In the past year, the rapid development of generative artificial intelligence (AIGC) technology and the emergence of various large-scale models have triggered global discussions on whether the era of general artificial intelligence (AGI) is coming. After AIGC large model public services are gradually accepted by the public dialectically, how to use AIGC technology to reshape enterprise intelligent services has become a deep water area.
Now, almost all companies are trying to implement AIGC technology in their own fields, but they will face various challenges and difficulties during the implementation process. However, we clearly see that the new artificial intelligence (AI) technology has changed the original link of enterprises to empower business through AI atomization capabilities. Based on the unified enterprise AI data infrastructure, the generalized intelligence accumulated in the large model is deeply integrated with the precise enterprise knowledge, and then condensed into specific scenario-based services, leading the enterprise into a truly comprehensive intelligent era.
In 2006, Geoffrey Hinton and others published the paper "Reducing the Dimensionality of Data with Neural Networks", which opened the era of "deep learning". Up to now, after nearly 20 years, more and more enterprises have built a relatively complete AI application development and operation and maintenance system. This system is usually divided into three layers: the bottom layer is the machine learning platform, the middle layer is the AI service, The top layer is the enterprise application based on AI service.
There are many implementation problems in AI in the era of deep learning: specifically, on the data side, due to the poor generalization ability of traditional small models, it is difficult to directly combine existing models with more complete enterprise private data to provide external services, and enterprise private data is in AI-oriented applications have not achieved connectivity and linking; in terms of AI service ecology, enterprises expect to be able to directly reuse existing AI services to quickly build applications, but in reality only a small number of services (such as OCR, speech recognition, etc.) are implemented In order to achieve high reusability, enterprises still need to continuously develop new algorithms and models for their own data and business scenarios, and the efficiency of AI application implementation needs to be improved.
With the rapid development of AIGC technology centered on large-scale model technology, the iPhone in the AI era has officially arrived, and AI is moving from the era of deep learning to the era of large-scale models. The brand-new technology paradigm is redefining the implementation method of enterprise AI applications, accelerating the comprehensive intelligent upgrade of enterprises, and will also bring new changes in the development and operation and maintenance of traditional AI applications.
Large models can be combined with more enterprise data for intelligent application development, not limited to very limited data for AI intelligence; the generalization capability of large models can enable a large model to handle multiple downstream AI tasks, saving model development time and the operation and maintenance costs of multiple models; at the same time, large models with tens of billions of parameters have generalization capabilities and intelligent emergence capabilities, and the model effect is greatly improved compared with traditional deep learning models. These large-scale model advantages also sounded the clarion call for AIGC's implementation in enterprises.
Paradigm shift of AI application landing from deep learning era to large model era
Due to its technical advantages, large-scale models have been gradually implemented by enterprises in combination with enterprise data. In practice, there are also many challenges:
-
It is difficult and costly to manage massive amounts of data in enterprises. How to make the rich data of enterprises ready for the implementation of AI?
-
Combining vector database and enterprise data, there are problems such as hallucinations and inexplicability in the process of landing large models. How to meet the needs of enterprises for precise knowledge and interpretability?
-
How to achieve faster model inference under limited resources?
-
How to open up the entire enterprise intelligence value chain from enterprise data to model services?
In order to meet the above challenges and realize the rapid implementation of AI applications, enterprises need to establish AI infrastructure for the era of large models, reduce the illusion of large models, optimize the use of GPU resources, and improve service reasoning capabilities; intelligent management and inventory to provide high-quality data input for large models; user-friendly engineering links need to be provided to open up the value link from enterprise data, model services to intelligent applications.
3.1 A unified AI infrastructure is more important than ever
Thanks to the generalization and multimodal capabilities of large models, in the future, except for a few companies using their own new training and fine tuning (fine tune) large models, most companies will be more based on large models (general large models or industry Large model), combined with a small number of small models for rapid application landing. In this context, enterprises are paying more attention to private data, more economical computing power, and out-of-the-box models than ever before.
In the case of large models, enterprises also need to solve the problems of large model hallucinations, long-term memory and reasoning of large models, and at the same time improve the effect of service reasoning based on limited GPU resources. Therefore, AI infrastructure for unified data, computing power, and models for large models will become a standard requirement for future enterprises in the era of large models.
3.2 AI is the final export of data, and high-quality data can help enterprises create high-quality local models and AI applications
Currently, business intelligence (BI) is still one of the important exports of data. With the rise of large model technology, the integration of BI and AI will be accelerated, and AI will become the ultimate value export of data. From the perspective of BI's data usage history, high-quality data reports depend on high-quality, cleaned structured data. Similar to BI, high-quality multimodal data will also be an important factor affecting the quality of AI models and applications.
Enterprises will shift from regulatory-based data governance to business-driven intelligent data asset inventory, providing high-quality data input for large models and business scenarios. However, it is difficult for the traditional data governance platform that focuses on manual implementation to complete these tasks efficiently. Therefore, driven by the capabilities of large models, the construction of a unified intelligent data asset platform will effectively guarantee high-quality enterprise data, thereby accelerating the implementation of AI applications in enterprises in the era of large models Ability.
3.3 A low-code platform that connects data and model factories can really help enterprises quickly implement AI applications
Before the era of large models, due to the limited generalization ability of AI models, different AI applications needed to establish new AI models separately in many cases, which led to the development of new AI applications relying on professional algorithm engineers and senior developers. In the era of large models, enterprise personnel can pay more attention to their private data and choose a suitable large model (or a combination of large and small models), and link the model with local high-quality data through prompt engineering or model fine-tuning together.
Furthermore, the large model can understand business scenarios, and weave different models according to customer data and business scenarios to jointly achieve business goals. Enterprise personnel will be more inclined to use low-code methods to associate their data and models, so as to realize the rapid implementation of AI capabilities in the business. By "weaving" the multi-modal data of the enterprise and utilizing the capabilities of the continuously upgraded AI model, the enterprise service can be reshaped, and new business value can be generated in the end.
4.1 Fabarta Product Matrix
Based on the understanding of the AI application landing paradigm in the era of large models, Fabarta combined the business pain points and needs of many large-scale top financial and manufacturing customers currently serving, and after a certain product polishing, he proposed a product matrix of "one body with two wings" . This product matrix aims to realize the integration of data, computing power, and models in the era of large models, build infrastructure in the era of large models, and help enterprises quickly build AI applications in the era of large models.
Fabarta "One Body Two Wings" product matrix
"Integration" refers to the unified construction of the enterprise's infrastructure in the large model, to realize the management of data, computing power, and model runtime, and to realize AI-ready infrastructure (Infrastructure Ready for AI) from the basic level, which integrates graph and The multi-modal intelligent engine of vector computing can not only realize the long-term memory in the process of model reasoning, but also optimize the reasoning framework of the model so that it has better logical reasoning ability; the "two wings" refer to data and AI respectively, where data The side provides comprehensive multi-modal data management functions, effectively helping enterprises to manage and inventory data assets. These data can be stored in the "one" as the private data of the enterprise, providing high-quality data for AI, and realizing AI-ready data (Data Ready for AI) from the data side; the AI side has opened up the integration of model factory capabilities and private knowledge of enterprises capabilities, and adopts a low-code approach to help enterprises quickly use large models, and implement AI applications based on fine-tuning or prompt mode combined with enterprise private knowledge (AI Ready for Apps).
4.2 Fabarta ArcNeural Multimodal Intelligence Engine
In the AI era, a variety of multi-modal databases have been born. Many traditional databases can also support different forms of data storage through expansion. However, the essence of these multi-modal databases is still to realize the storage and unified access of multiple data. . In the era of large models, what Fabarta has always thought about is, in addition to supporting multi-modal data storage and unified access, what else can be done to support large models?
-
Data challenge: How can we help enterprises build a unified private multi-modal data layer and combine this private data with large models well?
-
Computing power challenge: If an enterprise privatizes and deploys large models, how can it support higher concurrency with limited computing power?
-
Model challenge: the reasoning ability of large models is limited, how to help improve the reasoning ability of large models? Enterprises have a very high deterministic demand for generative answers from large models. How to effectively reduce the problem of random generation of large models? How to help enterprises achieve explainable intelligence?
ArcNeural is an intelligent engine built with Data-Centric AI as the core for processing symbolic data graphs (Graph) and vectors (Vector). AI applications provide private memory and accurate and interpretable reasoning. ArcNeural is an infrastructure built on the three elements of AI data, computing power, and models, providing support for upper-level AI intelligent applications and accelerating the process of business intelligence innovation.
For example, in the knowledge base intelligent question answering system, first import all enterprise data (original raw data, such as CRM, ERP data, product manuals, etc.) into ArcNeural, and the engine will automatically model and generate symbolic data (Embedding&Graphing). When users ask questions, ArcNeural uses interpretable symbolic computing (graph computing) and vector computing to analyze problems, find relevant high-value data, and provide an optimized runtime environment to support large models for content generation, induction, and summary. This not only ensures the accuracy, real-time and privacy of the answer, but also effectively avoids the "nonsense" of the large model, and provides users with intelligent and friendly services. At the same time, the flexible and scalable engine architecture also supports the scenario application of independent graph databases, graph computing, and enterprise-level vector databases, flexibly responding to the infrastructure needs of enterprises in terms of business intelligence.
Architecture Diagram of Fabarta ArcNeural Multimodal Intelligent Engine
4.3 Fabarta Multimodal Data Weaving Platform
The application of large-scale model technology needs to get through the massive private data of enterprises. However, the current enterprise data has many types, large quantities, and mixed quality. How to sort out "private data" for AI to use? This presents new challenges for data management:
-
Data form upgrade: From focusing on core business reports to comprehensively covering original raw data, such as CRM data, ERP data, product manuals, rules and regulations, pictures, video files, etc., not just from structured data to unstructured data Upgrading requires new technical methods for data connection and capture, as well as a more universal metadata management method;
-
Governance goal upgrade: The goal of governance is upgraded from the "DAMA six characteristics" (Completeness, Uniqueness, Timeliness, Validity, Accuracy, and Consistency) of data to understanding of semantics and Extraction of implicit relations;
-
Data service upgrade: The service object of data governance has been upgraded from BI to AI and then evolved into a large model. Therefore, the form of data service has also been upgraded from traditional two-dimensional tables to knowledge services suitable for large model ecology.
Fabarta's multi-modal data weaving platform is an AI-oriented data management platform. It connects, understands, and manages data more intelligently, transforms enterprise data into enterprise knowledge, and provides data driving force for the application of AI. It is also compatible with traditional data Governance scenarios. Based on the ArcNeural intelligent engine, the platform connects enterprise private data, automatically acquires and analyzes the metadata and data semantics, forms a data lineage and asset map, and provides intelligent data standard implementation, data quality analysis, and index link optimization on this basis , data classification and cataloging, and other functions to provide data services for business applications and large models. Its core modules include:
-
Graph-enhanced data governance: collect and identify metadata information of multi-mode data including structured databases, documents, and pictures, analyze data lineage through original technical information such as data processing scripts and system access logs, and comprehensively control data standards and data Quality, data security, and compatibility with the existing data governance platform of the enterprise;
-
Intelligent data asset inventory: use accurate metadata, kinship and other information to screen and classify massive enterprise data, and through the assistance of intelligent technology, understand the data content, extract the hidden data relationship, and restore the real data model (Data Model);
-
Multi-modal data service: Through index modeling, data virtualization, knowledge service and other technologies, it adapts to traditional BI, AI and large model scenarios at the same time, provides comprehensive data services, and can also be connected to the Fabarta enterprise intelligent service platform to quickly implement AI application.
Architecture Diagram of Fabarta Multimodal Data Weaving Platform
As the "data wing" in Fabarta's product matrix, the Fabarta multimodal data weaving platform makes full use of large model capabilities to realize intelligent data management, connects with the existing big data platform of the enterprise, sorts out and manages the massive multimodal data of the enterprise, and constructs data Asset maps, and provide an intelligent data foundation for the implementation of AI large models, and provide data ready for AI implementation (Data Reay for AI).
4.4 Fabarta Enterprise Intelligent Analysis Platform
In the past few decades, the main users of the AI platform were professional algorithm engineers and senior developers, and the platform provided them with complete AI development engineering links, mature algorithms, and efficient training and reasoning frameworks. The large model has attracted the attention of various industries with its excellent generalization reasoning ability, especially how to enable business users and engineers in the enterprise to use the accumulated private data in combination with the large model at a low threshold, directly Generate value for the business. From being oriented to specialized and sophisticated AI developers to implementing large-scale model capabilities oriented to enterprise business users, the challenges of this transformation are what the enterprise intelligence platform of this era needs to face:
-
Data preparation for the era of large models: how to better segment and reorganize enterprise multi-modal data into a data storage form suitable for large models in AI scenarios, and select fine-tuning or prompt modes in combination with business scenarios to reduce large models The hallucination problem to ensure enterprise-level landing;
-
Model preparation for the large-scale model era: How can enterprises choose a basic model that fits their own business scenarios as the enterprise intelligence base among the mushrooming open-source and closed-source large-scale model projects;
-
Business empowerment in the era of large models: How can enterprises help business personnel focus on their own business data, combine the respective advantages of large and small models, directly generate AI applications in a drag-and-drop manner, and complete inclusive large-scale model scenarios on a self-service basis landing.
Architecture Diagram of Fabarta Enterprise Intelligent Analysis Platform
As the "AI wing" in Fabarta's product matrix, Fabarta's enterprise intelligence analysis platform aims to help AI developers, business users, and application developers in enterprises to quickly implement in a low-code and no-code (Low-Code No-Code LCNC) manner AI capabilities. Its characteristics are as follows:
-
Acceleration of large-scale model implementation: Help enterprises select the optimal large-scale model based on their own business and data, support self-owned data and mainstream large-scale models for fine-tuning, and directly use large-scale models and local knowledge bases to build knowledge services;
-
Support the evolution of large-scale model applications: Enterprise-level AI applications rely more on specific large models and penetrate into all aspects of business by combining multiple small models. The platform fully supports large-scale modelization of enterprise-level AI applications through model factories and model orchestration;
-
Full coverage of enterprise smart user ecology: Accelerate the whole process of data from development to business through the three-tier architecture of data and AI development, business analysis, service and application. Help data generate business value for industries and scenarios;
-
Series connection of LCNC analysis capabilities: LCNC applications are not only drag-and-drop of front-end components, but also help business users focus more on data and business itself by pre-packaging complex business analysis logic, and directly produce business applications through platform capabilities. Comprehensively improve the inclusiveness of AI;
-
Precipitation and reuse of industry capabilities: Help enterprises to deposit their own industry-specific business knowledge and technical atomic capabilities, and use platform capabilities to support the reusability of their business scenarios when they expand horizontally, and accelerate the implementation of industry capabilities;
-
Interpretable graph intelligence: Use the natural interpretability of graph data, combined with graph computing algorithms, decision engines, analysis canvas, graph BI, data exploration and other capabilities to help end business users know what is happening and why.
Using its core products, Fabarta has helped several leading enterprise customers to carry out intelligent transformation:
-
Using the data lineage link tracking technology empowered by graph and large model technology, it helps a leading city merchant to fully trace the data lineage and improve data insight, so as to ensure that decisions are made based on reliable data; through in-depth analysis of data lineage, data can be quickly located Quality issues, realize the exploration of data, and lay a solid foundation for building data weaving;
-
Using the enterprise intelligent analysis platform, combined with the graph technology and graph algorithm in the multi-mode intelligent engine, helped a leading commercial bank to quickly develop post-loan risk control applications, and realize early warning configuration, risk evaluation, risk investigation, customer view, etc. in risk management function, reduce the cost of risk monitoring, and achieve one-stop management of risk monitoring;
-
Using the multi-model intelligent engine that integrates vectors and graphs, it helps a leading manufacturing company to quickly connect to the internal knowledge base, build an intelligent question-and-answer system, realize the empowerment of large models to enterprise data, and fully tap and exert the value of enterprise data.
Towards a new paradigm in the era of large models, Fabarta manages symbolic data in the era of large models, mainly vectors and graphs, through a multi-mode intelligent engine, provides computing power and model acceleration support, and serves as the AI infrastructure of the new era; at the same time, through data weaving The platform realizes the exploration, intelligent inventory and use of data, and provides high-quality enterprise data for large models; uses the enterprise intelligent analysis platform to help enterprises quickly connect to local data, and uses large models to empower business applications.
Adhering to the concept of "connecting the world, seeing the future with wisdom", Fabarta is committed to building AI infrastructure in the era of large models, and working with partners and customers to create intelligent enterprises in the era of large models.
If you are interested in our products or technologies, please contact us at [email protected].
Fabarta is an AI infrastructure company that helps enterprises achieve continuous intelligence-driven innovation by exploring and connecting data resources. At present, Fabarta's product system is divided into three layers: at the engine layer, it builds an AI-oriented technical infrastructure and provides an ArcNeural multi-modal intelligent engine that supports the integration of graphs, vectors, and AI reasoning capabilities; at the platform layer, it accelerates through the enterprise intelligent analysis platform Interpretable graph intelligence and the new generation of AI technology are implemented in enterprise scenarios. At the same time, the multi-modal data weaving platform is used to help enterprises sort out multi-modal data assets, allowing enterprises to give full play to the value brought by data flow; in addition, Fabarta can be based on multi-modal The modal intelligence engine, enterprise intelligent analysis platform, and multi-modal data weaving platform work with customers and partners to build industry applications and accelerate the digital transformation of enterprises and the implementation of AI technology.