作者 | 宋慧出品 | CSDN（ID：CSDNnews）
After the battle of 100 models in 2023, the large model in 2024 has swept the entire industry and become one of the core technologies of the new quality productivity of the digital economy. Now, all parties are more pragmatically exploring how to truly implement large-scale model technology, and take into account performance and cost performance, to serve enterprise scenario applications, in order to improve business efficiency and innovation.
The technologies involved in large models are huge and complex, and opportunities and challenges coexist for the development of AI applications in the industry. According to the latest CSDN AI Developer Ecosystem Report, multiple problems coexist in the development of large model technology, and the proportion of problems is average, such as the lack and difficulty in recruiting AI technical talents, the difficulty in finding suitable business scenarios, the insufficient quantity and quality of AI data and infrastructure computing power, and AI regulatory and legal risks and compliance issues.
The more realistic and direct factors that enterprises need to face are that the enterprise technical team needs to choose to accurately meet the needs of industry-specific business scenarios in terms of the scale, accuracy, and generalization performance of large model parameters, and how to quickly integrate large model capabilities to improve the AI capabilities of enterprise businesses and products.
Although there are many difficulties, the practice and innovative application of the large model industry are emerging. Recently, the innovative education platform "Chaos Academy" has launched a new AI function - AI Q&A Assistant, which adds AI Q&A capabilities for course content to the one-way transmission of knowledge in traditional education.
The Chaos Academy AI Q&A Assistant uses the technical capabilities of the Orion Star Model. Zeng Delong, CTO of Chaos Academy, said: "In terms of the accuracy of the AI Q&A assistant, the Orion Star Model RAG can provide fast and accurate answers when dealing with frequently asked questions about the course. ”
"In terms of the convenience of organizing and adjusting the knowledge base, the background UI of the Orion Star Model RAG Kit is simple and convenient, and it is very efficient in terms of information retrieval and content organization. Organizing our knowledge base in the form of cloud documents allows us to modify and adjust the content of our knowledge base more quickly. ”
"AI Tutor" is launched,
Why education companies can quickly acquire large-scale application development capabilities
In the landing scenarios of various industries of large models, education is a typical knowledge-intensive and knowledge-intensive scenario, where users (students, trainees) have problems, and the results obtained from online searches are usually only fixed standardized knowledge of "one face", and there is still a high cost of learning and understanding for learners. The large-scale AI Q&A function of Chaos Academy realizes more flexible and rich reply content, and provides more customized course services based on student user portraits.
For students and users of the Chaos platform, they can get customized course recommendations (what to learn) more accurately before class, use AI Q&A interaction in class to obtain learning and Q&A content more efficiently (learn thoroughly), and use AI Q&A interaction after class to promote the combination of learning and practice and quickly retrieve the learned knowledge base (flexible use + search and review at any time).
This new feature sounds more like a learner in addition to active learning and homework completion, a dedicated "AI learning tutor" is arranged close to him, this kind of "thousands of people, thousands of faces" customized service, not only improves the learning efficiency of chaotic users, the original intention of completing education, but also an innovative application of AI large models in the education industry.
AI Q&A has been launched on a number of well-known courses on the Chaos platform.
The measured response speed is timely and the generated content is accurate
In order to ensure the accuracy and effect of AI Q&A, in addition to using large models, the chaos technology team still has some manual participation in the review of professional course content. The first batch of online courses selected 17 classes with fresh content and high popularity, and in the future, the AI Academy and all other new courses that will be launched by Chaos Academy will be equipped with AI Q&A as a standard function.
In addition to AI Q&A, for Chaos' internal R&D team, large models also promote the dual benefits of innovation and efficiency in the development of other new applications. For example, in the just-concluded chaotic 5,000-person cruise ship offline event, an AI assistant (based on a large model) was introduced in the event sales stage, and the team imported the cruise ship corpus into the platform to provide users with cruise ship assistant services. As a result, 15% of the learners used the AI assistant, with about 3-8 session rounds per user, which not only improved the user experience of the event, but also saved a lot of communication labor costs for the event organization.
According to the Chaos Technology Team, it took less than 2 months to officially launch the R&D of AI Q&A after investigating multiple large models, which makes us wonder how a company focusing on education can quickly obtain the ability to develop large model applications and quickly implement them when large model developers are facing multiple R&D challenges.
Orion Star Large Model RAG Kit + "Turnkey" Customization
Accelerate the application of large models in the industry
According to reports, the AI Q&A of Chaos Academy adopts the popular Orion Star large model technology and service in China this year. Different from the industry's common RAG technical services, Orion Star Model provides a stronger and more professional RAG suite, including knowledge preprocessing, semantic understanding and rephrasing, hybrid search engine, automatic quality evaluation system and fine-tuning effect calibration.
Chaos Academy uses the RAG suite capabilities of the Orion Star Large Model, which involves data preprocessing, identification of user problems, rewriting and expansion of user problems, retrieval and ranking of vector knowledge bases, interaction and boundary control with large models, as well as session management and user session context data management in engineering.
RAG is a hot technology point for large models this year,Chaos Academy at the beginning of the AI large model landing exploration,The correct rate of course analysis and answer after using other large models in the industry to train and reason,52.88%,Even if the accuracy rate is increased to 80.77% after processing through RAG,However, the application of industry models has high requirements for professionalism,There is still a 20% fault tolerance rate that cannot be tolerated。
The Orion Star Large Model RAG Kit industry application and professional knowledge are deeply integrated, based on the Chaos Academy business to build applications and optimize the effect, after obtaining high-quality data and fine-tuning, the accuracy of course questions and answers is increased to 94.5%, ensuring the accuracy and professionalism of the industry model application. (After this interview, Chaos Academy continues to optimize data quality using the RAG suite, and the latest Q&A accuracy rate has increased to 98.7%. ）
In addition to RAG, there are also general dialogues, plug-ins, long tokens, knowledge extraction, generation of Q&A pairs, and fine-tuning in Japanese and Korean.
In this case, the Orion Star Large Model RAG Suite provides all technical processing capabilities for AI Q&A except data preprocessing, while the Chaos technical team only needs to do the preprocessing of its own course data and a small amount of API docking work. The backend UI of the Orion Star Model RAG Kit is simple and convenient, and it is very efficient in information retrieval and content organization. The way the knowledge base is organized in the form of cloud documents also allows the Chaos team to modify and adjust the content of the knowledge base more quickly. This is also an efficient way for an educational enterprise to quickly obtain the application development capability of AI large models.
Under the general trend of digital intelligence, the enterprise R&D team represented by the Chaos Academy needs to improve new technical capabilities such as AI large models, so in addition to API docking and using general services, Orion Star provides a complete turnkey service from the research and evaluation of the large model to the final delivery of the Chaos Academy, which runs through the enterprise from the preliminary evaluation, LLM strategy, to its own model selection, solution design, implementation, deployment and optimization, as well as delivery and maintenance, complete personal technical capabilities, enabling enterprise large model application development.
Through turnkey services, enterprises can obtain more flexible application customization development, and further explore and utilize large models to achieve accurate and efficient applications. For example, in the future, Chaos Academy will also enhance the intervention capabilities of all RAG technologies, such as enterprise design prompts and related orchestration, and Rank customized development based on its own deep understanding of business.
Behind the rapid implementation of enterprise model applications
The pragmatic path of the Orion Star Model
It is understood that the Chaos Academy AI Q&A uses the hot Orion Star model in the field of enterprise application models this year, and the Orion Star model is the latest achievement of Fu Sheng, chairman and CEO of Cheetah Mobile, after All in AI in 2016. After several years of accumulation, Orion Star has self-developed full-chain AI technology, such as full-chain voice technology (Orion Star is the technology provider behind Xiaomi Xiaoai, Huawei AI speakers, Himalayas, Midea and other speaker products), multi-sensor fusion navigation, consumer-grade robotic arms and physical service robot products.
In the current explosive battle of 100 models, the Orion Star Model has taken a different pragmatic path. At the beginning of the year, the Orion Star model shouted the slogan of "born for enterprise applications" and created the most cost-effective large model. Unlike other large models that engage in an arms race in chips and pursue the effect of hundreds of billions of models and then look for applicable scenarios, Orion Star works backwards from the scene, excavates enterprise applications, defines vertical scenarios, and practices surpassing GPT-4 in application effects through the polishing of (enterprise private) data and applications, which is by no means delusional.
After focusing on the micro-research and development work in seven application directions, the model accuracy and inference speed are excellent, the model is smaller and more conducive to private deployment, and the consumer-grade graphics card can also achieve the effect of 100 billion parameter models, which can be said to effectively reduce the cost of enterprise deployment of large models, thereby accelerating the implementation of large model industry enterprises.