On July 14, the Kyligence User Conference with the theme of "Unleashing Digital Intelligence Productivity" was successfully held in Shanghai. The latest Kyligence product family was released at the conference: the preview version of Kyligence Copilot, an AI digital assistant , the Cloud and Enterprise versions of Kyligence Zen, a one-stop indicator platform , and the latest version of Kyligence Enterprise, an enterprise-level OLAP platform .
Li Yang, co-founder and CTO of Kyligence, delivered a keynote speech on "Indicator Platform + AI Technology Implementation and Future Outlook" , which received enthusiastic responses from the audience. Here is a special summary of Li Yang's speech on the day The transcript of the speech is for reference. At present, Kyligence Copilot has also opened the application test, and you can visit the application channel at the end of the article.
The following is the transcript of Li Yang's speech that day
The Live Demo demonstrated by Luke just now is very exciting. Next, let me share how Kyligence implements the indicator platform + AI technology, and our views on some key technologies in the future.
Number of people used = AI Copilot + index system + reasonable cost
Let's start with the problem, the highest ideal of "everyone uses numbers". First of all, I will share why Copilot + indicator system + reasonable cost is a relatively perfect combination that can achieve the goal of "everyone uses numbers"; then I will share several typical landing scenarios, including SaaS mode, local deployment and embedded deployment; and finally share the future outlook on these key technologies.
The indicator system gives us a common data language . When each of us uses data to communicate, the first obstacle we encounter must be the lack of a common language. Just as Mandarin allows 1.3 billion Chinese to communicate freely, it is also very important to have a standard and consistent caliber for data interpretation, which is the premise of enterprise data sharing and collaboration.
Although the AI assistant is the latest to arrive, it is like a bottle cap and is the most critical part of the puzzle of "numbers for everyone". Kyligence Copilot provides a zero-threshold data tool , which is simpler than the most widely used Excel, and users can solve the last mile of data tool connectivity by typing a single line.
Finally, it is a reasonable cost. We see that AI is very good, but everyone is looking for its landing scene. Why are you still looking? In fact, it is a question of cost. We can foresee the cost of investing in AI, so we must let AI empower business scenarios and bring greater value and returns .
Combining the three key parts of AI assistant, index system, and reasonable cost, we can see a complete puzzle. A zero-threshold data tool with a unified data language can communicate efficiently, and the final cost is economically controllable, creating real value for the enterprise business. By combining Kyligence Copilot, Metrics Store and high-performance OLAP engine, we can see a complete puzzle, and a new era for everyone to use numbers has opened.
How did these technologies land? Everyone knows that Kyligence is a technology-based company, and the core positioning of our products is "indicator platform and engine". In addition to this core, we will expand some indicator capabilities, such as attribution analysis, Copilot capabilities for natural language docking, and target dashboards.
The two circles in the middle are Kyligence's product line and the foundation of our foothold. The outer circle is to support more vertical industries, such as finance, retail, manufacturing, medicine, etc. We also hope to co-create with partners in more industries, so that Kyligence products can be implemented in more industry scenarios.
On the right is a large language model. Kyligence will not make a general large language model, but we will make a domain language model on the basis of a general large model. Just like students will receive general education first, and then go to in-depth study of a certain professional field, Kyligence Copilot is based on the large language model, becoming a professional digital assistant for indicator analysis and data analysis , just like in the Live Demo. It can objectively help enterprises to do data analysis, and assist enterprises in business decision-making and operation management.
Landing plan
SaaS｜Localized deployment｜Embedded deployment
After sharing the positioning of Kyligence, let's take a look at the implementation plan of Kyligence indicator platform + AI.
- SaaS
What the Live Demo showed just now is actually the SaaS version. The core of this product is the Kyligence indicator platform and engine. Kyligence Enterprise is our OLAP engine. Kyligence Zen is the indicator platform product we launched today. We can see some surrounding the platform. Capabilities such as attribution analysis, goal board, dashboard and Copilot that everyone is paying more attention to today.
When talking about SaaS, we have to talk about data security. How does Kyligence ensure that users can trust us with their data?
- First of all, the entire Kyligence Zen production domain is completely isolated from the Kyligence office domain , including myself without permission to access the production environment.
- Data storage and computing resources are isolated by organization . For example, if different companies apply for registration, the resources of each applied organization are completely isolated from other organizations.
- The data is encrypted throughout the entire process , whether it is data storage or transmission, there is encryption control throughout the process, and only customers and Zen applications can access and process data.
- All access to the production environment has a process to ensure audit and trace review , and we have SOC 2 Type 2 and ISO certification to ensure compliance and safety and reliability.
The SaaS version is the easiest way to experience the new generation of data analysis capabilities of Kyligence Zen. Welcome to apply for a trial of Kyligence Zen at the end of this article.
- Localized Deployment Solution
For enterprises, data security is invaluable . Represented by the financial industry, many enterprises also hope to have a local deployment solution. Let's first talk about how data can not leave the local environment of the enterprise? In a real scenario, the simplest idea is to move the large language model to the private domain of the enterprise, but in fact, there may be a problem when it is implemented at this stage: ChatGPT represents the highest level in the industry. When using an alternative model, will it be weaker than ChatGPT?
This situation may indeed happen, so how can we guarantee language understanding and reliable performance under weaker premise? We can disassemble this action of Copilot into several language behaviors. In fact, we may not need such a powerful and perfect general language ability for each partial language behavior.
That is to say, when we narrow the scope, we can appropriately reduce the requirements for the general model. To give a specific example, when Kyligence Copilot executes an instruction, it actually divides into three steps:
- The first step is to review the question to see whether the question asked is legal, compliant, logical, etc. This is the first dialogue with the language model.
- The second step is instruction understanding , which is to map user requests into specific actions of an indicator platform. This step is usually difficult because users express various business needs.
- The third step is to execute the instructions on the indicator platform , which may be attribution analysis, target Kanban, etc. What is obtained in this step is only data.
- The fourth step is to use a language model to interpret the data and charts in natural language, and to feed back key insights to users.
In the actual implementation, the enterprise can replace a general large model with three local small models, which makes the implementation at this stage a relatively reliable and feasible solution; of course, if the enterprise has a large model that can complete these three tasks It is also possible. The above is a more feasible solution for local deployment of indicator platform + AI.
You can see that the ability of ChatGPT is quite good, how about the performance of other large models in terms of understanding and performance? We have also done some tests and research for you. We have used an open source large language model instead of ChatGPT to make some preliminary attempts. For example, LLaMA 13B and Falcon 40B can probably reach about 70% of ChatGPT 3.5 in terms of test command understanding ability. Capabilities, which can be understood as the minimum capability range that enterprises hope to implement Copilot, a large language model indicator, locally.
There are many suppliers of large language models in China. After our preliminary research, the interface of large language models is relatively unified and common. If the company owns or purchases a large model, Kyligence also quickly cooperates and connects. If you want to implement the capabilities you just saw in your company, we welcome everyone to come and communicate with our frontline personnel, and Kyligence can co-create with you.
- Embedded solution
Luke just mentioned Copilot as a Service, which is the third feasible solution. We can quickly embed Copilot's natural language capabilities into systems developed by enterprises themselves. Kyligence takes the indicator platform as the core, and has some indicator capabilities around it. These capabilities are designed and provided in an open API componentized manner, which can become a part of enterprise applications to directly serve customers, and can also be embedded in our partners. Industry solutions, enterprise data platforms within large enterprises, etc.
We refer to Kyligence Zen's attribution analysis, Copilot, target Kanban and other capabilities as open data products. From a technical point of view, we have a YAML-like open indicator definition language ZenML; at the same time, we have an open API, Web widgets, and Copilot as a Service (ChatUI) are all embeddable controls.
Only ten lines of code are needed to embed capabilities such as Copilot into existing enterprise applications . You can see these operation guides in the Kyligence Zen user manual, which is also applicable to various SaaS applications, industry privatization solutions and internal enterprise data systems.
Future Prospects for Key Technologies
After sharing the main deployment plan, I will further look forward to our views on these key technologies and directions along the indicator system, Copilot and OLAP engine with reasonable cost.
- Index system
Let me talk about the indicator system first. I think it is very interesting that in our practice we gradually discovered that when everyone can use this data, data governance becomes a more difficult problem. We say that order and innovation are actually a pair of contradictions. The most stringent data governance is strictly controlled. There is a standardized indicator system throughout the enterprise. It has the advantage of strict management, but it will also stifle innovation.
Many companies hope to find some reference and reproducible index systems to quickly replicate a proven success, which is an orderly part. With the foundation, the second step is to innovate. Enterprises have their own business characteristics, and they hope to innovate in addition to the industry-based general index system. How to find a balance between order and innovation, we found that the indicator system is a very good tool. Kyligence Zen is a low-code indicator platform, which can help everyone quickly accumulate an indicator system based on business innovation on top of the general indicator system, realize rapid replication and promotion, and raise the baseline of digital intelligence management in the industry.
- AI Copilot
From the perspective of the stability of language model training, the language ability of indicators must belong to a subfield of NLP To SQL. We think that this direction limited to the indicator field will be the first to mature, because SQL, as a query language, has an infinite degree of free space, and the degree of freedom of natural language is also infinite, and stability is bound to have problems.
However, if we narrow the target area of this problem, for example, focusing on the indicators that enterprises care most about, based on these indicators, enterprises will conduct some attribution analysis, cross-time period, and cross-dimensional analysis. So when we limit it to a target domain, the mapping of the entire language model to the instruction training in the indicator platform domain will be much easier. Therefore, Kyligence has considerable confidence in the stability and accuracy of our products after practice.
Finally, we have a set of tools and capabilities for training large language models into domain indicator models on this basis, called the Byzer-LLM toolbox. Starting from the basic model, we can do prompt-tune, plus the user's own data to form The indicator knowledge base is used for Fine-tune, and finally becomes a language model in the available indicator field. We are also continuing to incubate this set of tool chains, and it is now available for initial use.
- reasonable cost
When everyone starts to use data, and when communication barriers are lifted by the indicator standard language, we can expect that the load on the analysis engine may increase by a hundred or even a thousand times . The person in charge of the IT department may start to count the money. How much concurrency is currently supported by the enterprise, and if the resources are doubled, the cost may greatly exceed expectations. This is also the direction that Kyligence has been working on, that is, how to use an ultra-high concurrent OLAP engine technology to support a hundred times the load.
In addition, we continue to improve the performance of the computing engine. The vectorized Spark engine technology we developed, Kyligence Turbo, has more than doubled the speed of the standard Spark, which can help enterprises save about 50% of computing power and costs.
Based on the TPC-H 100 test, Kyligence Turbo takes 55.72% less time than Apache Spark SQL 3.3.1 . This test can be stably reproduced on AWS EC2. Friends who are interested in technology can visit the homepage of Kyligence Turbo, which has open sourced the entire testing process and can reproduce this experiment. We will continue to incubate this new technology, and expect that all scenarios that use Spark today can achieve an immediate 100% performance improvement or 50% cost reduction with the support of the vectorization engine.
In terms of OLAP engine, we should optimize it for cost. Nanjing University mentioned in a paper released last year that enterprises should not only consider performance when evaluating OLAP engines, because after resource elasticity on the cloud, as long as the money is in place, performance can always be achieved. We should look at it from the perspective of cost. A simple verification was done in this paper. The horizontal axis is the number of queries. It can be seen that the cost of the OLAP engine using Kylin is relatively stable. Kyligence can take a hundred times the load, our engine will be very advantageous.
In short, we see that when the three technical points of Copilot + indicator platform + high-concurrency OLAP engine are in place at the same time, a new era of data usage by everyone has opened, and everyone can quickly experience the SaaS version, and can also deploy locally Or embedded mode can be empowered to our application system. Trials of Kyligence Zen and Kyligence Copilot are now open, and you are welcome to click the link to apply for a trial.
About Kyligence
Founded in 2016 by the founding team of Apache Kylin, Kyligence is a leading provider of big data analysis and indicator platforms, providing enterprise-level OLAP (multidimensional analysis) product Kyligence Enterprise and one-stop indicator platform Kyligence Zen for users Provide enterprise-level business analysis capabilities, decision support systems and various data-driven industry solutions.
Kyligence has served many customers in banking, securities, insurance, manufacturing, retail, medical and other industries in China, the United States, Europe and Asia Pacific, including China Construction Bank, Ping An Bank, Shanghai Pudong Development Bank, Bank of Beijing, Bank of Ningbo, Pacific Insurance, China UnionPay, SAIC, Changan Automobile, Starbucks, Anta, Li Ning, AstraZeneca, UBS, MetLife and other world-renowned companies, and reached global partnerships with Microsoft, Amazon Cloud Technology, Huawei, Ernst & Young, Deloitte, etc. Kyligence has received multiple investments from institutions such as Redpoint, Broadband Capital, Shunwei Capital, Eight Roads Capital, Coatue, SPDB International, CICC Capital, Gopher Assets, and Guofang Capital.