Beijing Bank deeply practices the concept of “One Bank, One Data, One Platform,” focusing on the “dual customer” experience and aiming for value creation. It extensively applies AIGC technology to accelerate the construction of a knowledge-driven large model application system.
In 2023, the AIGC and large model industries experienced explosive growth, driving artificial intelligence to penetrate various sectors, triggering a revolution in productivity and creativity. Beijing Bank adheres to systematic thinking, deeply applies AIGC technology, and comprehensively lays out the deep integration of artificial intelligence and financial scenarios. It gradually forms a large model application system centered around core algorithms, with multi-model adaptation, multi-task plugins, and multi-source knowledge complementarity, supporting business departments in reconstructing scenarios and experiences with AIGC capabilities.
Chief Information Officer of Beijing Bank, Gong Weihua
Reconstructing the New Paradigm of Industry-Technology Integration with Large Model Technology
Since February 2023, when Beijing Bank began exploring the application of large models in the financial sector, it launched its first large model Q&A tool, Jingzhi Assistant, in April 2023. After multiple version iterations, it successively introduced open-source large models such as Vicuna, ChatGLM, and Baichuan, combining the internal knowledge base with large models to achieve internal knowledge retrieval and Q&A. It also integrated external data queries and RPA plugins, exploring the application of large models in data queries and task execution.
In August 2023, based on the preliminary construction results, Beijing Bank initiated the construction of the AIB financial intelligence application platform, collaborating with 11 business departments to compile over 160,000 pieces of financial knowledge. The platform provides real-time online support for wealth management product queries, business problem resolutions, financial information aggregation, customer marketing scripts, macro policy research, and industry development forecasts for positions such as wealth managers, lobby managers, customer managers, comprehensive tellers, and remote customer service personnel. The AIB platform gradually builds an intelligent application system that covers the front, middle, and back offices, creating AI tools that everyone can use, will use, wants to use, and finds useful, thereby enhancing the professionalism of business operations.
The large model application system includes layers from the bottom up: computing layer, data layer, framework layer, model layer, and application layer. Specifically, the computing layer includes GPU, NPU, CPU, memory, and other computing resources, supporting large model training and inference based on innovative AI computing. The data layer establishes a comprehensive knowledge base, integrating internal regulations, product knowledge, business processes, and other information, loading it into a vector database for unified storage and management. The framework layer constructs the inference and training capabilities of large models, supporting model fine-tuning, model evaluation, model inference, search enhancement, prompt engineering, and other foundational capabilities of large models. The model layer supports the deployment of various mainstream foundational large models across industries, allowing for model fine-tuning with integrated financial knowledge based on specific financial scenarios and tasks, generating large models for the financial industry and scenario task models, ensuring optimal use based on business scenario needs. The application layer connects business scenarios and application systems, providing large model application services.
Technological Innovation Resonating with Business Scenarios
To better integrate with business scenarios, Beijing Bank employs technologies such as Agent intelligent applications, financial knowledge bases, intelligent retrieval, and model fine-tuning in its large model technology practice, enhancing flexibility, interpretability, safety, and adaptability of large models in financial applications.
It establishes an Agent intelligent application capability driven by large models, enabling rapid integration of large model capabilities. The self-developed Agent intelligent application capability offers comprehensive functional support for process orchestration, plugin development, and Agent deployment. The Agent possesses capabilities for semantic understanding, task planning, and plugin invocation, supporting applications in business scenarios such as intelligent Q&A, knowledge retrieval, data analysis, and task execution. Agent technology is an intelligent entity that can achieve environmental perception, decision-making, and action execution. Its core driving force is the large model, but in addition, Beijing Bank has added key components such as planning, memory, and tool usage, enabling the Agent to not only understand and respond to the environment but also to think independently and take action based on goals. Through its modular capabilities, it achieves human-machine collaboration, thereby enhancing business value.
It establishes a diversified financial knowledge base. By employing technologies such as vector databases and knowledge graphs, it forms a comprehensive and three-dimensional financial knowledge graph system covering key business aspects from marketing strategies to operational regulations, regulatory policies to internal audits, ensuring that all business decisions and content generation have a solid data foundation. It breaks down “data silos” and achieves data interoperability with various business systems, building a dedicated large model ecosystem for Beijing Bank, truly serving the front line and business needs.
It independently develops data querying and search engines to enhance the accuracy of financial applications. It builds a search engine that integrates forward and inverted indexing, capable of quickly and accurately locating and extracting the financial knowledge points required by large models, generating content based on the latest and most authoritative data, thereby improving the professionalism and credibility of the content. At the same time, the system automatically cites knowledge sources, enhancing the transparency and traceability of the generated content. Additionally, it independently develops a Text2SQL algorithm and constructs an expert model in the field of indicator knowledge, connecting the indicator knowledge base with large models, decoupling complex underlying data structures, and avoiding issues such as “AI hallucinations” fabricating fields and fictitious data extraction standards.
Based on innovative computing power and framework, it constructs L1 industry large model frameworks and L2 scenario task models. By utilizing innovative computing power and framework, it fine-tunes large models according to the characteristics of banking business, locking certain parameters of the pre-trained model and adding a bypass matrix trained specifically with the accumulated financial knowledge base, ensuring that the model retains its general capabilities while enhancing its generalization and professional adaptability when addressing financial scenario issues, ensuring that the generated content better meets the actual needs of banking operations and effectively serves complex business scenarios.
Challenges of Large Model Applications in Small and Medium-Sized Banks
As a highly regulated industry, banking has always adhered to high standards and requirements in terms of policy. Small and medium-sized banks face numerous challenges in exploring large model applications, including insufficient computing resources, lack of training data, and algorithm safety concerns.
Insufficient computing resources. Training and inference of large models require sufficient AI computing support. The procurement and maintenance of hardware devices necessitate substantial financial investment, placing significant cost pressure on financial institutions. At the same time, against the backdrop of a shortage of high-end GPU chips, there is a significant gap in the demand for mid-to-high-end AI computing power among financial institutions. The current application ecosystem of innovative AI chips is still not fully developed, with adaptation involving multiple levels such as CPU, operating systems, cloud platforms, AI frameworks, acceleration frameworks, and algorithm models, making the adaptation process complex and challenging.
Lack of training data. Data is the foundation for training large models. To effectively address financial business issues, a substantial amount of high-quality, multi-domain financial data is needed for incremental training of large models tailored to specific task scenarios. The data resources available to individual financial institutions are relatively limited, which affects the effectiveness of large model applications to some extent. Financial data is highly sensitive, and challenges exist in data classification management, data desensitization and cleansing, and preventing data bias and misuse.
Algorithm safety concerns. In terms of safety and trust, large models inherently have hallucination issues, their training data is difficult to trace, generated content is untrustworthy, and the computational processes are not explainable, making direct application in financial scenarios with high accuracy requirements and complex business processes challenging. Additionally, in terms of technological ethics, large models may also raise risks of algorithmic discrimination, ethical dilemmas, and fraud, impacting the healthy and sustainable development of the financial industry.
Keeping Up with Industry Exploration in Platform Construction and Scenario Innovation
With the in-depth development of vertical large models in the financial sector, large model applications in banking will increasingly touch on core banking operations. Beijing Bank will keep pace with the developments in the large model industry, strengthen the research and application of multi-modal large models, and continue to explore in both platform construction and scenario innovation.
In terms of platform construction, it will continue to build large model application platforms, coordinating large model computing power, model, and data management, gradually accumulating large model training data, enhancing the security of large models, and utilizing Agent capabilities to achieve rapid integration with other central capabilities within the bank, providing technical support for various business lines to access data and execute business processes using large models.
In terms of scenario innovation, it will explore innovative applications of large models across all scenarios and multiple domains, deeply integrating with scenarios such as credit, collaborative office work, precise marketing, and intelligent customer service, using large model technology to realize functions such as generating credit reports, intelligent meeting minutes, marketing copy generation, and summarizing customer service tickets, empowering the front line to reduce costs and increase efficiency, solidifying the foundation of “big technology,” further enhancing digital transformation capabilities, and comprehensively building a commercial bank driven by artificial intelligence.
(This article was published in the June 2024 issue of Financial Electrification)
Liu Weiguang: Born to Innovate – Financial-Level Cloud Native
Recommended Reading (Click the image to view exciting content)
New Media Center
Director / Kuang Yuan
Editors / Yao Liangyu, Fu Tiantian, Zhang Jun, Tai Siqi
