Business processing system based on heterogeneous computing power

Abstract

The invention provides a business processing system based on heterogeneous computing power, which comprises a computing power scheduling component, a knowledge engineering component, a large model component and an intelligent decision component, and is characterized in that the computing power scheduling component constructs a uniform resource model, maps heterogeneous computing power resources into standard computing units, and allocates corresponding standard computing units for computing tasks; the knowledge engineering component constructs a multilingual knowledge graph, extracts incremental knowledge from the service interaction data and updates the incremental knowledge to the multilingual knowledge graph; the large model component trains the general large model according to the corpus data of the target domain to generate a vertical domain large model; the intelligent decision-making component receives a service instruction, calls a vertical field large model to be combined with a multilingual knowledge graph for reasoning, and generates an executable decision-making flow; and calling an external service system to execute specific operation according to the executable decision flow, and feeding back execution result feedback data to flow back. According to the invention, the overall computing power utilization rate and the cross-language business processing capability are improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and cloud computing technology, and in particular to a business processing system based on heterogeneous computing power. Background Technology

[0002] In the context of globalization, some companies possess large market size and potential consumer groups, enabling them to achieve rapid business growth by expanding into global markets. With the widespread adoption of large-scale modeling technologies and big data applications, companies need to continuously train and deploy more complex models, leading to a dramatic increase in the demand for computing resources from these large-scale models.

[0003] Currently, some enterprises have relatively weak infrastructure in certain regions, and local computing power is often insufficient to independently support their high computing demands. At the same time, existing general-purpose large-scale models and data analysis tools generally suffer from incomplete multilingual coverage and cultural context misunderstandings, making it difficult to support enterprises in effectively processing multilingual business data and providing decision support. Summary of the Invention

[0004] This invention provides a business processing system based on heterogeneous computing power to solve the technical problems of insufficient local computing power and difficulty in cross-language and cross-cultural data processing in enterprises in some regions in the prior art.

[0005] This invention provides a business processing system based on heterogeneous computing power, including a computing power scheduling component, a knowledge engineering component, a large model component, and an intelligent decision-making component, wherein: The computing power scheduling component is used to construct a unified resource model, map heterogeneous computing power resources to standard computing units under the unified resource model, and allocate corresponding standard computing units to computing tasks in the knowledge engineering component, the large model component, and the intelligent decision-making component. The knowledge engineering component is used to construct a multilingual knowledge graph and extract incremental knowledge from business interaction data to update the multilingual knowledge graph. The large model component is used to train a general large model based on corpus data of the target domain to generate a vertical domain large model; The intelligent decision-making component is used to receive business instructions, call the vertical domain big model and combine it with the multilingual knowledge graph to perform reasoning, generate an executable decision flow; and call an external business system to perform specific operations according to the executable decision flow, and feed back the execution result data to the knowledge engineering component and the big model component.

[0006] According to the present invention, a business processing system based on heterogeneous computing power is provided, wherein the computing power scheduling component includes a heterogeneous computing power supply engine, and the heterogeneous computing power supply engine is used for: Configure a standard resource access interface and load a computing power adaptation plugin to map different computing resource parameters to standard computing units in the unified resource model; and, By deploying a lightweight agent on edge devices, the resource idle status and network connection status of edge devices are detected, and the computing power of edge devices that meet the preset conditions are registered as edge computing service nodes in the unified resource model.

[0007] According to the present invention, a business processing system based on heterogeneous computing power is provided, wherein the computing power scheduling component further includes an intelligent computing power scheduling engine, the intelligent computing power scheduling engine being used for: Receive tasks to be processed and extract task feature data of the tasks to be processed; the task feature data includes computational intensity parameters and data compliance tags; Obtain the computing unit characteristic data of currently available standard computing units; the computing unit characteristic data includes resource status data, network connection data, energy consumption cost data, and data location identifier; Based on the task feature data and the computing unit feature data, the task to be processed is sent to a matching target available standard computing unit.

[0008] According to a business processing system based on heterogeneous computing power provided by the present invention, the knowledge engineering component includes a global data flywheel engine, which is used for: Collect implicit business data and perform streaming training on the corresponding vertical domain large model based on the implicit business data; Collect explicit business data, perform privacy field filtering and multilingual semantic alignment processing on the explicit business data, and perform offline full-scale fine-tuning on the vertical domain large model based on the processed explicit business data.

[0009] According to a business processing system based on heterogeneous computing power provided by the present invention, the knowledge engineering component further includes a knowledge engineering engine, the knowledge engineering engine being used for: The system calls the translation management system to translate source language data into text in the target language and writes it back to the corresponding language branch, as well as to identify the changed target text and perform incremental semantic alignment on the target text. Triples are extracted from the text based on the entity recognition model and stored in the multilingual knowledge graph; the triples include entities, relations and attributes. When the vertical domain big model receives an input query, it retrieves the associated nodes and document fragments that match the input query from the multilingual knowledge graph, and combines the associated nodes and document fragments into a prompt word context input to the vertical domain big model.

[0010] According to the present invention, a business processing system based on heterogeneous computing power is provided, wherein the large model component includes a corpus construction module, a targeted fine-tuning module, a word segmentation optimization module, and a reinforcement learning module, wherein: The corpus construction module is used to construct corpus data for the target domain; The targeted fine-tuning module is used to fine-tune the parameters of the pre-trained general large model based on the corpus data of the target domain, and generate a vertical domain large model. The word segmentation optimization module is used to segment the corpus data of the target domain into words based on text segmentation rules; The reinforcement learning module is used to obtain reward signals based on feedback data in the business scenario, and to iteratively optimize the large model of the vertical domain based on the reward signals.

[0011] According to a business processing system based on heterogeneous computing power provided by the present invention, the intelligent decision-making component includes an intelligent agent infrastructure module, which includes a perception unit, a decision-making unit, an execution unit, and a learning unit, wherein: The sensing unit is used to receive and process multimodal data from the target domain and generate an environmental state representation; The decision-making unit is used to generate an executable decision flow corresponding to the business instruction based on the environmental state characterization. The execution unit is used to parse the executable decision flow, call the external business system to perform specific operations through a predefined interface protocol, and monitor the execution status; The learning unit is used to collect feedback data from external business systems and update the decision parameters of the decision unit based on the feedback data.

[0012] According to a heterogeneous computing power-based business processing system provided by the present invention, the intelligent decision-making component further includes a toolchain module for: Establish a standard communication protocol between the intelligent agent and external business systems, and maintain a tool registry to manage the calling interfaces and parameter configurations of external tools; and An isolated execution environment is established. Before the execution unit calls the external business system, the specific operation is pre-run in the isolated execution environment, and the specific operation is risk-verified and blocked based on the security policy.

[0013] According to the business processing system based on heterogeneous computing power provided by the present invention, the intelligent decision-making component further includes an intelligent agent operation and maintenance module and an intelligent agent evaluation module, wherein: The intelligent agent operation and maintenance module is used to provide a visual orchestration interface to configure the workflow logic of the intelligent agent, and to perform deployment management and operation log analysis; The intelligent agent evaluation module is used to construct a multi-dimensional evaluation system and generate optimization suggestion data based on the evaluation results corresponding to the multi-dimensional evaluation system.

[0014] According to the present invention, a business processing system based on heterogeneous computing power is provided, the system further comprising an application execution component, the application execution component comprising: The supply chain management intelligent agent is used to generate inventory optimization strategies based on historical sales data and market trend data, and to call the regulatory knowledge base to perform compliance verification operations. The intelligent customer service agent is used to identify the inquiry intent of multilingual users based on the multilingual knowledge graph and output response data in the corresponding language. The intelligent logistics agent is used to determine the optimal transportation route based on the real-time status data of the logistics service provider, and to adjust the optimal transportation route based on real-time road condition data. The intelligent recommendation agent is used to generate a product recommendation strategy based on the matching degree between user preference features and product features, and to recommend products according to the product recommendation strategy.

[0015] The business processing system based on heterogeneous computing power provided by this invention constructs a unified resource model through the system's computing power scheduling component. This model maps dispersed heterogeneous computing resources to standard computing units, thereby shielding the differences and limitations of underlying heterogeneous hardware. It achieves unified management and efficient allocation of computing resources even in environments with weak infrastructure, avoiding computing bottlenecks caused by weak infrastructure in a single region and improving overall computing power utilization. The multilingual knowledge graph constructed by the knowledge engineering component, combined with a data flywheel mechanism, can continuously extract incremental knowledge from business interaction data and update the graph, providing knowledge support for business processing that is tailored to the culture and language habits of different regions. Large-scale... The model component trains a general-purpose large model using target domain corpus to generate a vertical domain large model. This enhances the model's understanding of industry-specific terms, transaction rules, and marketing logic in cross-border e-commerce scenarios, avoiding the generalization problem of the general-purpose large model in vertical scenarios. After receiving business instructions, the intelligent decision-making component calls the vertical domain large model and combines multilingual knowledge graph reasoning to generate an executable decision flow. This ensures that the decisions meet the business scenarios and language and cultural requirements of different regions. At the same time, its execution result feedback mechanism drives the continuous iteration of the knowledge graph of the knowledge engineering component and the model optimization of the large model component, continuously improving the efficiency of business data processing and the accuracy of decision support. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is one of the structural diagrams of a business processing system based on heterogeneous computing power provided in an embodiment of the present invention.

[0018] Figure 2 This is the second structural schematic diagram of the business processing system based on heterogeneous computing power provided in the embodiments of the present invention.

[0019] Figure 3 This is the third structural schematic diagram of the business processing system based on heterogeneous computing power provided in the embodiments of the present invention.

[0020] Figure 4 This is the fourth structural schematic diagram of the business processing system based on heterogeneous computing power provided in the embodiments of the present invention.

[0021] Figure 5 This is the fifth structural schematic diagram of the business processing system based on heterogeneous computing power provided in the embodiments of the present invention.

[0022] Figure 6 This is the sixth structural schematic diagram of the business processing system based on heterogeneous computing power provided in the embodiments of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0024] It should be noted that in the description of this invention, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. The terms "upper," "lower," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Unless otherwise expressly specified and limited, the terms "installed," "connected," and "linked" should be interpreted broadly, for example, as a fixed connection, a detachable connection, or an integral connection; a mechanical connection or an electrical connection; a direct connection or an indirect connection through an intermediate medium; or a connection within two elements. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0025] The business processing system based on heterogeneous computing power in this embodiment of the invention, such as Figure 1 As shown, it includes a computing power scheduling component 10, a knowledge engineering component 20, a large model component 30, and an intelligent decision-making component 40.

[0026] The computing power scheduling component is used to construct a unified resource model, map heterogeneous computing power resources to standard computing units under the unified resource model, and allocate corresponding standard computing units to computing tasks in the knowledge engineering component, the large model component, and the intelligent decision-making component.

[0027] In this embodiment, the computing power scheduling component is responsible for the supply and scheduling of computing resources for the entire system. Its goal is to shield the physical differences of the underlying heterogeneous computing power resources and build a logically unified computing power resource pool that can be allocated on demand.

[0028] Here, heterogeneous computing resources refer to computing hardware resources of different architectures and types, including but not limited to: general computing resources, parallel computing resources, neural network-specific computing resources, programmable heterogeneous computing resources, etc.

[0029] Specifically, the computing power scheduling component constructs a unified resource model. This unified resource model is a standardized abstract framework for computing power resources, defining the descriptive dimensions of computing power resources, such as computing power performance, storage capacity, latency indicators, energy consumption parameters, and other descriptive dimensions, units of measurement, and scheduling interfaces, thus shielding the underlying differences of heterogeneous computing power.

[0030] This unified resource model can define one or more standard computing units. A standard computing unit is the smallest logical unit for computing power scheduling, defined by a set of parameters such as memory size, GPU computing power, network bandwidth, and cost per unit time. For example, the computing power of one GPU card can be mapped to N standard computing units, and one CPU server can be mapped to M standard computing units, facilitating the unified allocation of computing power across different types.

[0031] Based on this, virtual machine instances from different cloud service providers, or physical servers in enterprise-built data centers, with their different configuration parameters, can all be converted into several standard computing units through a mapping process.

[0032] When other components in the system have computational tasks to perform, such as a large model component needing to train a model, or a knowledge engineering component needing to process a batch of user interaction data, they will send resource requests to the computing power scheduling component. This resource request will describe the characteristics of the computational task; for example, a model training task is computationally intensive and requires a high-performance GPU, while a data processing task might be I / O intensive. Upon receiving the resource request, the computing power scheduling component will allocate one or more suitable standard computing units to the computational task based on the real-time status of its maintained global standard computing unit resource pool. For example, a high-performance standard computing unit cluster containing GPU resources might be allocated to the model training task, while a set of standard computing units with high I / O performance might be allocated to the data processing task. In this way, upper-layer applications do not need to care about the physical location or hardware on which their tasks run, thus decoupling computing resources from business logic.

[0033] The knowledge engineering component is used to construct a multilingual knowledge graph and extract incremental knowledge from business interaction data to update the multilingual knowledge graph.

[0034] In this embodiment, the goal of the knowledge engineering component is to build and maintain a dynamic, global knowledge brain for the entire intelligent agent system, enabling it to understand and utilize business-related professional knowledge.

[0035] Here, incremental knowledge refers to new knowledge that is not included in the current multilingual knowledge graph, such as newly emerging business terms, newly established corporate partnerships, and newly released product attributes.

[0036] Specifically, the knowledge engineering component constructs a multilingual knowledge graph. Here, the multilingual knowledge graph is a structured semantic network used to store entities, relationships between entities, and entity attributes. It should be understood that multiple languages ​​refer to the same entity or relationship, resulting in representations in multiple languages. For example, the entity "T-shirt" might be associated with tags in English, Spanish, and Thai, ensuring the system accurately understands the semantic meaning of instructions in different languages.

[0037] To keep the multilingual knowledge graph up-to-date, the knowledge engineering component also employs a data flywheel mechanism for data updates. Specifically, when the intelligent agent system interacts with external users or business systems, it generates a large amount of business interaction data, such as user query logs, customer service conversation records, product reviews, and order data. The data flywheel mechanism continuously extracts new incremental knowledge from this flowing business interaction data automatically or semi-automatically through technologies such as natural language processing and data mining. For example, by analyzing a large number of customer service conversations, the system discovers that many inquiries about product B are related to waterproof performance. The system can then extract the incremental knowledge that "product B has a highly relevant attribute: waterproof performance" and update it in the multilingual knowledge graph.

[0038] The large model component is used to train a general large model based on corpus data of the target domain to generate a vertical domain large model.

[0039] Here, the general large model refers to a basic model pre-trained on massive amounts of internet text. It has general language understanding, generation, and reasoning capabilities, but is not optimized for a specific domain.

[0040] In this embodiment, the target domain refers to the specific business scenario that the system targets, such as cross-border e-commerce business scenario, industrial manufacturing business scenario, etc. Specifically, target corpus data for the specific business scenario that the system targets is collected in advance, including but not limited to professional terminology, historical business records, and industry standard documents in the target domain. Then, based on the general large model, the target domain corpus data is used for fine-tuning to obtain a vertical domain large model. This vertical domain large model retains the language capabilities of the general large model, while also possessing the professional knowledge and business logic of the target domain.

[0041] The intelligent decision-making component is used to receive business instructions, call the vertical domain big model and combine it with the multilingual knowledge graph to perform reasoning, generate an executable decision flow; and call an external business system to perform specific operations according to the executable decision flow, and feed back the execution result data to the knowledge engineering component and the big model component.

[0042] Here, business instructions refer to externally inputted business needs, such as querying the market share of a product, generating a quarterly sales strategy, or executing a refund process for a customer order. These are the triggering conditions for intelligent decision-making.

[0043] In this embodiment, after receiving a business instruction, a large model of the vertical domain is invoked for logical reasoning. At the same time, in order to ensure the accuracy and timeliness of the decision, information from a multilingual knowledge graph is also used as context input. After reasoning and analysis, an executable decision flow is generated.

[0044] Here, an executable decision flow refers to a business process with clearly defined execution steps; it is the vehicle that transforms abstract decisions into concrete operations. For example, a refund process decision flow might include steps such as verifying order information, confirming refund conditions, accessing the financial system, and sending a refund notification.

[0045] In this embodiment, after obtaining the executable decision flow, the external business system, such as the ERP system, WMS warehousing system, TMS logistics system, etc., is called through the predefined API interface to perform specific operations according to the above decision flow.

[0046] Furthermore, after the external business system finishes execution, it will return corresponding execution result feedback data, such as successful replenishment or logistics delays, and feed this data back to the knowledge engineering component and the large model component. The knowledge engineering component can extract incremental knowledge to update the graph, and the large model component can use this data to further optimize model performance.

[0047] The heterogeneous computing power-based business processing system of this invention constructs a unified resource model through its computing power scheduling component, mapping dispersed heterogeneous computing resources to standard computing units. This shields the differences and limitations of underlying heterogeneous hardware, enabling unified management and efficient allocation of computing resources even in environments with weak infrastructure. It avoids computing bottlenecks caused by weak infrastructure in a single region, improving overall computing power utilization. The multilingual knowledge graph constructed by the knowledge engineering component, combined with a data flywheel mechanism, continuously extracts incremental knowledge from business interaction data and updates the graph, providing knowledge support for business processing that aligns with the cultures and language habits of different regions. Large-scale... The model component trains a general-purpose large model using target domain corpus to generate a vertical domain large model. This enhances the model's understanding of industry-specific terms, transaction rules, and marketing logic in cross-border e-commerce scenarios, avoiding the generalization problem of the general-purpose large model in vertical scenarios. After receiving business instructions, the intelligent decision-making component calls the vertical domain large model and combines multilingual knowledge graph reasoning to generate an executable decision flow. This ensures that the decisions meet the business scenarios and language and cultural requirements of different regions. At the same time, its execution result feedback mechanism drives the continuous iteration of the knowledge graph of the knowledge engineering component and the model optimization of the large model component, continuously improving the efficiency of business data processing and the accuracy of decision support.

[0048] It should be noted that each implementation method of this application can be freely combined, rearranged, or executed individually, and does not need to rely on or depend on a fixed execution order.

[0049] In some embodiments, reference Figure 2 As shown, the computing power scheduling component includes a heterogeneous computing power supply engine 110, which is used for: Configure a standard resource access interface and load a computing power adaptation plugin to map different computing resource parameters to standard computing units in the unified resource model; and, By deploying a lightweight agent on edge devices, the resource idle status and network connection status of edge devices are detected, and the computing power of edge devices that meet the preset conditions are registered as edge computing service nodes in the unified resource model.

[0050] In this embodiment, a cross-border e-commerce scenario is used as an example for explanation. In the cross-border e-commerce scenario, computing resources are usually multi-level: one end is the high-performance computing power provided by centralized cloud vendors, and the other end is the edge devices scattered in overseas warehouses and logistics stations.

[0051] The heterogeneous computing power supply engine first supports the access of general-purpose computing power by configuring a standard resource access interface. This standard resource access interface is a predefined set of standardized computing power resource access specifications, including data interaction protocols, parameter transmission formats, authentication and authorization mechanisms, etc., to ensure that computing power resources from different sources can be identified and managed in a unified manner.

[0052] To adapt to the computing resources of different cloud service provider platforms, the heterogeneous computing power supply engine adopts a plug-in architecture, developing a computing power adaptation plug-in for each cloud service provider platform.

[0053] Specifically, for general-purpose computing power, such as computing resources from different cloud service provider platforms, the heterogeneous computing power supply engine loads corresponding computing power adaptation plugins to extract common attributes of multiple cloud platforms and achieve unified modeling and allocation of resources. For example, an AWS adaptation plugin calls AWS's native API to obtain specific parameters such as the CPU model, GPU card type, memory size, and network bandwidth of its EC2 instance. Then, through preset conversion rules, it abstracts and maps these heterogeneous resources into standard computing units under a unified resource model. For supercomputing and intelligent computing resources, such as national supercomputing centers or enterprise-built AI computing clusters, the heterogeneous computing power supply engine, through its matching computing power adaptation plugins, first converts the heterogeneous resources into an intermediate representation independent of hardware architecture. Then, the unified intermediate representation is further compiled and adapted for efficient execution in this specific environment, thereby ensuring the efficient execution of computing tasks in the heterogeneous resource pool.

[0054] Furthermore, for edge devices, such as industrial control computers in factories, edge node servers in smart cities, and local computing terminals in retail stores, the computing power of these devices is typically used to handle low-latency, localized computing tasks. In this embodiment, a lightweight agent is deployed on various edge devices. The lightweight agent monitors the resource idle status and network connection status of the edge devices in real time. The lightweight agent reports the collected status data to the heterogeneous computing power supply engine. The heterogeneous computing power supply engine determines whether the edge device has the capability to contribute computing power based on preset conditions. For edge devices that meet the preset conditions, the heterogeneous computing power supply engine registers its computing power as an edge computing service node and maps it to the corresponding standard computing unit, incorporating it into the global computing power resource pool.

[0055] The heterogeneous computing power-based business processing system of this invention defines a standard resource access interface through a heterogeneous computing power supply engine, supports unified abstraction of heterogeneous computing power resources, flexibly adapts to the management modes and service differences of different computing power platforms, and shields the complexity of underlying technologies through unified resource description and interface specifications.

[0056] In some embodiments, reference Figure 2 As shown, the computing power scheduling component further includes an intelligent computing power scheduling engine 120, which is used for: Receive tasks to be processed and extract task feature data of the tasks to be processed; the task feature data includes computational intensity parameters and data compliance tags; Obtain the computing unit characteristic data of currently available standard computing units; the computing unit characteristic data includes resource status data, network connection data, energy consumption cost data, and data location identifier; Based on the task feature data and the computing unit feature data, the task to be processed is sent to a matching target available standard computing unit.

[0057] In this embodiment, when the intelligent computing power scheduling engine receives a task to be processed, it extracts the task's feature data, including but not limited to: Computational intensity parameters: These are used to measure the intensity of a task's consumption of computing resources, such as floating-point operations, parallel computing requirements, peak memory usage, and storage I / O throughput. For example, large model training is a high computational intensity task, while ordinary data cleaning tasks are low computational intensity tasks. Data compliance labels: These label the data security and compliance requirements that must be followed when performing tasks, such as data location restriction labels, privacy protection level labels, and access permission requirement labels.

[0058] Simultaneously, the intelligent computing power scheduling engine will obtain computing unit characteristic data of all currently available standard computing units from the unified resource model in real time, including but not limited to: Resource status data: Real-time hardware load of standard computing units, such as utilization rate, free memory capacity, remaining storage space, peak computing power, etc. Network connectivity data: Network transmission capabilities and stability metrics of standard computing units, such as network bandwidth, latency, packet loss rate, and distance to data storage nodes; Energy cost data: Operating costs and energy consumption indicators of standard computing units, such as the billing price per unit of computing power, hourly power consumption, energy efficiency ratio, etc. Data location identifier: Physical deployment location information of standard computing units, used to match data compliance tags for tasks, such as geographical identifiers like Beijing edge node, Shanghai cloud node, etc.

[0059] The intelligent computing power scheduling engine inputs the above task feature data and computing unit feature data into a multi-objective optimization function. This multi-objective optimization function has multiple built-in optimization objectives and weight coefficients to calculate the optimal match between the task to be processed and the currently available standard computing units. Here, the optimization objectives include, but are not limited to, optimal performance, lowest cost, and strongest compliance. The function will dynamically adjust the weight of each objective according to the task priority and output the optimal matching result, that is, the target available standard computing units matched by the current task to be processed.

[0060] The heterogeneous computing power-based business processing system of this invention coordinates heterogeneous computing power resources through an intelligent computing power scheduling engine. It comprehensively considers multiple factors such as task characteristics, resource status, network conditions, and energy consumption costs, and automatically selects the optimal execution location and path of computing power according to a multi-objective optimization mechanism to meet the diverse needs of business scenarios.

[0061] In some embodiments, reference Figure 3 As shown, the knowledge engineering component includes a global data flywheel engine 210, which is used for: Collect implicit business data and perform streaming training on the corresponding vertical domain large model based on the implicit business data; Collect explicit business data, perform privacy field filtering and multilingual semantic alignment processing on the explicit business data, and perform offline full-scale fine-tuning on the vertical domain large model based on the processed explicit business data.

[0062] Here, implicit business data refers to indirect data that is not directly generated during business interactions and has no explicit structured form, including but not limited to: User behavior data: such as data on e-commerce websites, including every product click, page dwell time, adding to cart, placing an order, and even return behavior. Intermediate log data: such as the number of times the content generated by the large model was corrected, and the time distribution data of the inference path; Operational status data: such as the success rate of execution of certain decision-making instructions, resource usage fluctuation data, etc.

[0063] It should be noted that, due to the large volume and time-sensitive nature of such implicit business data, the global data flywheel engine in this embodiment does not perform deep cleaning on it. Instead, it directly transforms it into a vector stream through feature engineering and inputs it into the corresponding vertical domain large model for streaming training. Here, streaming training means that it does not need to wait for the full accumulation of data, but can receive data streams in real time and continuously and slightly update the parameters of the model so that the model can quickly adapt to business dynamics.

[0064] Here, explicit business data refers to data that is directly generated in the business scenario and has clear business characteristics, including but not limited to: multilingual customer service chat logs, user comments on social media, detailed product reviews, and voice quality inspection recordings from call centers.

[0065] In this embodiment, the global data flywheel engine periodically collects explicit business data in batches. This explicit business data is unstructured or semi-structured data rich in semantic information. Before being used for training, the explicit business data undergoes preprocessing operations such as privacy field filtering and multilingual semantic alignment. Specifically, the system calls the entity recognition model to automatically identify and de-identify personal identity information in the data, such as names, phone numbers, and addresses, ensuring that the training process complies with data privacy regulations. Then, different languages ​​are mapped to the same semantic space to eliminate language barriers. Finally, the processed explicit business data is used to perform offline full-scale fine-tuning of the large-scale model in the vertical domain.

[0066] The business processing system based on heterogeneous computing power in this embodiment of the invention improves the model performance of large models in vertical domains through the dual-loop parallel data processing flow of the global data flywheel engine.

[0067] In some embodiments, reference Figure 3 As shown, the knowledge engineering component further includes a knowledge engineering engine 220, which is used for: The system calls the translation management system to translate source language data into text in the target language and writes it back to the corresponding language branch, as well as to identify the changed target text and perform incremental semantic alignment on the target text. Triples are extracted from the text based on the entity recognition model and stored in the multilingual knowledge graph; the triples include entities, relations and attributes. When the vertical domain big model receives an input query, it retrieves the associated nodes and document fragments that match the input query from the multilingual knowledge graph, and combines the associated nodes and document fragments into a prompt word context input to the vertical domain big model.

[0068] Here, source language data refers to the original business text data to be translated, and target language text refers to the corresponding language text obtained after the source language data has been converted by the translation management system. Language branches refer to independent storage branches in the multilingual knowledge graph divided by language, such as Chinese branches, English branches, etc. Knowledge texts of different languages ​​will be classified and stored in the corresponding branches to realize the language-structured management of the knowledge graph.

[0069] In this embodiment, the knowledge engineering engine integrates a three-layer architecture of a translation management system, a knowledge graph, and retrieval enhancement generation. Specifically, in order to build and maintain the consistency of multilingual knowledge, the knowledge engineering engine calls a translation management system to translate source language data into target language text, such as translating Chinese product details or regulations into English product details or regulations; and writes it back to the corresponding language branch in the knowledge graph.

[0070] Furthermore, to reduce the cost of multilingual maintenance, the knowledge engineering engine employs an incremental semantic alignment strategy during the process of translating the target language text and writing it back to the corresponding language branch. Instead of performing a full translation every time the source text changes, it only identifies the changed target text and triggers incremental semantic alignment of these changed portions, avoiding a full recalculation and saving resources.

[0071] To transform unstructured text data into machine-understandable structured knowledge, the knowledge engineering engine also utilizes pre-trained entity recognition models. For example, based on the BERT model to process various text data, the entity recognition model can automatically extract triples of entities, relations, and attributes from the text. For instance, based on the entity recognition model, the regulatory nodes in the text are semantically linked with document nodes such as product models and test reports to form regulatory-product-market triples that can jump across languages. These triples are stored in the graph database under the corresponding language branch to build a multilingual knowledge graph.

[0072] To enable large-scale models to accurately utilize this knowledge, the knowledge engineering engine can also implement a retrieval-enhanced generation mechanism. When a vertical domain large-scale model receives a user's input query, the system does not directly allow the model to perform reasoning. Instead, the knowledge engineering engine first performs semantic analysis on the input query and searches the multilingual knowledge graph to retrieve the most matching related nodes and document fragments. Here, related nodes refer to entity nodes in the multilingual knowledge graph that semantically match the input query, and document fragments refer to domain-specific document content fragments related to the input query. After retrieving the related nodes and document fragments, these nodes and fragments are combined into a structured contextual prompt, which is then input into the vertical domain large-scale model along with the input query. This provides clear knowledge support for the model's reasoning process, improving the accuracy and professionalism of the model's responses.

[0073] The business processing system based on heterogeneous computing power in this embodiment of the invention not only realizes unified management and dynamic updating of multilingual knowledge through a knowledge engineering engine, but also improves the accuracy and reliability of large models in answering questions in vertical domains.

[0074] In some embodiments, reference Figure 4 The large model components include a corpus construction module 410, a targeted fine-tuning module 320, a word segmentation optimization module 330, and a reinforcement learning module 340, wherein: The corpus construction module is used to construct corpus data for the target domain; The targeted fine-tuning module is used to fine-tune the parameters of the pre-trained general large model based on the corpus data of the target domain, and generate a vertical domain large model. The word segmentation optimization module is used to segment the corpus data of the target domain into words based on text segmentation rules; The reinforcement learning module is used to obtain reward signals based on feedback data in the business scenario, and to iteratively optimize the large model of the vertical domain based on the reward signals.

[0075] In this embodiment, the target domain is cross-border e-commerce, which is used as an example for explanation. The original corpus data can be product descriptions, user reviews, customer service dialogues, social media content, etc. of cross-border e-commerce platforms. After collecting the original corpus data in the cross-border e-commerce domain, the original corpus data is preprocessed such as cleaning, noise reduction, and annotation. Then, the preprocessed corpus data is screened by combining the pre-set domain expert rules and preset cleaning rules to ensure the instructions and accuracy of the corpus data, thereby obtaining the processed target corpus data.

[0076] Here, domain expert rules refer to the corpus screening standard rules formulated by professionals in the field, while preset cleaning rules refer to the automated screening rules built into the system, such as filtering invalid text that is too short, removing general content that is irrelevant to the target domain, and correcting typos.

[0077] After obtaining the corpus data of the target domain, the parameters of the pre-trained general model are fine-tuned using few-shot learning or zero-shot learning, allowing the model to quickly adapt to specific domain tasks. Compared with full-parameter training, this can significantly reduce the consumption of computing resources and the training cycle.

[0078] Furthermore, in this embodiment, during model training, the corpus data is segmented into words based on preset text segmentation rules to improve the model's ability to understand domain-specific terms and process multilingual text. Specifically, corresponding text segmentation rules are designed for the grammatical features and lexical structures of different languages. For example, for Chinese text, text segmentation rules based on word frequency and context can be used; for English text, text segmentation rules based on spaces and punctuation marks can be used. Simultaneously, for special text content such as code snippets and formula symbols, multimodal lexical mapping can be used to transform the code syntax structure and mathematical formula logic into lexical sequences that the model can understand.

[0079] Furthermore, in order to achieve continuous optimization of the model, the large model component in this embodiment will also obtain reward signals based on feedback data in the business scenario. For example, in the intelligent customer service scenario, if the customer service personnel adopt the answer generated by the model, a positive reward signal will be generated; if the customer service personnel modify the answer generated by the model, a negative reward signal will be generated based on the difference data before and after the modification. Finally, the large model will be iteratively optimized based on the corresponding reward signals.

[0080] The business processing system based on heterogeneous computing power in this invention improves the task processing capability of large models in the target domain through corpus construction, targeted fine-tuning strategies, word segmentation optimization and reinforcement learning mechanisms.

[0081] In some embodiments, reference Figure 5 The intelligent decision-making component includes an intelligent agent infrastructure module 410, which includes a perception unit, a decision-making unit, an execution unit, and a learning unit, wherein: The sensing unit is used to receive and process multimodal data from the target domain and generate an environmental state representation; The decision-making unit is used to generate an executable decision flow corresponding to the business instruction based on the environmental state characterization. The execution unit is used to parse the executable decision flow, call the external business system to perform specific operations through a predefined interface protocol, and monitor the execution status; The learning unit is used to collect feedback data from external business systems and update the decision parameters of the decision unit based on the feedback data.

[0082] Here, "environmental state representation" refers to the abstract description of the business environment generated after fusing, analyzing, and structuring multimodal data.

[0083] In this embodiment, we will continue to use the cross-border e-commerce field as an example to explain the concept. The perception unit collects different types of input data from cross-border e-commerce business, such as market data, user behavior data, product data, and logistics data. Through data fusion, feature extraction, noise filtering, and other processing, the scattered multimodal data is transformed into a structured environmental state representation to present key information about the current business environment and provide input basis for the decision-making unit.

[0084] The decision-making unit leverages the reasoning capabilities of a large model, combined with the business rules and strategies of cross-border e-commerce, to conduct intelligent decision-making using a directed graph or state graph visualization orchestration method. Specifically, the decision-making unit first breaks down complex business instructions into logically clear and step-by-step executable decision flows based on environmental state representations. Each decision node in the executable decision flow is associated with a corresponding business rule verification step, ensuring both the accuracy of the decision results and achieving full traceability from instruction input to decision output. Taking pricing strategy formulation as an example, the agent in the decision-making unit can automatically integrate environmental state representations such as cost, market, competitors, and cross-border supply chains based on the above decision logic to perform comprehensive reasoning, ultimately generating a reasonable pricing scheme that conforms to business objectives and market rules.

[0085] The execution unit parses the executable decision flow, transforms it into specific operation instructions, and executes the corresponding tasks. Through predefined interface protocols, it calls external business systems, such as order management systems, logistics tracking systems, and payment systems, to achieve automated task execution. During the execution process, it also monitors the execution status of the tasks in real time.

[0086] The learning unit optimizes the decision-making parameters of the agents in the decision-making unit by collecting feedback data during task execution, such as user feedback and system performance indicators, thus guiding the agents in the decision-making unit to continuously optimize their decision-making strategies.

[0087] In some embodiments, reference Figure 5 The intelligent decision-making component further includes a toolchain module 420, used for: Establish a standard communication protocol between the intelligent agent and external business systems, and maintain a tool registry to manage the calling interfaces and parameter configurations of external tools; and An isolated execution environment is established. Before the execution unit calls the external business system, the specific operation is pre-run in the isolated execution environment, and the specific operation is risk-verified and blocked based on the security policy.

[0088] Here, the standard communication protocol defines the communication specifications and interface standards between the agent and external business systems. Adopting a unified standard communication protocol ensures seamless integration between the agent and various external business systems, improving system compatibility and scalability. For example, defining connection protocols based on RESTful APIs or message queues enables data interaction between the agent and order management systems, logistics tracking systems, etc.

[0089] In this embodiment, the toolchain module maintains a tool registry to manage the metadata of all external tools, including but not limited to the tool's name, function description, API URL, HTTP method, and detailed parameter configuration. When an agent needs to call an external tool, it can search for a suitable tool in the tool registry through a tool discovery mechanism. For example, when querying logistics information, the agent can find a logistics tracking tool in the tool registry and call its API to obtain the logistics information.

[0090] To ensure operational security, the toolchain module establishes an isolated execution environment. Before the execution unit actually calls the production environment interface of the external business system, the specific operation instructions generated by the decision-making unit are first sent to this isolated environment for pre-running. This isolated environment can be a containerized environment that deploys a simulated interface of the external business system or connects to its test environment. During pre-running, the toolchain module performs risk verification and blocking on the operation based on preset security policies. These security policies may include: Permission verification: Check whether the agent has permission to perform this operation. For example, a customer service agent should not have permission to modify product prices. Parameter validation: Check whether the parameters of the API call are compliant, for example, to prevent SQL injection and check whether the price modification range is within a reasonable range; High-risk operation blocking: Identify and block operations that may cause serious consequences, such as bulk data deletion.

[0091] Only when an operation has been successfully pre-run in the isolated execution environment and has passed all risk checks will the execution unit be authorized to call the real production environment interface.

[0092] Furthermore, in this embodiment, the toolchain module also possesses data access and memory capabilities, responsible for managing and accessing the data required by the agent, including business data, user data, market data, etc. Through a unified data access interface, the agent can easily obtain the necessary data. Simultaneously, the data access mechanism controls access permissions to ensure data security. The memory mechanism is responsible for storing the agent's historical decisions and execution records, providing data support for the agent's learning and optimization. For example, the agent can remember users' historical purchase records and preferences, providing a basis for subsequent recommendations.

[0093] The heterogeneous computing power-based business processing system of this invention provides support for the decision-making and execution of intelligent agents through multiple aspects, including connection protocols, tool registration and discovery mechanisms, isolated execution environments, data access and memory capabilities, and security and compliance control mechanisms.

[0094] In some embodiments, reference Figure 5 The intelligent decision-making component further includes an intelligent agent operation and maintenance module 430 and an intelligent agent evaluation module 440, wherein: The intelligent agent operation and maintenance module is used to provide a visual orchestration interface to configure the workflow logic of the intelligent agent, and to perform deployment management and operation log analysis; The intelligent agent evaluation module is used to construct a multi-dimensional evaluation system and generate optimization suggestion data based on the evaluation results corresponding to the multi-dimensional evaluation system.

[0095] In this embodiment, the intelligent agent operation and maintenance module provides a visual orchestration interface, further productizing the intelligent agent framework and toolchain to form a one-stop capability from design, testing, deployment to operational observation. In the visual orchestration interface, users can easily design the behavior and decision-making logic of the intelligent agent, configure and adjust the intelligent agent through the interface; it also provides rich testing tools and environments, allowing users to test the intelligent agent in different scenarios to verify its performance and stability; simultaneously, it supports one-click deployment, quickly deploying the intelligent agent to the actual business environment, and enabling users to monitor the operational status and performance indicators of the intelligent agent in real time, promptly identifying and resolving problems.

[0096] Specifically, during the design phase, a visual design tool for the agent framework and toolchain is provided in the visual orchestration interface, allowing users to quickly build agents through drag-and-drop, configuration, and other methods. The design tool includes a task orchestrator, node editor, and connector designer, enabling users to intuitively design the agent's decision-making process and execution logic. Simultaneously, the design tool provides a pre-evaluation function for agent performance, helping users optimize the agent's design.

[0097] During the testing phase, a comprehensive suite of testing tools and test environment deployment tools are provided within the visual orchestration interface, supporting testing of the agent's functionality, performance, and security. Testing tools include unit testing tools, integration testing tools, and stress testing tools, capable of simulating various business scenarios and anomalies to comprehensively test the agent. The test environment provides a runtime environment similar to the production environment, ensuring the accuracy of test results.

[0098] During the deployment phase, a visual orchestration interface provides deployment tools for agents and deployment schemes, supporting one-click deployment and rapid deployment of agents. Deployment tools include packaging, configuration, and installation tools, which can package agents and their dependent components into executable files for easy deployment. Deployment schemes provide detailed deployment guidelines and configuration instructions to help users successfully deploy agents.

[0099] During the operational observation phase, a visual orchestration interface provides tools for monitoring and analyzing the agent's operational status, performance metrics, and error messages in real time. The operational monitoring tools display the agent's running status, such as task execution progress and resource usage, helping users to promptly identify and resolve issues. The log analysis tools allow for the retrieval and analysis of the agent's operational logs, providing detailed log reports and statistical information.

[0100] Furthermore, in addition to the intelligent agent operation and maintenance module, the intelligent decision-making component in this embodiment also includes an intelligent agent evaluation module, which constructs a multi-dimensional evaluation system from three levels: offline capabilities, online behavior, and end-to-end business results, to achieve a comprehensive and objective evaluation of the intelligent agent's performance and provide a basis for the optimization and improvement of the intelligent agent.

[0101] Specifically, offline capability assessment primarily evaluates the agent's atomic capabilities, including reasoning ability, retrieval ability, tool selection and invocation correctness, etc. Automated testing tools are used to perform regression testing on each atomic capability of the agent to ensure that upgrades do not introduce degradation. For example, in reasoning capability assessment, a series of test cases can be designed to evaluate the agent's reasoning accuracy and logic in complex scenarios. Online behavior assessment focuses on the agent's success rate, path efficiency, average cost and latency under real traffic, as well as its error self-healing and escalation rates. Online behavior data is collected by monitoring the agent's operational status and performance indicators in real time to evaluate the agent's online behavior. Simultaneously, observable call chains are used to locate bottlenecks and analyze the agent's performance bottlenecks and optimization points during task execution. End-to-end business assessment is based on business objectives, measuring task first-pass rate, human intervention rate, compliance risk event rate, and customer experience indicators. By comparing business data before and after the agent executes the task, the contribution and optimization effect of the agent to business objectives are evaluated. For example, in customer experience metrics evaluation, user surveys and feedback analysis are used to assess the impact of intelligent agents on customer satisfaction and loyalty.

[0102] Based on the multi-dimensional evaluation system built upon the above three levels, the system evaluates intelligent agents and generates comprehensive evaluation results by integrating standardized datasets and task sets, reproducible environments and benchmarks, and in-process and post-process auditing mechanisms. Furthermore, it automatically generates optimization suggestions based on the evaluation results. For example, if low customer satisfaction in end-to-end business results is identified, and online behavioral data analysis pinpoints the cause as excessive latency in a specific API call, the system will generate optimization suggestions for that API call.

[0103] In some embodiments, reference Figure 1 As shown, the system further includes an application execution component 50, which includes: The supply chain management intelligent agent is used to generate inventory optimization strategies based on historical sales data and market trend data, and to call the regulatory knowledge base to perform compliance verification operations. The intelligent customer service agent is used to identify the inquiry intent of multilingual users based on the multilingual knowledge graph and output response data in the corresponding language. The intelligent logistics agent is used to determine the optimal transportation route based on the real-time status data of the logistics service provider, and to adjust the optimal transportation route based on real-time road condition data. The intelligent recommendation agent is used to generate a product recommendation strategy based on the matching degree between user preference features and product features, and to recommend products according to the product recommendation strategy.

[0104] Specifically, the supply chain management intelligent agent dynamically integrates multi-dimensional data such as historical sales data (e.g., sales volume, sales period, regional distribution, return rate, etc.) and market trend data (e.g., competitor sales dynamics, changes in consumer preferences, seasonal demand fluctuations) across the supply chain to generate inventory optimization strategies, such as calculating the optimal safety stock and replenishment point for each overseas warehouse. Furthermore, before executing replenishment decisions, the supply chain management intelligent agent automatically invokes a regulatory knowledge base to perform compliance checks. This regulatory knowledge base is a pre-set structured knowledge base containing information on tariff rules, product compliance standards, and logistics regulatory requirements for the target market. It is used to verify whether inventory strategies and commodity circulation comply with relevant regulations. For example, it checks whether the goods to be purchased meet the latest import standards of the target country, thereby avoiding inventory backlog and losses due to compliance issues.

[0105] When receiving inquiries from users in different countries, the intelligent customer service system first utilizes a multilingual knowledge graph to identify the user's inquiry intent. For example, even if a user asks "When will it arrive?" in a Thai dialect, the system can accurately understand that their intent is to inquire about the logistics status. Then, by combining real-time information on products, orders, and logistics, the system calls upon a large-scale model in the vertical domain to output response data in the corresponding language, providing fluent, accurate, and culturally appropriate answers.

[0106] In the transportation and customs clearance process, intelligent logistics agents, combined with large-scale modeling technology, connect in real-time with logistics service providers' status data, such as capacity, pricing, and estimated delivery time for each route. This allows for real-time planning of optimal transportation routes for vehicles. By monitoring transportation demand and road conditions in real time, the system dynamically analyzes the timeliness, cost, and risk of transportation routes to calculate the shortest transit time, automatically recommending the optimal route. During transit, it continuously monitors real-time road conditions along the recommended optimal route and dynamically adjusts it accordingly. For example, if a planned transit airport is detected to be closed due to heavy snow, the system automatically replans the route and notifies relevant parties, thereby minimizing transportation delays.

[0107] Intelligent recommendation systems monitor and predict consumer behavior and preferences. By collecting and analyzing data such as browsing history, purchase history, and search keywords, they deeply analyze user preference characteristics and accurately predict consumer demand. Simultaneously, they extract product features from product information. Deep learning models calculate the matching degree between user preference features and product features to generate product recommendation strategies. For example, after a user browses beach shorts, sunglasses and sunscreen will be prioritized for recommendations. Finally, the system recommends products based on the product recommendation strategy, displaying a customized product list to each user across different touchpoints such as the website homepage and app feeds, thereby improving user experience and purchase conversion rates.

[0108] The heterogeneous computing power-based business processing system of this invention integrates intelligent technologies from warehousing, logistics, and user services into its execution components. This enables capabilities such as demand forecasting and inventory optimization, logistics route planning and inventory layout, compliance risk management, and multilingual support. It improves the efficiency and accuracy of supply chain management, enhances transportation timeliness, reduces logistics costs, operating costs and risks, increases customer service response speed and quality, and improves user experience and purchase conversion rates. This provides users with more intelligent, efficient, convenient, on-demand, and user-friendly services. Simultaneously, the intelligent agent connects with external systems such as ERP (Enterprise Resource Planning), WMS (Warehouse Management System), TMS (Transportation Management System), and payment and risk control systems to complete long-chain tasks such as inventory optimization, price monitoring, contract generation, customs clearance, and reconciliation. Furthermore, it effectively controls cross-border legal and tax compliance by combining a localized knowledge base and verifiable identity.

[0109] For ease of understanding, this embodiment uses a cross-border e-commerce scenario as an example for explanation. (Reference) Figure 6 As shown, this embodiment constructs a business processing system based on heterogeneous computing power. Through cross-border computing power integration, it achieves efficient scheduling and maximizes resource utilization of computing resources, breaks through the limitations of computing resources in a single region, avoids the impact of insufficient local computing power on business deployment and operational efficiency, and improves the utilization rate of computing resources. By integrating specialized large-scale model capabilities, it achieves multi-language and multi-cultural interaction under a unified architecture, enabling enterprises to achieve a higher degree of uniformity and flexibility in a complex global environment. The cross-border e-commerce intelligent agent connects with external business systems such as ERP enterprise resource planning system, WMS warehouse management system, TMS transportation management system, payment and risk control system to complete long-link tasks such as inventory optimization, price monitoring, contract generation, and reconciliation settlement, thereby improving decision-making efficiency in complex scenarios.

[0110] Specifically, the system includes integrated computing power and cross-border intelligent computing components. For computationally intensive tasks such as large-scale modeling and big data analytics, it distributes data storage and computing tasks across multiple data centers globally, reducing data transmission distance and latency. This ensures that large models can provide rapid response and real-time service to end users in various locations, breaking through the limitations of computing resources in a single region, reducing costs, improving computing resource utilization, and ensuring the flexibility and continuity of cross-border business. The integrated computing power and cross-border intelligent computing components specifically include a heterogeneous computing power supply engine and an intelligent computing power scheduling engine.

[0111] The heterogeneous computing power supply engine primarily utilizes a plug-in architecture to define standard resource access interfaces, supporting unified abstraction of multi-cloud, intelligent / supercomputing, and edge resources. It flexibly adapts to the management models and service differences of various computing power platforms, shielding underlying technical complexity through unified resource descriptions and interface specifications. The intelligent computing power scheduling engine collaboratively manages multiple resources across cloud, edge, and endpoint platforms. It comprehensively considers multiple factors such as task characteristics, resource status, network conditions, energy consumption costs, and data compliance. Relying on intelligent algorithms and multi-objective optimization mechanisms, it constructs a rapid-response, cost-controllable, and energy-optimized computing power scheduling system. It automatically selects the optimal execution location and path for computing power, providing users with seamless, cross-domain, end-to-end computing power services. It supports the efficient completion of large-scale computing, cloud-edge-endpoint collaboration, and other tasks, as well as global resource optimization, to meet diverse business scenario needs.

[0112] The system also includes a knowledge engineering and global data flywheel component, constructing a global multilingual knowledge base. It connects entity-relationship-attribute triples through a knowledge graph, enabling large models to understand the relationships between concepts. This supports downstream retrieval, question answering, and reasoning. Furthermore, based on data flywheel technology, it precipitates real-world online interaction data into new knowledge nodes, which are then fed back into the database, achieving dynamic expansion and accuracy improvement of knowledge assets, transforming scattered data into sustainably reusable intelligent assets. Specifically, the knowledge engineering and global data flywheel component includes a knowledge engineering engine and a global data flywheel engine.

[0113] The knowledge engineering engine employs a three-layer architecture: a translation management system, a knowledge graph, and retrieval-enhanced generation. The translation management system automatically translates source language content and writes it back to the corresponding language branch, while maintaining version tracking and quality verification rules to ensure consistency of the same knowledge across multiple languages. The knowledge graph semantically links regulatory nodes in the text with document nodes such as product models and test reports. Retrieval-enhanced generation, during the inference phase, retrieves nodes and document fragments from the knowledge graph in real time based on user questions, inputting the retrieved results as context into the large model to generate answers that conform to local language conventions and include regulatory basis. The global data flywheel engine uses a dual-ring parallel architecture: the inner ring performs online incremental learning, while the outer ring performs offline full-scale fine-tuning, progressively enhancing the multilingual recognition capabilities of the large model.

[0114] The system also includes a large model training and optimization component. Through corpus construction, targeted fine-tuning strategies, word segmentation optimization, and reinforcement learning mechanisms, it forms a complete closed loop of data, training, adaptation, and iteration, enabling the general model to have more accurate and efficient task processing capabilities in the cross-border e-commerce field.

[0115] The system also includes an agent decision-making and evaluation component, which integrates the reasoning capabilities of large models with external business systems, business data, and business processes, enabling models to move beyond simply answering questions and become perceptible, plannable, actionable, and retrospective around business objectives. This agent decision-making and evaluation component specifically includes an agent infrastructure framework, an agent toolchain, agent operations and maintenance, and agent evaluation.

[0116] The intelligent agent framework, through a closed-loop learning system comprising a perception layer, a decision-making layer, an execution layer, and a learning layer, enables the agent to perceive, plan, act, and reflect around business objectives. Based on real-time information acquired from the business environment, combined with its own knowledge base and reasoning capabilities, it formulates optimal action plans and invokes appropriate tools and resources to translate decisions into concrete actions. Simultaneously, it reflects and analyzes the execution results, continuously optimizing its decision-making model and behavioral strategies, thereby improving the accuracy and adaptability of its decisions. The intelligent agent toolchain supports the agent's decision-making and execution through multiple aspects, including connection protocols, tool registration and discovery mechanisms, execution sandboxes, data access and memory capabilities, and security and compliance control mechanisms. Intelligent agent operations and maintenance further productizes the intelligent agent framework and toolchain, forming a one-stop capability from design, testing, deployment to operational observation. In agent operation and maintenance, users can easily design the agent's behavior and decision-making logic, configure and adjust the agent through a visual interface, and access a wealth of testing tools and environments. Users can test the agent in different scenarios to verify its performance and stability. One-click deployment is supported, allowing for rapid deployment of the agent to real-world business environments. Users can also monitor the agent's operational status and performance metrics in real time, enabling timely problem identification and resolution. Agent evaluation establishes a consistent measurement loop across three levels: offline capabilities, online behavior, and end-to-end business results. This comprehensive and objective evaluation of agent performance provides a basis for optimization and improvement. Offline capability evaluation focuses on testing and analyzing the agent's model performance and inference capabilities; online behavior evaluation focuses on the agent's performance in real-world business environments, such as response speed and decision accuracy; and end-to-end business evaluation focuses on linking the agent's decisions to the ultimate business objectives, assessing the agent's contribution to those objectives.

[0117] The system also includes cross-border e-commerce scenario application components. By integrating intelligent technologies in links such as warehousing, logistics, and customer service, it realizes capabilities such as demand forecasting and inventory optimization, logistics path planning and inventory layout, compliance risk management, and multilingual support, improving the efficiency and accuracy of supply chain management, enhancing transportation timeliness, reducing logistics costs, operating costs, and risks, increasing the response speed and service quality of customer service, improving user experience and purchase conversion rate, and providing users with more intelligent, efficient, convenient, on-demand, and user-friendly services. At the same time, the intelligent agent connects with external business systems such as the ERP enterprise resource planning system, WMS warehouse management system, TMS transportation management system, payment and risk control system, etc., to complete long-chain tasks such as inventory optimization, price monitoring, contract generation, customs clearance and reconciliation settlement, and effectively control cross-border regulations and tax compliance by combining local knowledge bases and verifiable identities. The cross-border e-commerce scenario application components specifically include a supply chain intelligent management intelligent agent, an intelligent transportation and customs clearance intelligent agent, an intelligent customer service intelligent agent, and an intelligent recommendation intelligent agent.

[0118] Among them, the supply chain intelligent management intelligent agent realizes significant cost reduction, efficiency improvement, and risk control in supply chain management through intelligent algorithms and data fusion. In the supply chain link, it dynamically integrates multi-dimensional data such as global consumption trends, logistics status, and real-time environmental information, as well as historical sales data and market trends, etc., to build a prediction model, realizing demand forecasting and inventory optimization, and providing accurate inventory layout and replenishment strategies for overseas warehouses. In terms of compliance, according to the automatically updated multi-source regulation knowledge base, it automatically completes product certification, tax declaration, and data compliance review, realizing integrated management from risk warning to independent response. The intelligent transportation and customs clearance intelligent agent combines large model technology in the transportation and customs clearance link, connects with the data of logistics service providers in real time, plans the optimal path for transportation vehicles in real time, dynamically analyzes the timeliness, cost, and risk of the transportation route by monitoring transportation demand and road conditions in real time, calculates the path with the shortest transportation duration, automatically recommends the optimal solution, and can be adjusted in real time, greatly improving transportation timeliness, reducing logistics costs and delay rates. The intelligent customer service intelligent agent combines technologies such as machine learning, big data processing, natural language processing, and semantic analysis. According to various contents such as pictures, texts, and voices in the consumer's conversation, it identifies the consumer's consultation needs, and then uses a reply template to achieve rapid response, and can continuously self-adjust and optimize according to the consumer's consultation and feedback. At the same time, aiming at the language difference problem in cross-border transactions, it provides multilingual support, automatically identifies the language used by consumers, and uses the corresponding language for reply, and ensures the accuracy and fluency of translation through semantic analysis technology. The intelligent recommendation intelligent agent combines big data technology to monitor and predict consumers' behavior habits and preferences. By collecting and analyzing data such as consumers' browsing records, purchase history, and search keywords, it deeply analyzes users' behavior and preferences, accurately predicts consumers' needs, and provides users with accurate personalized product recommendations.

[0119] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0120] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in each of the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of each embodiment of the present invention.

Claims

1. A business processing system based on heterogeneous computing power, characterized in that, It includes computing power scheduling components, knowledge engineering components, large model components, and intelligent decision-making components, among which: The computing power scheduling component is used to construct a unified resource model, map heterogeneous computing power resources to standard computing units under the unified resource model, and allocate corresponding standard computing units to computing tasks in the knowledge engineering component, the large model component, and the intelligent decision-making component. The knowledge engineering component is used to construct a multilingual knowledge graph and extract incremental knowledge from business interaction data to update the multilingual knowledge graph. The large model component is used to train a general large model based on corpus data of the target domain to generate a vertical domain large model; The intelligent decision-making component is used to receive business instructions, call the vertical domain big model and combine it with the multilingual knowledge graph to perform reasoning, generate an executable decision flow; and call an external business system to perform specific operations according to the executable decision flow, and feed back the execution result data to the knowledge engineering component and the big model component.

2. The business processing system based on heterogeneous computing power according to claim 1, characterized in that, The computing power scheduling component includes a heterogeneous computing power supply engine, which is used for: Configure a standard resource access interface and load a computing power adaptation plugin to map different computing resource parameters to standard computing units in the unified resource model; and, By deploying a lightweight agent on edge devices, the resource idle status and network connection status of edge devices are detected, and the computing power of edge devices that meet the preset conditions are registered as edge computing service nodes in the unified resource model.

3. The business processing system based on heterogeneous computing power according to claim 1, characterized in that, The computing power scheduling component further includes an intelligent computing power scheduling engine, which is used for: Receive tasks to be processed and extract task feature data of the tasks to be processed; the task feature data includes computational intensity parameters and data compliance tags; Obtain the computing unit characteristic data of currently available standard computing units; the computing unit characteristic data includes resource status data, network connection data, energy consumption cost data, and data location identifier; Based on the task feature data and the computing unit feature data, the task to be processed is sent to a matching target available standard computing unit.

4. The business processing system based on heterogeneous computing power according to claim 1, characterized in that, The knowledge engineering component includes a global data flywheel engine, which is used for: Collect implicit business data and perform streaming training on the corresponding vertical domain large model based on the implicit business data; Collect explicit business data, perform privacy field filtering and multilingual semantic alignment processing on the explicit business data, and perform offline full-scale fine-tuning on the vertical domain large model based on the processed explicit business data.

5. The business processing system based on heterogeneous computing power according to claim 1, characterized in that, The knowledge engineering component further includes a knowledge engineering engine, the knowledge engineering engine being used for: The system calls the translation management system to translate source language data into text in the target language and writes it back to the corresponding language branch, as well as to identify the changed target text and perform incremental semantic alignment on the target text. Triples are extracted from the text based on the entity recognition model and stored in the multilingual knowledge graph; the triples include entities, relations and attributes. as well as, When the vertical domain big model receives an input query, it retrieves the associated nodes and document fragments that match the input query from the multilingual knowledge graph, and combines the associated nodes and document fragments into a prompt word context input to the vertical domain big model.

6. The business processing system based on heterogeneous computing power according to claim 1, characterized in that, The large model components include a corpus construction module, a targeted fine-tuning module, a word segmentation optimization module, and a reinforcement learning module, among which: The corpus construction module is used to construct corpus data for the target domain; The targeted fine-tuning module is used to fine-tune the parameters of the pre-trained general large model based on the corpus data of the target domain, and generate a vertical domain large model. The word segmentation optimization module is used to segment the corpus data of the target domain into words based on text segmentation rules; The reinforcement learning module is used to obtain reward signals based on feedback data in the business scenario, and to iteratively optimize the large model of the vertical domain based on the reward signals.

7. The business processing system based on heterogeneous computing power according to claim 1, characterized in that, The intelligent decision-making component includes an intelligent agent infrastructure module, which comprises a perception unit, a decision-making unit, an execution unit, and a learning unit, wherein: The sensing unit is used to receive and process multimodal data from the target domain and generate an environmental state representation; The decision-making unit is used to generate an executable decision flow corresponding to the business instruction based on the environmental state characterization. The execution unit is used to parse the executable decision flow, call the external business system to perform specific operations through a predefined interface protocol, and monitor the execution status; The learning unit is used to collect feedback data from external business systems and update the decision parameters of the decision unit based on the feedback data.

8. The business processing system based on heterogeneous computing power according to claim 7, characterized in that, The intelligent decision-making component also includes a toolchain module for: Establish a standard communication protocol between intelligent agents and external business systems, and maintain a tool registry to manage the calling interfaces and parameter configurations of external tools; as well as An isolated execution environment is established. Before the execution unit calls the external business system, the specific operation is pre-run in the isolated execution environment, and the specific operation is risk-verified and blocked based on the security policy.

9. The business processing system based on heterogeneous computing power according to claim 1, characterized in that, The intelligent decision-making component also includes an intelligent agent operation and maintenance module and an intelligent agent evaluation module, wherein: The intelligent agent operation and maintenance module is used to provide a visual orchestration interface to configure the workflow logic of the intelligent agent, and to perform deployment management and operation log analysis; The intelligent agent evaluation module is used to construct a multi-dimensional evaluation system and generate optimization suggestion data based on the evaluation results corresponding to the multi-dimensional evaluation system.

10. The business processing system based on heterogeneous computing power according to any one of claims 1 to 9, characterized in that, The system further includes an application execution component, which includes: The supply chain management intelligent agent is used to generate inventory optimization strategies based on historical sales data and market trend data, and to call the regulatory knowledge base to perform compliance verification operations. The intelligent customer service agent is used to identify the inquiry intent of multilingual users based on the multilingual knowledge graph and output response data in the corresponding language. The intelligent logistics agent is used to determine the optimal transportation route based on the real-time status data of the logistics service provider, and to adjust the optimal transportation route based on real-time road condition data. The intelligent recommendation agent is used to generate a product recommendation strategy based on the matching degree between user preference features and product features, and to recommend products according to the product recommendation strategy.

Images