Business process execution method and related equipment

Abstract

The invention provides a business process execution method and related equipment, and the method comprises the steps: carrying out the intention recognition of a business request, and obtaining a structured intention and key entity information; based on the structured intention, constructing a business process comprising a plurality of executable tasks; executing each executable task in the business process in sequence, and extracting a target cue word template matched with the structured intention from a cue word factory when a target task of a preset large model needs to be called in each executable task is executed; filling the key entity information into the target cue word template to obtain a parameterized cue word; and based on the parameterized cue word, calling a preset large model to execute the target task, and continuing to execute the remaining executable tasks in sequence after the target task is executed. In the scheme, dynamic binding of the cue word and the business scene is realized, and the accurately customized cue word can be automatically generated for different scenes, so that accurate adaptation to diversified business scenes in automatic execution of the business process is realized.

Description

A method for executing a business process and related equipment Technical Field

[0001] This invention relates to the field of computer technology, and more specifically to a method for executing a business process and related equipment. Background Technology

[0002] In traditional enterprise operations, business processes heavily rely on manual operation. This model is not only inefficient but also prone to errors due to human fatigue, making it difficult to guarantee process efficiency and consistency. To address this challenge, business process automation tools (such as RPA) have emerged. By simulating human operations at the user interface layer, they quickly automate various desktop tasks in a "non-intrusive" manner, significantly lowering the barrier to automation. With the rapid development of large language models, the capabilities of these models can be applied to specific business processes, allowing business process nodes to handle more complex tasks.

[0003] In existing technologies, business process automation tools rely on fixed prompts when calling large models, making it difficult to adapt to various business scenarios. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a method and related equipment for executing business processes, so as to achieve the purpose of accurate adaptation to diverse business scenarios in the automated execution of business processes.

[0005] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

[0006] This invention discloses a method for executing a business process, the method comprising:

[0007] Receive business requests input by the user;

[0008] The business request is subjected to intent recognition to obtain structured intent and key entity information;

[0009] Based on the structured intent, a business process containing multiple executable tasks is constructed;

[0010] Each executable task in the business process is executed sequentially. When executing the target task that requires calling the preset large model in each executable task, the target prompt word template that matches the structured intent is extracted from the pre-built prompt word factory. The prompt word factory stores prompt word templates corresponding to various structured intents.

[0011] The key entity information is filled into the target prompt word template to obtain parameterized prompt words;

[0012] Based on the parameterized prompts, a preset large model is invoked to execute the target task, and after the target task is completed, the remaining executable tasks are executed sequentially.

[0013] Optionally, the step of performing intent recognition on the service request to obtain structured intent and key entity information includes:

[0014] The business request is input into a pre-trained natural language understanding classification model for intent recognition, and the structured intent and key entity information output by the natural language understanding classification model are obtained.

[0015] Optionally, the step of constructing a business process containing multiple executable tasks based on the structured intent includes:

[0016] Extract a process blueprint that matches the structured intent from a pre-defined process template library; the process blueprint contains multiple service nodes.

[0017] Extract the metadata corresponding to each service node from the preset service node repository;

[0018] Based on the metadata corresponding to each service node, executable tasks corresponding to each service node are constructed to obtain a business process containing multiple executable tasks.

[0019] Optionally, the prompt word factory stores multiple versions of prompt word templates corresponding to various structured intents, and each version of the prompt word template is mapped to a comprehensive score that is updated periodically;

[0020] Accordingly, the step of extracting target prompt word templates that match the structured intent from the pre-built prompt word factory includes:

[0021] Extract multiple versions of prompt word templates corresponding to the structured intent from a pre-built prompt word factory;

[0022] From the multiple versions of prompt word templates corresponding to the structured intent, the prompt word template with the highest comprehensive score is selected as the target prompt word template for the structured intent adaptation.

[0023] Optionally, the method further includes:

[0024] Collect log data during the execution of the aforementioned business process;

[0025] The target prompt word template is scored based on predefined rules and the log data to obtain the current score;

[0026] Based on the current score, the overall score corresponding to the target prompt word template is updated.

[0027] A second aspect of this invention discloses a business process execution apparatus, the apparatus comprising:

[0028] The receiving unit is used to receive service requests input by the user;

[0029] The intent recognition unit performs intent recognition on the service request to obtain structured intent and key entity information;

[0030] A building unit is used to build a business process containing multiple executable tasks based on the structured intent;

[0031] An execution unit is used to sequentially execute each of the executable tasks in the business process, and when executing the target task that requires calling a preset large model in each of the executable tasks, it extracts the target prompt word template that matches the structured intent from the pre-built prompt word factory; the prompt word factory stores prompt word templates corresponding to various structured intents;

[0032] A parameterization unit is used to fill the key entity information into the target prompt word template to obtain parameterized prompt words;

[0033] The invocation unit is used to invoke a preset large model to execute the target task based on the parameterized prompt words, and to continue to execute the remaining executable tasks in sequence after the target task is completed.

[0034] Optionally, the intent recognition unit is specifically used for:

[0035] The business request is input into a pre-trained natural language understanding classification model for intent recognition, and the structured intent and key entity information output by the natural language understanding classification model are obtained.

[0036] Optionally, the building unit is specifically used for:

[0037] Extract a process blueprint that matches the structured intent from a pre-defined process template library; the process blueprint contains multiple service nodes.

[0038] Extract the metadata corresponding to each service node from the preset service node repository;

[0039] Based on the metadata corresponding to each service node, executable tasks corresponding to each service node are constructed to obtain a business process containing multiple executable tasks.

[0040] Optionally, the prompt word factory stores multiple versions of prompt word templates corresponding to various structured intents, and each version of the prompt word template is mapped to a comprehensive score that is updated periodically;

[0041] Accordingly, the execution unit used to extract multiple versions of prompt word templates corresponding to the structured intent from the pre-built prompt word factory is specifically used for:

[0042] From the multiple versions of prompt word templates corresponding to the structured intent, the prompt word template with the highest comprehensive score is selected as the target prompt word template for the structured intent adaptation.

[0043] Optionally, the device further includes:

[0044] The scoring unit is used to collect log data of the business process execution; score the target prompt word template based on predefined rules and the log data to obtain the current score; and update the comprehensive score corresponding to the target prompt word template based on the current score.

[0045] Based on the above-described embodiment of the present invention, a business process execution method and related equipment are provided, which receive a business request input by a user; perform intent recognition on the business request to obtain structured intent and key entity information; construct a business process containing multiple executable tasks based on the structured intent; sequentially execute each executable task in the business process, and when executing a target task that requires calling a preset large model in each executable task, extract a target prompt word template that matches the structured intent from a pre-built prompt word factory; the prompt word factory stores prompt word templates corresponding to various structured intents; fill the target prompt word template with the key entity information to obtain parameterized prompt words; based on the parameterized prompt words, call the preset large model to execute the target task, and continue to sequentially execute the remaining executable tasks after the target task is completed. In this solution, the business request is first parsed into structured intent and key entity information, then the most suitable target prompt word template is matched according to the identified intent, and finally the key entity information is filled to generate parameterized prompt words. This mechanism enables dynamic binding of prompt words to business scenarios, and can automatically generate precise and customized prompt words for different scenarios, thereby achieving accurate adaptation to diverse business scenarios in the automated execution of business processes. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present 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 only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0047] Figure 1 is a flowchart of a business process execution method disclosed in an embodiment of the present invention;

[0048] Figure 2 is an architecture diagram of a business process execution system disclosed in an embodiment of the present invention;

[0049] Figure 3 is a flowchart of the operation of a business process execution system disclosed in an embodiment of the present invention;

[0050] Figure 4 is a structural diagram of a business process execution device disclosed in an embodiment of the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] In this application, 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 limitation, 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.

[0053] As can be seen from the background technology, in the existing technology, business process automation tools rely on fixed prompt words to call large models, which makes it difficult to adapt to various business scenarios.

[0054] Therefore, this invention discloses a method and related equipment for executing a business process. In this solution, the business request is first parsed into structured intent and key entity information. Then, the most suitable target prompt word template is matched based on the identified intent. Finally, the key entity information is filled in to generate parameterized prompt words. This mechanism achieves dynamic binding between prompt words and business scenarios, enabling the automatic generation of precisely customized prompt words for different scenarios, thereby achieving accurate adaptation to diverse business scenarios in the automated execution of business processes.

[0055] Figure 1 shows a flowchart of a business process execution method disclosed in an embodiment of the present invention. The method mainly includes the following steps:

[0056] Step S101: Receive the service request input by the user.

[0057] In step S101, the user submits a business request through the application layer (such as a chat window or form). The business request is unstructured data.

[0058] For example, a user might enter the following business request: "Help me write the market summary section of a quarterly sales report, with detailed data."

[0059] Step S102: Perform intent recognition on the business request to obtain structured intent and key entity information.

[0060] In step S102, the business request is classified into a preset intent (such as #complaint#, #query#) using a trained NLU (Natural Language Understanding) classification model, and key entity information is extracted.

[0061] In the specific implementation of step S102, the business request is input into a pre-trained natural language understanding classification model for intent recognition, and the structured intent and key entity information output by the natural language understanding classification model are obtained.

[0062] For example, a business request might be: "Help me write the market summary section of a quarterly sales report, with detailed data." Intent recognition would then yield the corresponding structured intent and key entity information, such as: Structured intent = Write report section, Key entity information = Section name: Market Summary, Style: Detailed data.

[0063] Step S103: Based on the structured intent, construct a business process containing multiple executable tasks.

[0064] In step S103, based on the identified structured intent, the process template library and service node library are queried to dynamically generate an executable task sequence (business process).

[0065] In the specific implementation of step S103, a process blueprint adapted to the structured intent is extracted from a preset process template library. The process blueprint contains multiple service nodes. Metadata corresponding to each service node is extracted from a preset service node repository. Based on the metadata corresponding to each service node, executable tasks corresponding to each service node are constructed to obtain a business process containing multiple executable tasks.

[0066] The process blueprint adapted to the structured intent can be determined by a pre-built knowledge graph or rule engine.

[0067] It should be noted that the process template library is used to store predefined process blueprints in formats such as DAG (Directed Acyclic Graph), JSON, or YAML. These process blueprints are the basis for dynamically assembling business processes.

[0068] Service node repository: also known as the registry center, stores the metadata of all available atomic service nodes, including function descriptions, input / output formats, and calling addresses.

[0069] Step S104: Execute each executable task in the business process in sequence, and when executing the target task that requires calling the preset large model in each executable task, extract the target prompt word template that matches the structured intent from the pre-built prompt word factory.

[0070] The cue word factory stores cue word templates corresponding to various structured intents. Specifically, the cue word template that matches the structured intent can be determined through predefined rules.

[0071] It is understandable that when executing executable tasks that do not require calling large models, other external services (such as databases or RPA robots) are called to execute the corresponding executable tasks.

[0072] Optionally, the cue word factory stores multiple versions of cue word templates corresponding to various structured intents, and each version of the cue word template is mapped to a comprehensive score that is updated periodically.

[0073] Accordingly, target cue word templates adapted to structured intents are extracted from the pre-built cue word factory, including:

[0074] Extract multiple versions of cue word templates corresponding to the structured intent from the pre-built cue word factory; select the cue word template with the highest comprehensive score from the multiple versions of cue word templates corresponding to the structured intent as the target cue word template for structured intent adaptation.

[0075] It should be noted that the prompt word factory is responsible for the creation, storage, version control, parameterization configuration, and effect evaluation of prompt words.

[0076] Specifically, the prompt factory performs dynamic full lifecycle management of prompts, including upgrading traditional fixed and discrete prompts into manageable and measurable high-performance digital assets (prompt templates). That is, the prompt factory designs prompt templates with variables, supports A / B testing to achieve version management and establish a quantitative system for multi-dimensional evaluation, making them digital assets that can be adapted to multiple scenarios and continuously iterated.

[0077] The process of processing traditional fixed and discrete prompt words is as follows:

[0078] 1. Prompt word template: First, separate the "fixed core instruction" and "scenario-specific variation part" of the prompt word, then define the variation part as variables (such as user context, style requirements), and finally combine them into a reusable prompt word template.

[0079] 2. Prompt word factory management:

[0080] ① Version management, which means creating multiple versions (Ver A, Ver B, Ver C...) of the prompt word template corresponding to the structured intent. For example, for the "customer service reply" template, Ver A is more concise, while Ver B is more humanized.

[0081] ② A / B testing and multi-dimensional evaluation: In real-world scenarios, different versions of prompt word templates are periodically and randomly or selectively assigned to users for deployment. The entire process is monitored and logged, including inputs using the template version, variable population values, outputs from the large model, and performance metrics. The analyzer automatically calculates the overall score for each version, or it can be manually evaluated using annotations and rules. Simultaneously, based on the evaluation results, the prompt word factory may combine the advantages of VerA and VerB to create a more powerful VerC, which then proceeds to the next round of A / B testing.

[0082] Step S105: Fill the key entity information into the target prompt word template to obtain parameterized prompt words.

[0083] In step S105, key entity information is used as variables to fill the target prompt word template to obtain parameterized prompt words.

[0084] Step S106: Based on the parameterized prompt words, call the preset large model to execute the target task, and after the target task is completed, continue to execute the remaining executable tasks in sequence.

[0085] It should be noted that after the business process is completed, the business process execution result is obtained and then fed back to the user.

[0086] In one embodiment, log data of the business process execution is collected; the target prompt word template is scored based on predefined rules and the log data to obtain a current score; and the comprehensive score corresponding to the target prompt word template is updated based on the current score.

[0087] Specifically, the entire process is monitored and recorded, including complete log data for each execution, such as the version of the prompt words used, the inputs and outputs of each node, the final results, and performance data. This log data is stored in a structured manner, forming the basis for optimization analysis. Combining the collected log data, analysis is performed using predefined rules or algorithm models to update the overall score of the target prompt word template. The optimal version of the prompt word template is then reconfirmed, and the prompt word factory switches to the prompt word template that best matches the corresponding intent (with the highest overall score).

[0088] Based on the above-described business process execution method disclosed in this embodiment of the invention, this solution first parses the business request into structured intent and key entity information, then matches the most suitable target prompt word template according to the identified intent, and finally fills in the key entity information to generate parameterized prompt words. This mechanism realizes the dynamic binding of prompt words with business scenarios, and can automatically generate accurately customized prompt words for different scenarios, thereby achieving precise adaptation to diverse business scenarios in the automated execution of business processes.

[0089] Corresponding to the business process execution method disclosed in the above embodiments of the present invention, Figure 2 shows an architecture diagram of a business process execution system disclosed in an embodiment of the present invention.

[0090] (a) User / System Layer

[0091] The primary function of this layer is to serve as the system's sole entry point, handling all internal and external requests. Its main responsibilities include authentication, permission verification, request rate limiting, load balancing, and routing, ensuring system security and stability.

[0092] (ii) Intelligent Understanding and Assembly Layer

[0093] The central control unit acts as the brain of the platform, responsible for decision-making and coordinating various modules, including the intent recognition module, the context management module, and the process assembler.

[0094] 1. Intent Recognition Module: This module is mainly a pre-trained NLU (Natural Language Understanding) classification model that classifies user input into preset intents (such as #complaint#, #query#) and extracts key entity information.

[0095] 2. Context Management Module: Manages the context information of the session, maintains the state and variables throughout the process, and ensures that information is smoothly transmitted between different nodes.

[0096] 3. Process Assembler: Based on the identified intent and current context, it queries the process template library and dynamically generates an executable task sequence (process instance). Decision logic may be based on a graph or rule engine.

[0097] (III) Process Orchestration and Execution Engine

[0098] 1. Process Template Library: Stores predefined process blueprints in formats such as DAG (Directed Acyclic Graph), JSON, or YAML. These blueprints form the basis for dynamic assembly.

[0099] 2. Service Node Repository: The registry center stores the metadata of all available atomic service nodes, including function descriptions, input / output formats, and calling addresses.

[0100] 3. Workflow Execution Engine: The core executor, responsible for interpreting and executing the process instances generated by the assembler. It schedules nodes sequentially, manages dependencies, data flow, loops, conditional statements, and other logic between nodes, and maintains the execution status.

[0101] 4. Node Adapter: An abstraction layer that translates unified node call commands into different target services, such as gRPC, RESTful API, MQ messages, RPA clients, etc., to achieve seamless integration of the specific protocols and data formats required by the above-mentioned services.

[0102] (iv) Prompt word management engine

[0103] 1. Template Designer and Repository: Provides a UI or DSL (Domain-Specific Language) for professionals to design and store parameterized prompt word templates, and uses a Git-like mechanism for version control.

[0104] 2. Performance evaluator: Automatically scores the output of the large model through a rule engine (such as regular expression matching, keyword hit) or by calling an independent evaluation model, and also receives human scoring feedback.

[0105] 3. A / B Testing and Release Management: Manage the release strategy for the prompt word version, and conduct A / B testing and canary releases based on conditions such as percentage traffic and user tags.

[0106] 4. Performance Indicator Dashboard: Visually displays the success rate, time consumption, and business indicator conversion rate of different prompt word versions, providing an intuitive basis for decision-making.

[0107] (v) External Resources and Services

[0108] 1. Model Pool and Gateway: Connects to multiple large language models, provides a unified adaptation interface, and realizes model load balancing, failover, and cost optimization.

[0109] 2. Cue word renderer: Before calling the model, the parameterized cue word template is rendered into complete text that can be understood by the model based on the specific values ​​obtained from the context.

[0110] 3. Large language models: such as GPT-4, ChatGLM, etc., provide the core large model capabilities for the embodiments of this invention.

[0111] 4. External systems: These include various APIs, databases, RPA robots, etc., which are the objects of the final execution of the process.

[0112] Based on the above architecture, as shown in Figure 3, this is a flowchart of the operation of a business process execution system disclosed in an embodiment of the present invention. The main process of the system is as follows:

[0113] Step 1 (Process Trigger): The user submits a request through the application layer (such as a chat window or form).

[0114] The second step (intent recognition): The central controller's intent recognition module parses the user input and converts it into structured intents (#intent#) and entities ({entity}). If this fails, a clarification request is sent.

[0115] Step 3 (Dynamic Assembly): Based on the identified intent, the process assembler queries the process template library and service node library to dynamically assemble an optimal, executable task chain.

[0116] Step 4 (Execution and Invocation):

[0117] The scheduler executes the assembled processes sequentially. When a process requires large model capabilities, the scheduler requests a specific version of a parameterized prompt from the prompt factory. This prompt is then used to invoke the Large Language Model (LLM) and obtain the results. The process may also invoke other external services (such as databases and RPA robots).

[0118] Step 5 (Returning Results): Return the final result of the process execution to the user.

[0119] Step 6 (Log Recording): This step is crucial. The system will structurally save the entire data of this request, including the version of the prompt words used, the input and output of each node, the time taken, the final result, etc.

[0120] Step 7 (Analysis and Feedback): The optimization analyzer regularly analyzes log data and cross-analyzes it with indicators such as manual evaluation and business KPIs to assess the effectiveness of prompts and node strategies in this process.

[0121] Step 8 (Closed-Loop Optimization): The analyzer drives system iteration based on the analysis results, including prompt word optimization and process optimization. Prompt Word Optimization: Automatically switches prompt word traffic, disables ineffective versions, or triggers manual intervention. Process Optimization: Suggests or automatically adjusts process templates, such as optimizing node order and bypassing invalid nodes. Process Restart: The optimized prompt words and process strategies will be applied to subsequent user requests, thus starting a new, better loop.

[0122] Step 9 (Process Restart): The optimized prompts and process strategies will be applied to subsequent user requests, thereby starting a new and better loop.

[0123] This invention reduces system complexity and integration costs through deep collaboration between a prompt word management engine (prompt word factory) and a process orchestration engine. Simultaneously, a central linkage controller is introduced as the control center to unify the scheduling intent recognition, context management, process assembly, and execution feedback, avoiding performance bottlenecks and logical confusion caused by loose coupling between system modules, and achieving truly intelligent decision-making and coordination.

[0124] This invention employs a dual-pronged approach: full lifecycle management of dynamic prompt words and intent-driven dynamic process orchestration. In terms of full lifecycle management of dynamic prompt words, it achieves parameterization, versioning, and quantifiable effects. Parameterization and versioning transform prompt words from static text into dynamically configurable digital assets that support A / B testing and canary releases, enhancing adaptability to multiple business scenarios. Regarding quantifiable effects, a combination of a rule engine and manual evaluation enables objective assessment of different prompt word versions, ensuring that prompt word iteration is based on evidence and reducing over-reliance on expert experience. In intent-driven dynamic process orchestration, it primarily uses an NLU model to parse user intent in real time, moving away from predefined fixed processes and dynamically assembling service nodes, significantly improving system response flexibility and scenario coverage. Simultaneously, support from a service node repository and process template library enables "plug-and-play" and flexible reconfiguration of business processes, adapting to rapidly changing business needs.

[0125] This invention effectively implements a complete feedback mechanism for data closure. The system possesses a complete feedback loop, automatically extracting performance data from execution logs and automatically optimizing prompts and process strategies through an analyzer, achieving self-evolution capabilities. This forms a closed loop of "execution, feedback, optimization, and re-execution," significantly reducing the cost of manual intervention in later operation and maintenance. The adoption of a model pool and gateway design supports integration with various Large Language Models (LLMs), enabling load balancing, failover, and cost control, thereby improving system reliability and economy.

[0126] Corresponding to the business process execution method disclosed in the above embodiments of the present invention, Figure 4 shows a structural diagram of a business process execution device disclosed in an embodiment of the present invention, comprising:

[0127] The receiving unit 401 is used to receive service requests input by the user;

[0128] The intent recognition unit 402 performs intent recognition on the business request to obtain structured intent and key entity information;

[0129] Building unit 403 is used to build a business process containing multiple executable tasks based on structured intents;

[0130] The execution unit 404 is used to execute each executable task in the business process in sequence. When executing the target task of each executable task that requires calling the preset large model, it extracts the target prompt word template that matches the structured intent from the pre-built prompt word factory. The prompt word factory stores prompt word templates corresponding to various structured intents.

[0131] The parameterization unit 405 is used to fill key entity information into the target prompt word template to obtain parameterized prompt words;

[0132] Calling unit 406 is used to call a preset large model to execute the target task based on parameterized prompt words, and continue to execute the remaining executable tasks in sequence after the target task is completed.

[0133] Optionally, the intent recognition unit 402 is specifically used for:

[0134] The business request is input into a pre-trained natural language understanding classification model for intent recognition, and the structured intent and key entity information output by the natural language understanding classification model are obtained.

[0135] Optionally, building block 403 is specifically used for:

[0136] Extract process blueprints that match the structured intent from the pre-set process template library; the process blueprint contains multiple service nodes;

[0137] Extract the metadata corresponding to each service node from the preset service node repository;

[0138] Based on the metadata corresponding to each service node, executable tasks corresponding to each service node are constructed to obtain a business process containing multiple executable tasks.

[0139] Optionally, the cue word factory stores multiple versions of cue word templates corresponding to various structured intents, and each version of the cue word template maps to a comprehensive score that is updated periodically;

[0140] Correspondingly, the execution unit 404, which extracts multiple versions of prompt word templates corresponding to structured intents from the pre-built prompt word factory, is specifically used for:

[0141] From the multiple versions of cue word templates corresponding to the structured intent, the cue word template with the highest comprehensive score is selected as the target cue word template for structured intent adaptation.

[0142] Optionally, the device may also include:

[0143] The scoring unit is used to collect log data during the execution of business processes; to score the target prompt word template based on predefined rules and log data, and to obtain the current score; and to update the comprehensive score corresponding to the target prompt word template based on the current score.

[0144] Based on the business process execution device disclosed in the above embodiments of the present invention, this solution first parses the business request into structured intent and key entity information, then matches the most suitable target prompt word template according to the identified intent, and finally fills in the key entity information to generate parameterized prompt words. This mechanism realizes the dynamic binding of prompt words with business scenarios, and can automatically generate accurately customized prompt words for different scenarios, thereby achieving accurate adaptation to diverse business scenarios in the automated execution of business processes.

[0145] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and relevant parts can be referred to the descriptions in the method embodiments. The systems and system 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 creative effort.

[0146] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0147] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for executing a business process, characterized in that, The method includes: receiving a business request input by a user; performing intent recognition on the business request to obtain a structured intent and key entity information; constructing a business process containing multiple executable tasks based on the structured intent; sequentially executing each executable task in the business process, and when executing a target task in each executable task that requires calling a preset large model, extracting a target prompt word template adapted to the structured intent from a pre-built prompt word factory; the prompt word factory stores prompt word templates corresponding to various structured intents; filling the target prompt word template with the key entity information to obtain parameterized prompt words; calling the preset large model to execute the target task based on the parameterized prompt words, and continuing to sequentially execute the remaining executable tasks after the target task is completed.

2. The method according to claim 1, characterized in that, The step of performing intent recognition on the business request to obtain structured intent and key entity information includes: inputting the business request into a pre-trained natural language understanding classification model for intent recognition, and obtaining the structured intent and key entity information output by the natural language understanding classification model.

3. The method according to claim 1, characterized in that, The step of constructing a business process containing multiple executable tasks based on the structured intent includes: extracting a process blueprint adapted to the structured intent from a preset process template library; the process blueprint contains multiple service nodes; extracting metadata corresponding to each service node from a preset service node repository; and constructing executable tasks corresponding to each service node based on the metadata corresponding to each service node, thereby obtaining a business process containing multiple executable tasks.

4. The method according to claim 1, characterized in that, The prompt word factory stores multiple versions of prompt word templates corresponding to various structured intents. Each version of the prompt word template is mapped to a periodically updated comprehensive score. Accordingly, the step of extracting a target prompt word template that matches the structured intent from the pre-built prompt word factory includes: extracting multiple versions of prompt word templates corresponding to the structured intent from the pre-built prompt word factory; and selecting the prompt word template with the highest comprehensive score from the multiple versions of prompt word templates corresponding to the structured intent as the target prompt word template for the structured intent.

5. The method according to claim 4, characterized in that, The method further includes: collecting log data of the business process execution; scoring the target prompt word template based on predefined rules and the log data to obtain a current score; and updating the comprehensive score corresponding to the target prompt word template based on the current score.

6. A business process execution device, characterized in that, The device includes: a receiving unit for receiving a service request input by a user; an intent recognition unit for recognizing the intent of the service request to obtain a structured intent and key entity information; a construction unit for constructing a business process containing multiple executable tasks based on the structured intent; an execution unit for sequentially executing each executable task in the business process, and when executing a target task in each executable task that requires calling a preset large model, extracting a target prompt word template adapted to the structured intent from a pre-built prompt word factory; the prompt word factory stores prompt word templates corresponding to various structured intents; a parameterization unit for filling the key entity information into the target prompt word template to obtain parameterized prompt words; and a calling unit for calling a preset large model to execute the target task based on the parameterized prompt words, and continuing to sequentially execute the remaining executable tasks after the target task is completed.

7. The apparatus according to claim 6, characterized in that, The intent recognition unit is specifically used to: input the business request into a pre-trained natural language understanding classification model for intent recognition, and obtain the structured intent and key entity information output by the natural language understanding classification model.

8. The apparatus according to claim 6, characterized in that, The construction unit is specifically used to: extract a process blueprint that matches the structured intent from a preset process template library; the process blueprint contains multiple service nodes; and extract the metadata corresponding to each service node from a preset service node repository. Based on the metadata corresponding to each service node, executable tasks corresponding to each service node are constructed to obtain a business process containing multiple executable tasks.

9. The apparatus according to claim 6, characterized in that, The prompt word factory stores multiple versions of prompt word templates corresponding to various structured intents. Each version of the prompt word template is mapped to a periodically updated comprehensive score. Correspondingly, the execution unit used to extract multiple versions of prompt word templates corresponding to the structured intent from the pre-built prompt word factory is specifically used to: select the prompt word template with the highest comprehensive score from the multiple versions of prompt word templates corresponding to the structured intent as the target prompt word template for the structured intent adaptation.

10. The apparatus according to claim 9, characterized in that, The device further includes: a scoring unit, used to collect log data of the business process execution; score the target prompt word template based on predefined rules and the log data to obtain a current score; and update the comprehensive score corresponding to the target prompt word template based on the current score.

Images