AI agent building training method and system in work interaction scene

Abstract

The invention is suitable for the technical field of AI agent training, and particularly relates to an AI agent building training method and system in a work interaction scenario, and the method comprises the steps: recognizing access posts of an AI agent, configuring a work flow of each access post, generating a flow chain, clustering the access posts into a plurality of categories, and constructing a plurality of AI agent construction training modes; initial data in each category is collected, attribute information is configured, and the attribute information at least comprises a type and a source; and creating branch nodes in one-to-one correspondence with the categories. According to the method, the use scene of the AI intelligent agent can be greatly expanded, the generalization ability can be enhanced and the training efficiency can be improved by constructing the universal training set, the service fitting degree of the AI intelligent agent can be remarkably improved by constructing the optimized training set, the differentiation ability of the AI intelligent agent can be improved, the customization requirement under multiple scenes can be met, and the application prospect is wide. The understanding ability of the AI intelligent agent for specific users, services or scenes is improved, and the use experience is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of AI agent training, and particularly to a method and system for building and training an AI agent in a work interaction scenario. Background Art

[0002] In a work interaction scenario, an AI agent is an artificial intelligence system that can simulate a "digital employee", with the ability to understand business content, actively collaborate, and complete tasks. It can assist employees in various positions in the office process to improve work efficiency. For example, using an AI agent to arrange meetings, screen resumes, and file materials, etc.; however, in the construction and training of AI agents, there are still deficiencies in customization, and it is impossible to perform personalized training on AI agents according to different positions, which also makes AI agents lack dynamic adaptability.

[0003] Therefore, "how to pre-train an AI agent using general data and fine-tune it using personalized data" is the technical problem to be solved by the present invention. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for building and training an AI agent in a work interaction scenario to solve the problems raised in the above background art.

[0005] To achieve the above purpose, the present invention provides the following technical solutions: A method and system for building and training an AI agent in a work interaction scenario, the method includes: Identifying the access positions of the AI agent, configuring the work processes of each access position, generating a process chain, clustering the access positions into several categories, collecting the initial data in each category, and configuring the attribute information, where the attribute information at least includes: type and source; Creating branch nodes corresponding to the categories one by one, calculating the number of access positions in each category, mounting the same number of sub-nodes to the branch nodes, and integrating the branch nodes and sub-nodes to generate a training tree; Marking the process chain in the training tree. When the initial data uploaded by the access position is collected, based on the process chain, marking the transfer position of the initial data, and defining the sub-node corresponding to the transfer position as the target node, sending an information collection pop-up window to the user terminal corresponding to the target node to obtain the changed version of the initial data, establishing a correspondence between the target node and the changed version, collecting the feedback of the target node, and writing it into the changed version, integrating the initial data and the changed version to generate a general training set, and training the AI agent; Receive the work interaction requirements uploaded by the access positions, find the initial data of the same type based on the attribute data, select the excellent versions from the corresponding changed versions, generate an optimized training set using the initial data of the same type and the excellent versions, and fine-tune the AI agent.

[0006] Further, the steps of identifying the access positions of the AI agent and configuring the work processes for each access position include: Set the work types for each access position, and configure the transfer relationships between the access positions based on the work processes; Integrate the access positions and the transfer relationships to generate a process chain, where each work type corresponds to a process chain.

[0007] Further, the steps of generating the process chain, clustering the access positions into several categories, and collecting the initial data in each category include: Insert tags generated by the source into the initial data; Integrate all the process chains to obtain a chain group, and construct a backtracking mechanism based on the tags and the chain group.

[0008] Further, the steps of integrating the branch nodes and the child nodes to generate a training tree include: Calculate the update frequencies of each child node and branch node, delete the child nodes and branch nodes with update frequencies less than the threshold, and reconstruct the training tree; Open the data interface of the AI agent to the training tree and build a communication link.

[0009] Further, the steps of sending an information collection pop-up window to the user terminal corresponding to the target node to obtain the changed versions of the initial data include: Define the processed initial data as the changed versions, where each target node corresponds to a changed version; Traverse the process chain, identify the transfer order, and integrate all the changed versions to obtain a version set.

[0010] Further, the steps of selecting the excellent versions from the corresponding changed versions include: Extract the feedback opinions in the changed versions and input them into a pre-constructed natural language analysis model to output semantic features; Query the preset sentiment dictionary, identify the sentiment values of each semantic feature, and calculate the total sentiment value of each feedback opinion; Sort all the changed versions in descending order according to the total, generate a queue, and select a preset number of changed versions from the front of the queue and define them as excellent versions.

[0011] Further, the method further includes: Configure the trigger rule for optimizing the training set, where the trigger rule is: when the AI agent receives a work interaction requirement, input the optimized training set into the AI agent and perform training and fine-tuning.

[0012] Furthermore, the system includes: A configuration module, used to identify the access positions of the AI agents, configure the work processes of each access position, generate a process chain, cluster the access positions into several categories, collect the initial data in each category, and configure the attribute information, where the attribute information at least includes: type and source; A generation module, used to create branch nodes corresponding one by one to the categories, calculate the number of access positions in each category, mount the same number of child nodes to the branch nodes, integrate the branch nodes and child nodes, and generate a training tree; A general training module, used to mark the process chain in the training tree. When the initial data uploaded by the access position is collected, based on the process chain, mark the transfer position of the initial data, and define the child node corresponding to the transfer position as the target node. Send an information collection pop-up window to the user terminal corresponding to the target node, obtain the changed version of the initial data, establish the corresponding relationship between the target node and the changed version, collect the feedback opinions of the target node, and write them into the changed version. Integrate the initial data and the changed version, generate a general training set, and train the AI agent; An optimization training module, used to receive the work interaction requirements uploaded by the access position, based on the attribute data, find the initial data of the same type, select the excellent versions from the corresponding changed versions, use the initial data of the same type and the excellent versions to generate an optimized training set, and fine-tune the AI agent.

[0013] Furthermore, the configuration module includes: A setting unit, used to set the work types of each access position, and configure the transfer relationship between the access positions based on the work process; An integration unit, used to integrate the access positions and the transfer relationship to generate a process chain, where each work type corresponds to a process chain; An insertion unit, used to insert the label generated by the source into the initial data; A construction unit, used to integrate all the process chains to obtain a chain group, and construct a backtracking mechanism based on the label and the chain group.

[0014] Furthermore, the generation module includes: A reconstruction unit, used to calculate the update frequencies of each child node and branch node, delete the child nodes and branch nodes with update frequencies less than the threshold, and reconstruct the training tree; A building unit for opening the data interface of the AI agent to the training tree and building a communication link.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: By determining the initial data, the present invention can clarify the training objectives and application boundaries, improve the generalization ability of the AI agent, and provide a data basis for the training of the AI agent. By constructing a process chain, it can record the processing process of each initial data, ensure the traceability of the initial data, improve the data management level, and facilitate the iterative training of the AI agent. By constructing a training tree, it can clearly show the data evolution process and achieve the continuous optimization of the AI agent. By constructing a general training set, it can greatly expand the usage scenarios of the AI agent, enhance the generalization ability, and improve the training efficiency. By constructing an optimized training set, it can significantly improve the business compliance of the AI agent, help improve the differentiation ability of the AI agent, meet the customization requirements in multiple scenarios, enhance the understanding ability of the AI agent for specific users, businesses or scenarios, and enhance the user experience. Description of the Drawings

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention.

[0017] Figure 1 It is a flow chart of the method for building and training an AI agent in a work interaction scenario provided by an embodiment of the present invention; Figure 2 It is a first sub-flow chart of the method for building and training an AI agent in a work interaction scenario provided by an embodiment of the present invention; Figure 3 It is a second sub-flow chart of the method for building and training an AI agent in a work interaction scenario provided by an embodiment of the present invention; Figure 4 It is a third sub-flow chart of the method for building and training an AI agent in a work interaction scenario provided by an embodiment of the present invention; Figure 5 It is a fourth sub-flow chart of the method for building and training an AI agent in a work interaction scenario provided by an embodiment of the present invention; Figure 6 It is a block diagram of the composition of the AI agent building and training system in a work interaction scenario provided by an embodiment of the present invention; Figure 7 It is a block diagram of the composition of the configuration module in the AI agent building and training system in a work interaction scenario provided by an embodiment of the present invention; Figure 8It is a block diagram of the composition of the generation module in the AI intelligent agent construction and training system in the work interaction scenario provided by the embodiment of the present invention; Figure 9 It is a block diagram of the composition of the general training module in the AI intelligent agent construction and training system in the work interaction scenario provided by the embodiment of the present invention; Figure 10 It is a block diagram of the composition of the optimization training module in the AI intelligent agent construction and training system in the work interaction scenario provided by the embodiment of the present invention. Detailed implementation manners

[0018] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0019] Figure 1 It is the overall flowchart of the AI intelligent agent construction and training method and system in the work interaction scenario. In the embodiment of the present invention, an AI intelligent agent construction and training method in the work interaction scenario, the method includes: S100: Identify the access positions of the AI intelligent agent, configure the work processes of each access position, generate a process chain, cluster the access positions into several categories, collect the initial data in each category, and configure the attribute information, where the attribute information at least includes: type and source.

[0020] Identify all the positions that use the AI intelligent agent, that is, the access positions, such as customer service, human resources, and business, etc., and configure the corresponding standardized work processes for each access position; for example, after the customer service position receives a user request (product maintenance request), it sends the user request to the technical position, and the technical position formulates a maintenance plan, and then the logistics position retrieves the accessories according to the spare part list provided by the technical position, and the warehousing position packs and ships. After the customer service position receives the spare part shipment notice, it feedbacks to the user. This is also a work process composed of multiple access positions; simplify the work process and store it in a chain structure to obtain a process chain; for example, the process chain corresponding to the above work process is: customer service sends a request - technical position formulates a maintenance plan - logistics position retrieves accessories - warehousing position packs and ships - customer service position feedbacks.

[0021] All job positions are classified into multiple categories according to the type of each position or the work content, etc., such as the customer service category, the technical category, and the warehousing category, etc.; collect the initial data in each category, where the initial data is the existing data in each access position, and determine the type and source of each existing data, etc.; further, the type of existing data refers to the specific form and content category of the data, such as text, voice, image, structured table, and log file, etc., and the source refers to the data acquisition channel or system For example, all the existing data in the customer service position is called initial data, and the initial data includes: user questions, processing lists, and user satisfaction data, etc.

[0022] S200: Create branch nodes corresponding one-to-one with the categories, calculate the number of access positions in each category, mount the same number of child nodes into the branch nodes, and integrate the branch nodes and child nodes to generate a training tree.

[0023] Create a corresponding branch node for each category, where the branch node is a logical node, not an entity structure, count the number of access positions in each category, and generate the same number of child nodes, where the child nodes are also logical nodes and do not perform data processing. Each child node corresponds to a job person, mount the child nodes into the branch nodes corresponding to the categories, and integrate the branch nodes and child nodes to generate a training tree; the training tree is a tree-like data structure, only used to show the data flow relationship between different access positions.

[0024] S300: Mark the process chain into the training tree. When the initial data uploaded by the access position is collected, based on the process chain, mark the transfer position of the initial data, and define the child node corresponding to the transfer position as the target node. Send an information collection pop-up window to the user terminal corresponding to the target node to obtain the changed version of the initial data, establish the corresponding relationship between the target node and the changed version, collect the feedback opinions of the target node, and write them into the changed version. Integrate the initial data and the changed version to generate a general training set, and train the AI agent.

[0025] Each child node corresponds to a specific job operation in the process chain. Mark the process chain into the training tree, determine the transfer position according to the transfer process of the initial data, and define the corresponding child node in the training tree as the target node.

[0026] For example, the customer service staff receives a product maintenance request from the user, determines this product maintenance request as the initial data, the customer service staff writes the product maintenance request, specific product information, etc. into the maintenance request registration form, and sends it to the technical staff. The technical staff writes a delivery note and sends it to the logistics staff; during this process, the transfer positions include: the customer service staff, the technical staff, and the logistics staff; the corresponding child nodes of the three are the target nodes.

[0027] Determine the user terminal according to the user information corresponding to each child node, and send an information collection pop-up window to the corresponding user terminal to obtain the changed version of the initial data; establish the corresponding relationship between the target node and the changed version according to the flow of the initial data; when the changed version is generated, collect the feedback of the target node through the user terminal, where the feedback can be problem description, satisfaction score, and improvement suggestions; write the feedback into the changed version, and use the initial data and the changed version to generate a general training set, where the general training set is a set composed of several changed versions.

[0028] For example, in the above process, the changed versions corresponding to the product maintenance request include: maintenance request registration form, bill of lading, etc.; when the customer service staff completes the maintenance request registration form, it is evaluated by the technical staff to obtain feedback, and so on. Each changed version corresponds to a feedback, and the set composed of all the changed versions and feedback shows the flow process of the initial data; use this set as the pre-training data of the AI agent, so that the AI agent can automatically complete the extraction of user requests and operations such as maintenance request registration form and bill of lading, and the staff in the corresponding position only needs to review, thus reducing the workload of the staff.

[0029] S400: Receive the work interaction requirements uploaded by the access position, find the same type of initial data based on the attribute data, select the excellent versions from the corresponding changed versions, generate an optimized training set using the same type of initial data and the excellent versions, and fine-tune the AI agent.

[0030] When the staff uses the AI agent, initiate work interaction requirements, find all existing bills of lading and the corresponding initial data, and find the parts with better feedback according to the feedback, which are defined as excellent versions, integrate the excellent versions and the corresponding initial data, generate an optimized training set, and train the AI agent.

[0031] For example, the technical staff initiates the work interaction requirement of "filling out the bill of lading", extracts all existing bills of lading and the corresponding user request data from the technical position, and finds the bills of lading that are standard-compliant and information-complete according to the feedback, and inputs them into the AI agent to optimize the AI agent.

[0032] In Embodiment 2, Figure 2 The implementation process of the AI agent construction training method in the work interaction scenario provided by the embodiment of the present invention is shown. The steps of identifying the access position of the AI agent and configuring the work process of each access position are described in detail as follows: S101: Set the work types of each access position, and configure the transfer relationships between the access positions based on the work process.

[0033] Determine the work types of each work position. For example, set the customer service position as the work type of "customer request access and distribution", the technical position as the work type of "problem diagnosis and professional handling", and the logistics position as the work type of "material allocation and confirmation"; configure the data and task transfer relationships between the access positions, and clarify the transfer path of the initial data from one position to the next; for example, the transfer relationship corresponding to the user maintenance appointment is: customer service - technical - logistics.

[0034] S102: Integrate the access positions and transfer relationships to generate a process chain, where each work type corresponds to a process chain.

[0035] Use the work content and transfer relationships of the access positions to generate a process chain, and each work type corresponds to a process chain; for example, the process chains corresponding to the user maintenance appointment and the user replacement are different.

[0036] In Embodiment 3, Figure 2 Illustrates the implementation process of the AI intelligent agent construction and training method in the work interaction scenario provided by the embodiment of the present invention. The following details the steps of generating the process chain, clustering the access positions into several categories, and collecting the initial data in each category, as follows: S103: Insert tags generated by the source into the initial data.

[0037] Determine the source of the initial data and generate tags; the tags can be presented in ways such as links and remarks; the source can be user ID, channel, platform, etc.

[0038] S104: Integrate all the process chains to obtain a chain group, and construct a backtracking mechanism based on the tags and the chain group.

[0039] Integrate all the process chains to obtain a chain group, where the chain group is a set composed of process chains, and the backtracking mechanism is: restore the complete transfer path of the initial data in the chain group, and clarify which position, which person processes at each stage and what operations are performed.

[0040] In Embodiment 4, Figure 3 Illustrates the implementation process of the AI intelligent agent construction and training method in the work interaction scenario provided by the embodiment of the present invention. The following details the steps of integrating the branch nodes and sub-nodes to generate a training tree, as follows: S201: Calculate the update frequency of each sub-node and branch node, delete the sub-nodes and branch nodes whose update frequency is less than the threshold, and reconstruct the training tree.

[0041] Calculate the update frequencies of child nodes and branch nodes. For example, the frequency of using the AI agent for the cleaning position is relatively low and the requirement for professionalism is also relatively low. The corresponding child nodes and branch nodes can be deleted, and the generalization ability of the AI agent can be relied on to meet the usage requirements of the cleaning position.

[0042] S202: Open the data interface of the AI agent to the training tree and establish a communication link.

[0043] Open the data interface to the training tree and establish a stable communication link to ensure that the training tree can use the latest data to train the AI agent.

[0044] In Embodiment 5, Figure 4 The implementation process of the AI agent construction and training method in the work interaction scenario provided by the embodiment of the present invention is shown. The following details the step of sending an information collection pop-up window to the user terminal corresponding to the target node and obtaining the changed version of the initial data, as follows: S301: Define the processed initial data as the changed version, where each target node corresponds to a changed version.

[0045] According to the transfer process exemplified in S300, define the transferred initial data as the changed version.

[0046] S302: Traverse the process chain, identify the transfer order, and integrate all the changed versions to obtain a version set.

[0047] According to the transfer process of the initial data, combined with the process chain, determine the transfer order, and integrate all the changed versions according to the transfer order to obtain a unified version set; the version set not only retains the evolution process of the initial data at different stages but also reflects the decision-making basis and processing methods of different positions, thus providing a rich data basis for the training of the AI agent.

[0048] In Embodiment 6, Figure 5 The implementation process of the AI agent construction and training method in the work interaction scenario provided by the embodiment of the present invention is shown. The following details the step of selecting excellent versions from the corresponding changed versions, as follows: S401: Extract the feedback opinions in the changed version and input them into the pre-constructed natural language analysis model to output semantic features.

[0049] Input the feedback opinions into the natural language analysis model to output semantic features, which include but are not limited to emotion labels (such as positive, neutral, negative), quality evaluation, data problems, and response timeliness, etc.

[0050] S402: Query the preset sentiment dictionary, identify the sentiment values of each semantic feature, and calculate the total sentiment value of each feedback opinion.

[0051] Compare the sentiment dictionary composed of several semantic features, determine the sentiment value of each semantic feature, and calculate the total sentiment value.

[0052] S403: Sort all the changed versions in descending order according to the total sum, generate a queue, and select a preset number of changed versions from the front of the queue and define them as excellent versions.

[0053] Sort the changed versions in descending order according to the total sentiment value to generate a queue; according to the corresponding relationship between the sentiment value and the changed version, select a preset number of changed versions from the front of the queue and define them as excellent versions.

[0054] In Embodiment 7, different from Embodiment 1, in the embodiment of the present invention, the method further includes: configuring a trigger rule for optimizing the training set, where the trigger rule is: when the AI agent receives a work interaction requirement, input the optimized training set into the AI agent and perform training and fine-tuning.

[0055] In the pre-training of the AI agent, only the general training set is used for training. When the staff generates a new work interaction requirement, the optimized training set is used to fine-tune the AI agent.

[0056] In actual operation, when pre-training the AI agent, all data is used for training. After the staff puts forward a specific usage requirement, the same type of data is found to perform special training and optimization on the AI agent.

[0057] Figure 6 The composition structure block diagram of the AI agent construction training system in the work interaction scenario provided by the embodiment of the present invention is shown. The AI agent construction training system 1 in the work interaction scenario includes: A configuration module 11, used to identify the access positions of the AI agent, configure the work processes of each access position, generate a process chain, cluster the access positions into several categories, collect the initial data in each category, and configure the attribute information, where the attribute information at least includes: type and source; A generation module 12, used to create branch nodes corresponding to the categories one by one, calculate the number of access positions in each category, mount the same number of child nodes to the branch nodes, and integrate the branch nodes and child nodes to generate a training tree; The general training module 13 is used to mark the process chain into the training tree. After the initial data uploaded by the access position is collected, based on the process chain, mark the transfer position of the initial data, define the corresponding child node of the transfer position as the target node, send an information collection pop-up window to the user terminal corresponding to the target node, obtain the changed version of the initial data, establish the correspondence between the target node and the changed version, collect the feedback of the target node, and write it into the changed version, integrate the initial data and the changed version, generate a general training set, and train the AI agent; The optimization training module 14 is used to receive the work interaction requirements uploaded by the access position, find out the same type of initial data based on the attribute data, select the excellent version from the corresponding changed version, generate an optimization training set by using the same type of initial data and the excellent version, and fine-tune the AI agent.

[0058] Figure 7 The block diagram of the composition of the AI agent construction and training system in the work interaction scenario provided by the embodiment of the present invention is shown. The configuration module 11 includes: The setting unit 111 is used to set the work type of each access position and configure the transfer relationship between access positions based on the work process; The integration unit 112 is used to integrate the access position and the transfer relationship to generate a process chain, where each work type corresponds to a process chain; The insertion unit 113 is used to insert the label generated by the source into the initial data; The construction unit 114 is used to integrate all process chains to obtain a chain group, and construct a backtracking mechanism based on the label and the chain group.

[0059] Figure 8 The block diagram of the composition of the AI agent construction and training system in the work interaction scenario provided by the embodiment of the present invention is shown. The generation module 12 includes: The reconstruction unit 121 is used to calculate the update frequency of each child node and branch node, delete the child nodes and branch nodes with the update frequency less than the threshold, and reconstruct the training tree; The construction unit 122 is used to open the data interface of the AI agent to the training tree and construct a communication link.

[0060] Figure 9 The block diagram of the composition of the AI agent construction and training system in the work interaction scenario provided by the embodiment of the present invention is shown. The general training module 13 includes: The definition unit 131 is used to define the processed initial data as the changed version, where each target node corresponds to a changed version An obtaining unit 132 is configured to traverse the process chain, identify the transfer sequence, and integrate all changed versions to obtain a version set.

[0061] Figure 10 The following shows a block diagram of the composition of an AI intelligent agent construction and training system in a work interaction scenario provided by an embodiment of the present invention. The optimization training module 14 includes: An extraction unit 141 is configured to extract feedback opinions in the changed versions and input them into a pre-constructed natural language analysis model to output semantic features. A query unit 142 is configured to query a preset sentiment dictionary, identify the sentiment value of each semantic feature, and calculate the total sum of the sentiment values of each feedback opinion.

[0062] Among them, the configuration module 11 is mainly used to complete step S100, the generation module 12 is mainly used to complete step S200, the general training module 13 is mainly used to complete step S300, and the optimization training module 14 is mainly used to complete step S400. A setting unit 111 is mainly used to complete step S101, an integration unit 112 is mainly used to complete step S102, an insertion unit 113 is mainly used to complete step S103, and a construction unit 114 is mainly used to complete step S104. A reconstruction unit 121 is mainly used to complete step S201, and a construction unit 122 is mainly used to complete step S202. A definition unit 131 is mainly used to complete step S301, and an obtaining unit 132 is mainly used to complete step S302. An extraction unit 141 is mainly used to complete step S401, and a query unit 142 is mainly used to complete step S402.

[0063] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A method for building and training an AI intelligent agent in a work interaction scenario, characterized in that, The method includes: Identifying the access positions of the AI agents, configuring the work processes for each access position, generating a process chain, clustering the access positions into several categories, collecting the initial data in each category, and configuring the attribute information, where the attribute information at least includes: type and source; Creating branch nodes corresponding one-to-one with the categories, calculating the number of access positions in each category, mounting the same number of child nodes to the branch nodes, integrating the branch nodes and the child nodes, and generating a training tree; Marking the process chain in the training tree. When the initial data uploaded by the access position is collected, based on the process chain, marking the transfer position of the initial data, and defining the child node corresponding to the transfer position as the target node, sending an information collection pop-up window to the user terminal corresponding to the target node to obtain the changed version of the initial data, establishing the correspondence between the target node and the changed version, collecting the feedback opinions of the target node, and writing them into the changed version, integrating the initial data and the changed version, generating a general training set, and training the AI agent; Receiving the work interaction requirements uploaded by the access position, based on the attribute data, finding the initial data of the same type, selecting the excellent versions from the corresponding changed versions, generating an optimized training set using the initial data of the same type and the excellent versions, and fine-tuning the AI agent.

2. The AI intelligent agent construction and training method in the work interaction scenario according to claim 1, characterized in that The steps of identifying the access positions of the AI agents and configuring the work processes for each access position include: Setting the work type of each access position, and configuring the transfer relationship between the access positions based on the work process; Integrating the access positions and the transfer relationship to generate a process chain, where each work type corresponds to a process chain.

3. The AI intelligent agent construction and training method in the work interaction scenario according to claim 1, characterized in that The steps of generating the process chain, clustering the access positions into several categories, and collecting the initial data in each category include: Inserting the label generated by the source into the initial data; Integrating all the process chains to obtain a chain group, and constructing a backtracking mechanism based on the label and the chain group.

4. The AI intelligent agent construction and training method in the working interaction scenario according to claim 1, characterized in that, The steps of integrating the branch nodes and the child nodes to generate a training tree include: Calculating the update frequency of each child node and branch node, deleting the child nodes and branch nodes with the update frequency less than the threshold, and reconstructing the training tree; Opening the data interface of the AI agent to the training tree and building a communication link.

5. The AI intelligent agent construction and training method in the work interaction scenario according to claim 1, characterized in that, The steps of sending an information collection pop-up window to the user terminal corresponding to the target node to obtain the changed version of the initial data include: Defining the processed initial data as the changed version, where each target node corresponds to a changed version; Traversing the process chain, identifying the transfer order, and integrating all the changed versions to obtain a version set.

6. The AI intelligent agent construction and training method in the work interaction scenario according to claim 5, characterized in that, The steps of selecting the excellent versions from the corresponding changed versions include: Extracting the feedback opinions in the changed version and inputting them into the pre-constructed natural language analysis model to output semantic features; Querying the preset sentiment dictionary, identifying the sentiment value of each semantic feature, and calculating the sum of the sentiment values of each feedback opinion; Sorting all the changed versions in descending order according to the sum, generating a queue, and selecting a preset number of changed versions from the front of the queue and defining them as excellent versions.

7. The AI intelligent agent construction and training method in the working interaction scenario according to claim 1, characterized in that The method further includes: Configuring a trigger rule for optimizing the training set, where the trigger rule is: when the AI agent receives a work interaction requirement, input the optimized training set into the AI agent and perform training and fine-tuning.

8. An AI intelligent agent building and training system in a work interaction scenario, characterized in that, The system includes: A configuration module, used to identify the access positions of the AI agent, configure the work processes of each access position, generate a process chain, cluster the access positions into several categories, collect the initial data in each category, and configure the attribute information, where the attribute information at least includes: type and source; A generation module, used to create branch nodes corresponding one by one to the categories, calculate the number of access positions in each category, mount the same number of child nodes to the branch nodes, integrate the branch nodes and child nodes, and generate a training tree; A general training module, used to mark the process chain in the training tree. When the initial data uploaded by the access position is collected, based on the process chain, mark the transfer position of the initial data, define the child node corresponding to the transfer position as the target node, send an information collection pop-up window to the user terminal corresponding to the target node, obtain the changed version of the initial data, establish a correspondence between the target node and the changed version, collect the feedback of the target node, and write it into the changed version, integrate the initial data and the changed version, generate a general training set, and train the AI agent; An optimization training module, used to receive the work interaction requirements uploaded by the access position, based on the attribute data, find the initial data of the same type, select the excellent version from the corresponding changed version, use the initial data of the same type and the excellent version to generate an optimized training set, and fine-tune the AI agent.

9. The AI intelligent agent construction and training system in the work interaction scenario according to claim 8, characterized in that, The configuration module includes: A setting unit, used to set the work type of each access position and configure the transfer relationship between the access positions based on the work process; An integration unit, used to integrate the access positions and the transfer relationship to generate a process chain, where each work type corresponds to a process chain; An insertion unit, used to insert the label generated by the source into the initial data; A construction unit, used to integrate all the process chains to obtain a chain group, and construct a backtracking mechanism based on the label and the chain group.

10. The AI intelligent agent construction and training system in the work interaction scenario according to claim 8, characterized in that, The generation module includes: A reconstruction unit, used to calculate the update frequency of each child node and branch node, delete the child nodes and branch nodes with the update frequency less than the threshold, and reconstruct the training tree; A setup unit, used to open the data interface of the AI agent to the training tree and set up a communication link.

Images