an end-to-end search solution to quickly obtain the information they need.
Currently, Alibaba Cloud AI Search provides customers with all the AI search capabilities as shown in the figure below:
3.1 Alibaba Cloud AI Search Open Platform Product Overview
AI Search Open Platform, as a flagship product of Alibaba Cloud search team, is a precipitated fruit of more than 20 years of experience, which includes almost all services that can be used in all aspects of the current Alibaba Cloud AI search field, aiming to provide users with powerful search capabilities.
For customers who are familiar with development, using the platform API calls can achieve higher flexibility and speed. And for customers who want to deliver services quickly, Alibaba Cloud provides scenario-based products based on feedback. These products are categorized and developed for different business scenarios, and there is a corresponding product version for each scenario. Users only need to write data into the system to directly access the relevant results, and there is almost no need to write code.
3.2 Introduction to Alibaba Cloud ES version 8.15 features and application scenarios
The latest version 8.15 of Alibaba Cloud ES is a vector-enhanced version based on the latest kernel. This version improves performance by more than five times compared to version 8.9, supports data quantization, and can significantly reduce memory storage costs by 75%. In addition, this version natively supports vector search without the need for plug-ins, and supports the multi-way fusion sorting algorithm, which improves the fusion effect of vector search and traditional search results. This version also supports seamless integration with the AI search open platform.
In terms of search scenarios (RAG), Alibaba Cloud's ES can be widely used in the following scenarios:
Intelligent Customer Service: Through natural language processing, intelligent customer service can quickly respond to customer queries and provide accurate information. For example, a customer can ask about refund policy or express delivery status, and the system can instantly give relevant answers to improve customer satisfaction.
Big Data Retriveal: Enterprises can use AI search technology to build an internal knowledge base to help employees quickly find the information they need. This approach not only improves work efficiency, but also promotes knowledge sharing.
Vector Search: in e-commerce platforms, users can get personalized product recommendations through AI search. The system is able to provide accurate product suggestions based on the user's historical behavior and preferences, increasing the conversion rate.
Smart Q&A: In finance and other industries, AI search technology can handle complex queries, users can ask questions such as “how to apply for desktop computer”, the system will be based on historical data in the form of a table to output the relevant results. This ability allows users to compare and analyze data more intuitively, supporting more accurate investment decisions.
4）Alibaba Cloud AI Search Program
4.1 A full-link solution to build a RAG system based on Alibaba Cloud Elastisearch
This scheme shows how to build a full-link RAG (Retrieval-Augmented Generation) system by utilizing the modeling services provided by Alibaba Cloud Elasticsearch and AI search open platform.
First, customer data sources (e.g., PDF, Word documents) are imported into the system, and the information is recognized and extracted by document parsing services, followed by slicing and vectorization of the documents, and ultimately these vector data are stored in Alibaba Cloud ES for index construction.
Offline data writing mainly involves customers importing documents (e.g., word, PPT, PDF) into the AI search open platform and recognizing and extracting them by calling the document parsing service. For long documents, the system will be cut, such as the use of subheadings or subheadings and other structures for semantic or document structure cut. Then, the sliced text is transformed into dense vectors or sparse vectors and stored in Alibaba Cloud ES to construct indexes, completing the construction of original text indexes and dense vector and sparse data indexes.
When querying online, the user inputs a question, the internal service understands and expands the customer's question, then the query understanding service is introduced to determine the intent of the question and generate multiple possible question variants. Then, the problem is transformed into vector data, indexes are constructed, and multiplexed recall is performed to return Top N knowledge fragments. Subsequently, after rearranging the model, it is integrated into the larger model in a Prompt engineering manner for aggregation processing to form the final answer to the customer.
In addition, with version 8.15 of Alibaba Cloud ES, AI semantic search models can be created using the Influence API. The demonstration includes steps such as selecting a generic commercial version, configuring visualization control, setting up access whitelisting, selecting a model service, obtaining configuration information, and creating and debugging the model in ES. This process demonstrates how to seamlessly integrate Alibaba Cloud AI model services, reduce the threshold of model usage, and realize functions such as semantic search.
4.2 Effectiveness Evaluation and Optimization
To ensure the effectiveness of the AI search solution, Alibaba Cloud also provides a full-link effect evaluation service. The platform can help customers evaluate the relevance, illusion rate and credibility of answer results in a three-dimensional way. This evaluation mechanism not only supports the effect test of a single service, but also supports the synchronized evaluation of multiple services, which improves the decision-making efficiency of developers.
Through these evaluations, enterprises are able to identify problems and optimize in time to ensure the stability and accuracy of AI search services, thus improving user satisfaction.
Closing
Alibaba Cloud AI search solution version 8.15 adds the AI service center column and model management function on the original basis, supports one-click creation of AI open platform space and API, and maintains support for all the functions of version 8.13. 8.15 kernel also emphasizes its vector capability and AI search capability, and also supports basic application scenarios such as logging, which is suitable for customers with the need to reduce costs.
In terms of billing, it flexibly supports pay-per-volume or annual and monthly packages, and the AI platform is billed on a per-call basis, offering the first 100 free calls.
In addition, Alibaba Cloud has launched major promotional activities to help enterprises develop.
Alibaba Cloud Elacticsearch version 8.15 Full specification 50% off on new purchases of annual payment pass-through!
Alibaba Cloud Elacticsearch non-8.15 version, 2-core/4-core full specification new purchase annual payment pass through 50% off!
Top comments (0)