An artificial intelligence switching method and device, electronic equipment and medium

Abstract

This application provides an AI-powered call transfer method, apparatus, electronic device, and medium, comprising: responding to a detected call transfer to human agent, extracting session parameters from the current AI interaction process and querying full metadata parameters associated with the currently interacting user; fusing the session parameters and full metadata parameters to obtain a complete transfer parameter package; generating a transfer work order based on the transfer parameter package using a large language model; and sending the generated transfer work order to the target human agent's port. This invention achieves seamless integration from AI interaction to human service, significantly improves the integrity of context transmission, greatly reduces the number of times users repeatedly describe their problems, and effectively improves customer satisfaction and service efficiency.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to an artificial intelligence switching method, device, electronic device, and medium. Background Technology

[0002] Currently, in intelligent customer service systems, the collaborative work between AI agents and human agents has become crucial for improving service efficiency. However, existing transfer mechanisms often suffer from incomplete context transmission. When an AI agent cannot resolve a user's problem and needs to be transferred to a human agent, the user often needs to repeat the problem, and the human agent also needs to re-understand the background, significantly reducing service efficiency and user experience. Summary of the Invention

[0003] In view of this, the purpose of the present invention is to provide an artificial intelligence switching method, device, electronic device and medium to reduce the number of repeated inquiries and improve service efficiency.

[0004] Firstly, an artificial intelligence-based call transfer method is provided, applied to a customer service application platform within a performance system. The method includes: In response to the detection of a human intervention event, extract the session parameters of the current AI interaction process and query the full metadata parameters associated with the current interactive user; The session parameters are merged with the full set of metadata parameters to obtain a complete transfer parameter package; Based on the transfer parameter package, a large language model is used to generate a transfer work order; Send the generated transfer order to the target human agent's port.

[0005] Optionally, extracting session parameters during the current AI interaction process includes: The structured parameter extraction algorithm based on a large language model is used to extract the conversation parameters in the current AI interaction process, and generate an initial parameter set and parameter summary; Calculate the weighted coverage of each parameter in the initial parameter set; the weighted coverage is the ratio obtained by dividing the sum of the products of the preset parameter importance weights and the parameter non-empty states by the total weight value of all parameters. If the weighted coverage rate is greater than or equal to the preset first threshold, the parameter extraction is deemed qualified; otherwise, a secondary extraction mechanism is triggered or the missing key parameters are filled with the system default value. The parameter summary is compared with the set of key events in the historical conversation, and the percentage of intersection between the set of events mentioned in the summary and the set of key events in the conversation history is calculated. If the overlap ratio is greater than or equal to the preset second threshold, the summary quality is deemed acceptable; otherwise, the large language model is called again for extraction by adding prompt words until the requirements are met.

[0006] Optionally, the stored procedure for the full metadata parameters associated with the currently interacting user includes: Before a user initiates an interaction, the static personal information of employees stored in the performance system is populated into the initialization object based on the pre-configured key-value pair template to form basic metadata; During the session, dynamic information from user input is extracted in real time and mapped to empty or unlocked areas of key-value pairs; Identify sensitive fields in dynamic information and encrypt them. The basic metadata and encrypted dynamic information are encapsulated into a sessionMeta object and stored for later use.

[0007] Optionally, the session parameters and full metadata parameters can be merged to obtain a complete transfer parameter package, including: Iterate through the session parameters and full metadata parameters to be merged, and identify parameter items with the same name and identifier. For the identified parameter items with the same name, calculate their confidence level in the session parameters and full metadata parameters, and retain the information of the parameter items with the same name that have high confidence. The processed, conflict-free session parameters are merged with the full metadata parameters to generate a complete transfer parameter package.

[0008] Optionally, the processed, conflict-free session parameters are merged with the full metadata parameters to generate a complete transfer parameter package, including: Iterate through all parameters to be processed, filter out parameters whose final value is not empty, and sum up the preset importance weights corresponding to each of the filtered non-empty final values ​​to obtain the first value; The second value is obtained by summing the preset importance weights of all parameters; The fusion integrity score is determined based on the relative ratio of the first value to the second value. If the fusion integrity score is greater than or equal to the preset threshold, the parameter fusion is deemed qualified and the transfer operation is allowed. If the fusion integrity score is lower than the preset threshold, the parameter fusion is deemed unqualified, the transfer operation is prevented, and the completion process is triggered.

[0009] Optionally, the method may further include, before responding to a human intervention event being detected: After each round of AI dialogue, the semantic feature vector, user sentiment index, and current interaction frequency features of the current dialogue are extracted in real time. Based on semantic feature vectors, user sentiment index, and current interaction frequency features, a pre-built multi-dimensional trigger decision algorithm model is used to determine the probability that the current session needs to be transferred to a human agent. Compare the probability that the current session needs to be transferred to a human operator with the preset transfer threshold; When the probability of the current session needing to be transferred to a human operator exceeds a preset transfer threshold, a human operator transfer event is triggered.

[0010] Optionally, before sending the generated transfer ticket to the target human agent port, the following steps are also included: Obtain the session category, user priority, and status data of all available human agent ports for the current session. The status data should include at least the skill tag and the current waiting time. The matching degree between the skill tags of each human agent port and the question category of the current session is determined; Based on the user priority, the matching degree between skill tags and the question category of the current session, and the current waiting time, the routing score of each human agent port is determined; The agent with the highest routing score is selected as the target agent port.

[0011] Secondly, an artificial intelligence transfer device is provided for use in a customer service application platform within a performance system. The device includes: The extraction unit is used to respond to the detected human intervention event, extract the session parameters in the current AI interaction process, and query the full amount of metadata parameters associated with the current interactive user; The fusion unit is used to fuse session parameters with full metadata parameters to obtain a complete transfer parameter package; The generation unit is used to generate transfer work orders based on the transfer parameter package and a large language model. The sending unit is used to send the generated transfer work order to the target human agent port.

[0012] Thirdly, an electronic device is provided, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements any of the methods of the first aspect.

[0013] Fourthly, a computer-readable storage medium is provided, wherein a computer program is stored therein, and when the computer program is executed by a processor, it implements any of the methods of the first aspect.

[0014] This invention provides an AI-powered call transfer method, device, electronic device, and medium. In response to a detected call to a human agent, the method extracts session parameters from the current AI interaction and queries the full set of metadata parameters associated with the current user. The session parameters and metadata parameters are then fused to obtain a complete transfer parameter package. Based on this parameter package, a large language model is used to generate a transfer order. The generated transfer order is then sent to the target human agent's port. This invention achieves seamless integration from AI interaction to human service, significantly improves the integrity of contextual transmission, greatly reduces the number of times users repeatedly describe their problems, and effectively improves customer satisfaction and service efficiency.

[0015] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart of an artificial intelligence switching method provided by an embodiment of the present invention is shown; Figure 2 A flowchart of another artificial intelligence switching method provided by an embodiment of the present invention is shown; Figure 3 A flowchart of another artificial intelligence switching method provided by an embodiment of the present invention is shown; Figure 4 A schematic diagram of the structure of an artificial intelligence switching device provided in an embodiment of the present invention is shown; Figure 5 A schematic diagram of the structure of an electronic device provided in an embodiment of the present invention is shown. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0019] This invention provides an artificial intelligence-based call transfer method applied to a customer service application platform within a performance system. Before the platform is officially put into operation, an initialization configuration process needs to be executed, specifically including: 1. Define and configure a dedicated startup process for the specified communication channel (e.g., WeChat Work) in the customer service application platform; 2. Embed automated processing nodes in the startup process and configure corresponding Agentic applications for each node. Simultaneously, set the autonomous operation level for each Agentic application, including: fully autonomous, where the dialogue flow is dynamically orchestrated by the Agentic application without predefined processes; and orchestrated autonomous, where the dialogue flow is collaboratively controlled by the customer service application platform and the Agentic application, with some steps predefined.

[0020] 3. Supports flexible deployment strategies for Agentic applications, allowing users to choose either a local AI agent located in the same workspace as the customer service application platform or a remote AI agent deployed in a separate environment. 4. Establish a metadata key-value pair mapping mechanism to structurally convert user profile information and real-time session information from the customer service platform, ensuring that Agentic applications can efficiently access and parse relevant metadata, thereby achieving intelligent business collaboration.

[0021] After configuring the customer service application platform, as follows: Figure 1 As shown, the method includes the following steps: Step S101: In response to the detected human intervention event, extract the session parameters in the current AI interaction process and query the full metadata parameters associated with the current interactive user.

[0022] In this step, the transfer to human intervention event refers to the signal triggered when the system determines that the current AI interaction cannot meet the user's needs or meets specific transfer conditions.

[0023] Conversation parameters refer to dynamic data generated in real time during the current AI interaction, including but not limited to dialogue turns, user intent recognition results, sentiment analysis index, and key entity extraction results during the interaction process (such as time, location, and event type).

[0024] Full metadata parameters refer to static or historical background data that is stored in the performance system before the user initiates an interaction and is bound to the identity of the current interacting user. This includes, but is not limited to, employee ID, department affiliation, job level information, historical performance scores, historical interaction record summaries, and pre-configured business attribute tags.

[0025] Understandably, session parameters reflect "what the user is saying now" and "what the AI ​​just did," exhibiting high real-time performance and context-dependent characteristics; while full metadata parameters reflect "who the user is" and "the user's historical background," demonstrating relative stability and persistence across sessions. The combination of the two provides a complementary data foundation for subsequent fusion processing.

[0026] For example, taking a customer service application platform listening for a transfer request, when a user expresses "I want to transfer to a human agent" in the conversation or the system determines that the AI ​​cannot resolve the user's problem after two consecutive responses, a transfer to a human agent event is triggered. The system then initiates a parameter extraction process. On the one hand, it extracts the text of the most recent N rounds of conversation from the current conversation cache as the basis for the conversation parameters. On the other hand, it queries the performance database for the user's profile information based on the currently logged-in user's ID (such as employeeId) as the full metadata parameters.

[0027] Step S102: Merge the session parameters with the full metadata parameters to obtain the complete transfer parameter package.

[0028] In this step, the fusion process is not a simple data splicing, but a series of operations to clean, align, and complete the session parameters and full metadata parameters.

[0029] Specifically, the fusion process includes data cleaning, which involves removing invalid characters and formatting timestamps; data alignment, which involves mapping fields with the same meaning from different sources to a unified data structure; and data completion, which involves using static information from metadata to fill in missing non-critical information in session parameters.

[0030] The final generated complete transfer parameter package is a structured data object containing verified and integrated user panoramic information.

[0031] Step S103: Generate a transfer work order based on the transfer parameter package using a large language model.

[0032] In this step, a large language model is used for semantic understanding and content generation of the complete transfer parameter package. The complete transfer parameter package is used as part of the input Prompt to guide the large language model in generating transfer orders that conform to the reading habits of human agents.

[0033] The content of a transfer ticket typically includes: a summary of the user's core needs, solutions already attempted, user emotional state indicators, key background information (such as user level, historical issues), and suggested processing directions.

[0034] Unlike traditional template-based work order generation, generating transfer work orders using a large language model can flexibly adjust the focus of the work order based on the specific dialogue content and user background, avoiding monotonous formatted content and enabling human agents to grasp the core of the problem more quickly.

[0035] For example, for emotionally agitated users, the large language model can highlight "user's emotional index is high, it is recommended to calm them down first" in the work order; for complex technical issues, the large language model can generate more detailed problem descriptions and troubleshooting steps.

[0036] Step S104: Send the generated transfer work order to the target human agent port.

[0037] In this step, the target human agent port refers to the client interface or workbench system used by human customer service personnel.

[0038] The generated transfer order is pushed to the target human agent's port in real time through communication methods such as internal message queues, API interfaces or WebSocket connections.

[0039] Once the target agent receives a transfer ticket, they can directly interact with the user based on the ticket content without asking for basic information or repeatedly understanding the background of the problem.

[0040] Through the synergistic effect of steps S100 to S400 above, this embodiment achieves a seamless connection from AI interaction to human service, significantly improves the integrity of context transmission, greatly reduces the number of times users repeatedly describe problems, and effectively improves customer satisfaction and service efficiency.

[0041] Based on the above embodiments, extracting session parameters during the current AI interaction process includes: Step S101A: Use a structured parameter extraction algorithm based on a large language model to extract the conversation parameters in the current AI interaction process, and generate an initial parameter set and parameter summary.

[0042] Specifically, the structured parameter extraction algorithm based on the large language model is not a simple text summary, but rather guides the large language model to extract structured key-value pair information from unstructured dialogue history through a pre-defined Prompt template. For example, the Prompt can require the model to output data in JSON format, containing fields such as "user intent," "question category," "key entities," and "sentiment score." The generated initial parameter set is this JSON object, while the parameter summary is a short text that summarizes the core content of the dialogue.

[0043] Step S101B: Calculate the weighted coverage of each parameter in the initial parameter set; the weighted coverage is the ratio obtained by dividing the sum of the products of the preset parameter importance weights and the parameter non-empty states by the total weight value of all parameters.

[0044] Understandably, different parameters have different levels of importance in the transfer decision. For example, the "user intent" field is crucial for agents to determine the direction of the problem, and its importance weight can be set to high (e.g., a weight value of 3); while "emotional score" is important but not necessary, its weight can be set to medium (e.g., a weight value of 2); and auxiliary information such as "number of conversation turns" can be set to low weight (e.g., a weight value of 1).

[0045] Specifically, the formula for calculating the weighted coverage ratio is as follows:

[0046] in, This indicates the total number of key parameters to be extracted; Indicates the first The importance weights of each parameter; This represents the extracted value of the j-th parameter; This indicates an indicator function; its value is 1 when the condition within the parentheses is true, and 0 otherwise. That is, when the parameter... Take 1 if it is not empty, and take 0 if it is empty.

[0047] By using this weighted calculation method, the system can more accurately evaluate the quality of parameter extraction and avoid incomplete transfer information due to missing high-weight parameters.

[0048] Step S101C: If the weighted coverage rate is greater than or equal to the preset first threshold, the parameter extraction is deemed qualified; otherwise, a secondary extraction mechanism is triggered or the missing key parameters are supplemented with the system default value.

[0049] Step S101D: Compare the parameter summary with the set of key events in the historical conversation, and calculate the percentage of intersection between the set of events mentioned in the summary and the set of key events in the history of the conversation.

[0050] In addition to ensuring the integrity of the structured parameters, this embodiment also performs semantic-level quality checks on the generated parameter summary text to prevent the omission of key events. The intersection ratio is defined as an evaluation metric, and its calculation method is as follows:

[0051] The meanings of each parameter are defined as follows: This represents a set of key events pre-extracted from the current conversation history (e.g., {"inquiring about salary", "questioning performance rating", "mentioning equipment malfunction}"). This represents the set of events mentioned in the parametric summary text generated by the large language model; Indicates the number of elements in the set; This represents the intersection operation of sets, i.e., the number of key events actually covered in the summary.

[0052] Step S101E: If the intersection ratio is greater than or equal to the preset second threshold, the summary quality is deemed qualified; otherwise, the large language model is called again for extraction by adding prompt words until the requirements are met.

[0053] Furthermore, this embodiment introduces an event coverage-based verification mechanism to ensure the quality of the parameter summary. The system pre-extracts a set of key events from the dialogue history, such as {"querying salary", "questioning performance coefficient", "mentioning equipment malfunction"}. If the event set mentioned in the summary generated by the large language model is {"querying salary", "questioning performance coefficient"}, then the intersection ratio is 2 / 3 ≈ 0.67.

[0054] If the preset second threshold is 0.8, the summary is deemed unqualified because it omits the crucial information of "mentioning device malfunction." The system will then re-invoke the large language model to extract the information by adding prompts (such as "Please supplement the summary with the specific evidence information mentioned by the user") until the summary covers a sufficient proportion of the key events.

[0055] Through the aforementioned dual verification mechanism, namely weighted coverage verification and summary event coverage verification, this embodiment can effectively ensure that the extracted session parameters are both complete and accurate, avoiding the omission of key information due to the uncertainty of the content generated by the large language model. This provides a high-quality data foundation for subsequent parameter fusion and work order generation, thereby significantly improving the readability of transfer work orders and the service efficiency of human agents.

[0056] Based on the above embodiments, this embodiment describes in detail the process of constructing and storing full metadata parameters.

[0057] The storage procedure for the full set of metadata parameters associated with the current interactive user includes: Step A: Before the user initiates an interaction, the static personal information of employees stored in the performance system is populated into the initialization object according to the pre-configured key-value pair template to form basic metadata.

[0058] Specifically, the construction of full metadata parameters is a dynamic process that spans the entire session lifecycle, rather than a one-time query.

[0059] During the pre-population phase before user interaction, the system retrieves the user's static employee information from the performance system's database based on a pre-configured key-value pair template. This static information typically includes relatively stable data such as employee ID, name, department, job level, and start date. The system then populates this data into an initialization object, forming the basic metadata.

[0060] The purpose of this step is to establish the user's identity profile before the AI ​​interaction begins, avoiding repeated requests for basic information in subsequent interactions, thereby saving interaction rounds and improving the user experience.

[0061] For example, at 9:02 AM on Monday, Li Ming, an employee of a major manufacturing client (with 3,000 employees), sent a message to the HR Cloud AI assistant via WeChat Work: "It seems like my salary this month is less than expected. Could you help me check?"

[0062] The AI ​​assistant, Little HR, reads the basic metadata and knows that Li Ming is a P3 employee in the manufacturing department and the person involved this month. There is no need to ask "What is your employee number?" and it can directly proceed to business processing.

[0063] Step B: During the session, extract dynamic information from user input in real time and map it to empty or unlocked areas of key-value pairs.

[0064] As the conversation progresses, it enters the dynamic mapping phase. Users may reveal new information during the conversation, such as specific salary objections, reasons for questioning performance reviews, or changes in contact information.

[0065] Extract this dynamic information in real time and map it to empty or unlocked areas of key-value pairs.

[0066] In this step, the unlocked area refers to fields in the pre-configured template that are allowed to be overridden or supplemented. For example, the basic metadata may have a "Contact Information" field, but the data may be empty or expired. If the user provides a new mobile phone number in the conversation, the new information will be mapped to this field to achieve real-time data updates.

[0067] In a specific example, dynamic information is extracted in the following way.

[0068] Step B1: Scan the metadata value string with the regular expression \{\{([^}]+)\}\} and extract all expression fragments; Step B2: For each expression fragment, split it into an array of attribute paths using "." as the separator. ; Step B3: From the root context object Starting with recursively accessing the attribute path:

[0069] If any intermediate layer attribute does not exist, return an empty string. And log missing path warnings; Step B4: Parse the results Replace the original expression fragment to generate the final metadata value; Step S5: Calculate the batch parsing success rate:

[0070] in, This indicates the number of times the same round of session parsing was successful; This indicates the total number of times the session is parsed within the same round.

[0071] like This triggers an alarm and notifies the configuration administrator to check the validity of the expression path.

[0072] Step C: Identify sensitive fields in dynamic information and encrypt them.

[0073] During the dynamic mapping process, a "sensitive field identification and encryption" operation is performed simultaneously. Sensitive fields in dynamic information, such as ID card numbers, bank card numbers, and specific salary amounts, are identified and encrypted.

[0074] Encryption can employ symmetric encryption algorithms such as AES-256-GCM to ensure data security during storage and transmission. Even if the database is illegally accessed, attackers cannot directly read sensitive information in plaintext. This design meets data security compliance requirements and prevents user privacy leaks.

[0075] Step D: Encapsulate the basic metadata and the encrypted dynamic information into a sessionMeta object and store it for later use.

[0076] The sessionMeta object is a structured data container that persists throughout the session lifecycle and can be accessed by AI agent applications, utility code, or system memory via predefined paths.

[0077] This encapsulation method enables unified management of multi-source data and provides a standardized data interface for subsequent parameter fusion.

[0078] For example, take employee Li Ming's salary inquiry as an example. Before Li Ming initiates the interaction through WeChat Work, the AI ​​application agent reads static information such as his employee ID "EMP-20341", name "Li Ming", department "Manufacturing Department - Production Line 3", and job level "P3" from the performance database and populates it into the initialization object to form basic metadata.

[0079] During the conversation, Li Ming typed, "My salary this month seems to be less, why is the performance coefficient 0.8? I remember I met the target." The system extracted the dynamic information of "performance coefficient objection" in real time and mapped it to the "currentIssue" field of the sessionMeta object.

[0080] Meanwhile, the system recognized that the bank card number Li Ming might mention in subsequent conversations was a sensitive field and encrypted it.

[0081] Finally, the basic metadata and encrypted dynamic information are encapsulated into a sessionMeta object and stored in the session context. When a subsequent human intervention event is triggered, this sessionMeta object can be queried directly to quickly obtain Li Ming's full metadata parameters without having to query the database again or ask Li Ming again, thus significantly improving transfer efficiency and data accuracy.

[0082] Because the information in the session parameters may overlap with or contradict the information in the full metadata parameters (for example, a user claims to belong to "Department A" in the conversation, while the system profile shows that they belong to "Department B")

[0083] Therefore, based on the above embodiments, a conflict resolution mechanism is used to fuse session parameters with full metadata parameters to obtain a complete transfer parameter package, including: Step S102A: Traverse the session parameters and full metadata parameters to be merged, and identify parameter items with the same name and the same identifier.

[0084] Specifically, during the fusion process, the system first iterates through the session parameter set and the full metadata parameter set. Since session parameters originate from the user's current real-time input and inference from the large language model, while full metadata parameters originate from pre-stored records in the system database, there may be overlap in fields. For example, both may contain fields such as "department," "job title," or "contact information." The system identifies these parameter items with the same name using field identifiers (such as the JSON key name "department").

[0085] Step S102B: For the identified parameter items with the same name, calculate their confidence level in the session parameters and full metadata parameters, and retain the information of the parameter items with the same name that have high confidence.

[0086] For parameter items with the same name, this embodiment does not simply adopt a single strategy of "new overriding old" or "old as the standard", but introduces a confidence assessment mechanism.

[0087] Confidence level is a quantitative indicator that measures the accuracy and reliability of data. For conversation parameters, the confidence level can be determined by the probability score output by the large language model during extraction, or by the degree of explicitness of the user's confirmation in the conversation; for full metadata parameters, the confidence level can be determined by the authority of the data source (e.g., the confidence level of data synchronized by the HR core system is higher than that of data filled in by the user) and the last update time.

[0088] Compare the confidence levels of parameters with the same name in two data sources, and retain the information of the parameter with the same name that has the higher confidence level.

[0089] For example, if a user explicitly corrects erroneous department information in the system archive during a conversation, and the model recognizes the intention to correct the error with high confidence, the department information in the conversation parameters is retained; otherwise, if the user only vaguely mentions the error and the model has low confidence, the original record in the system archive is retained.

[0090] This mechanism effectively resolves data conflict issues, ensures the accuracy of information in the transfer parameter packet, and prevents human agents from making misjudgments due to receiving incorrect information.

[0091] Step S102C: Merge the processed, conflict-free session parameters with the full metadata parameters to generate a complete transfer parameter package.

[0092] By integrating the data, this embodiment can effectively avoid information gaps caused by missing or incorrect information from a single data source, ensuring that the information transmitted to the human agent includes both the context of the current problem and the user's background profile, thereby significantly improving the service efficiency after the transfer.

[0093] Based on the above embodiments, the processed conflict-free session parameters are merged with the full metadata parameters to generate a complete transfer parameter package, including: Step S102C1: Iterate through all parameters to be processed, filter out parameters whose final values ​​are not empty, and sum up the preset importance weights corresponding to each of the filtered non-empty final values ​​to obtain the first value.

[0094] Understandably, simply resolving conflicts is not enough to guarantee the quality of the transfer; it is also necessary to ensure the integrity of critical information.

[0095] The integration completeness score is a core indicator used to quantitatively assess the completeness of the transfer parameter package information. Each parameter has a pre-defined importance weight; for example, core fields such as "employee ID" and "problem description" have higher weights, while auxiliary fields such as "sentiment score" have lower weights.

[0096] Step S102C2: Sieve the preset importance weights of all parameters to obtain the second value.

[0097] Step S102C3: Determine the fusion integrity score based on the relative ratio of the first value and the second value.

[0098] The parameters whose final values ​​are not empty are selected, and their weights are summed to obtain the first value. This first value is then divided by the total weight of all parameters (the second value) to obtain the fusion integrity score. This score directly reflects the coverage ratio of effective information in the current transfer parameter packet.

[0099] Specifically, the formula for the fusion integrity score is as follows:

[0100] in, For the total number of parameters, For parameter importance; This indicates the final value of the parameter.

[0101] For example, Otherwise, the transfer will be blocked and the completion process will be triggered.

[0102] Step S102C4: If the fusion integrity score is greater than or equal to the preset threshold, the parameter fusion is deemed qualified and the transfer operation is allowed.

[0103] If the fusion integrity score is greater than or equal to the preset threshold (e.g., 0.9), it indicates that the key information has been collected completely, the system determines that the parameter fusion is qualified, and allows subsequent routing and transfer operations to be executed.

[0104] Step S102C5: If the fusion integrity score is lower than the preset threshold, the parameter fusion is deemed unqualified, the transfer operation is prevented, and the completion process is triggered.

[0105] If the score is below the threshold, it indicates that key information is missing (such as missing "problem description" or "employee ID"). In this case, the system will determine that the parameter fusion is unqualified, prevent the transfer operation from being executed, and automatically trigger the completion process.

[0106] The completion process may include: sending a completion extraction command to the AI ​​agent application, requesting it to ask targeted questions for the missing fields; or reading default values ​​from the system default configuration for filling (only applicable to non-critical fields).

[0107] Through this integrity verification mechanism, this embodiment constructs a quality firewall to ensure that the work orders delivered to human agents are available and complete, avoiding situations where human agents cannot process the work orders due to missing information or need to go back and ask the user again, thus significantly improving the transfer success rate and service efficiency.

[0108] For example, continuing from the previous example, employee Li Ming mentioned in the conversation that he belongs to "Production Line Group 4", while the full metadata parameters record his department as "Manufacturing Department - Production Line Group 3".

[0109] The system detected a parameter with the same name, "Department." The confidence level for "Production Line Group 4" in the session parameters was calculated to be 0.6 (due to uncertainty in the user's tone), while the confidence level for "Production Line Group 3" in the metadata parameters was 0.95 (from synchronization with the HR system). Determined to have a higher confidence level, "Production Line Group 3" was retained as the final value.

[0110] Subsequently, the integration integrity score was calculated. It was found that the "Problem Description" field was not empty (weight 0.3), the "Employee ID" field was not empty (weight 0.3), and the "Department" field was not empty (weight 0.2), but the "Contact Information" field was empty (weight 0.2). The first value was 0.8, the second value was 1.0, and the score was 0.8. If the preset threshold is 0.85, it is considered unqualified.

[0111] The system blocks the call transfer and triggers a completion process, instructing the AI ​​assistant to ask Li Ming, "What is your contact number?" After Li Ming replies, the system updates the parameter package, raising the score to 1.0, which meets the threshold, and allows the call transfer.

[0112] Traditional referral triggers often rely on users explicitly inputting the keyword "refer to human agent" or on simple round limits, which can easily lead to trigger delays or misjudgments.

[0113] Therefore, based on the above embodiments, this embodiment focuses on describing the intelligent judgment mechanism before the transfer of manual events is triggered.

[0114] Before responding to a human intervention event, such as Figure 2 As shown, the method also includes: Step S201: After each round of AI dialogue, extract the semantic feature vector, user sentiment index and current interaction frequency features of the current dialogue in real time.

[0115] Semantic feature vectors are used to characterize the deeper meaning of the user's current input statement, such as whether it contains an intention to question, deny, or ask for help; The user sentiment index quantifies the user's psychological state by analyzing the emotional tone of the text (such as anger, anxiety, and disappointment). Historical interaction frequency characteristics reflect the density of user interactions in the current or historical sessions. High-frequency interactions often indicate that the problem has not been effectively resolved.

[0116] Step S202: Based on semantic feature vectors, user sentiment index and current interaction frequency features, use a pre-built multi-dimensional trigger decision algorithm model to determine the probability that the current session needs to be transferred to a human agent.

[0117] This embodiment introduces a multi-dimensional trigger decision algorithm model, which quantifies the necessity of switching by comprehensively evaluating features in multiple dimensions.

[0118] First, keywords such as "transfer to human assistance" are extracted from the user's semantic feature vector, along with the user's question intent label. Then, the confidence score of the large language model for the correctness of the current question's answer is determined using the user sentiment index.

[0119] Then, using these four specific features, combined with the multi-dimensional trigger decision algorithm model, a four-dimensional feature weighted fusion method is used to calculate the comprehensive turn-to-trigger score. The specific formula is as follows:

[0120] Among them, the feature functions of each dimension The specific definitions and parameter meanings are explained below:

[0121]

[0122]

[0123]

[0124] in, This represents an exponential function, with a value of 1 when the text contains explicit semantic feature vectors indicating a human intervention request (such as "transfer to human intervention" or "find customer service"), and a value of 0 otherwise. The confidence score of the large language model for the correctness of its answer to the current question ranges from 0 to 1. This indicates the threshold for judging AI capabilities. Below the threshold, the lower the confidence level, the higher the tendency to transfer the call. Indicates the user's intent in the question; This represents the set of application knowledge domains preset by the system.

[0125] For the current round of dialogue, The initial threshold, The maximum allowed number of rounds; the weighting coefficients satisfy... Default value , , , .

[0126] Step S203: Compare the probability that the current session needs to be transferred to a human operator with the preset transfer threshold.

[0127] It should be noted that the preset transfer threshold can be dynamically adjusted according to actual business scenarios. For example, for high-priority VIP users, the system can appropriately lower the threshold to make it easier to trigger a transfer, thereby ensuring a high-quality service experience; while for general inquiries, the threshold can be appropriately raised to encourage users to prioritize AI self-service solutions, optimizing the allocation of human resources. Through this flexible threshold configuration mechanism, this embodiment achieves efficient utilization of human agent resources while ensuring user experience.

[0128] Step S204: When the probability of the current session needing to be transferred to a human operator exceeds the preset transfer threshold, a human operator transfer event is triggered.

[0129] For example, consider employee Li Ming's inquiry about a salary dispute. When Li Ming enters, "This is not right. I was out of work for a few days due to equipment failure, and I have a record of that. Can you help me resolve this?", the system performs multi-dimensional analysis: the semantic feature vector identifies signals containing questioning and requests for help, such as "This is not right" and "Can you help me resolve this?", outputting a high semantic transfer tendency score; the user emotion index analysis identifies "anxiety" and "dissatisfaction," further increasing the score; simultaneously, the historical interaction frequency shows that this is the third round of interaction and the problem is still unresolved. The multi-dimensional trigger decision algorithm model, combining the above features, calculates a transfer probability of 0.92, exceeding the preset transfer threshold of 0.85, thus automatically triggering a transfer to a human agent.

[0130] At this point, Li Ming did not explicitly request to be transferred to a human operator, but the system had already proactively identified his potential need, realizing a shift from "passive response" to "proactive service," significantly improving the accuracy of transfer decisions and user satisfaction.

[0131] Based on the above embodiments, before sending the generated transfer work order to the target human agent port, such as Figure 3 As shown, it also includes: Step S301: Obtain the session category, user priority, and status data of all available human agent ports for the current session. The status data shall include at least the skill tag and the current waiting time.

[0132] Specifically, routing decision is the final and crucial step in the transfer process. Its purpose is to select the most suitable agent from among the many available human agents to handle the current user's problem, thereby achieving optimal resource allocation and maximizing user experience.

[0133] The conversation category refers to the problem classification information carried in the transfer order, such as "salary dispute", "performance appeal", "system failure", etc. This information is usually automatically identified and labeled by the aforementioned large language model when generating the transfer order.

[0134] User priority refers to a user's status level in the system, such as VIP users, Premium users, Standard users, etc. Users with different priorities have different weights in the routing queue, and high-priority users are usually assigned first.

[0135] Status data refers to the real-time work status reported by human agents. Skill tags are used to identify the business areas that agents are good at, such as "payroll expert", "attendance specialist", "general customer service" etc. The current waiting time refers to how long it is expected that the task currently being processed by the agent will become available, or the number of users currently in the queue.

[0136] Step S302: Determine the matching degree between the skill tags of each human agent port and the question category of the current session.

[0137] Understandably, simple random or first-come-first-served assignment methods often fail to meet the needs of complex business scenarios. For example, assigning a complex salary dispute to an agent skilled in handling attendance issues can lead to low processing efficiency and may even require a second transfer; or making a VIP user wait in a regular queue for a long time will severely damage the user experience.

[0138] Therefore, this embodiment introduces a multi-dimensional routing scoring mechanism to determine the optimal target seat through quantitative calculation.

[0139] In this embodiment, the matching degree is first determined based on the skill tags of each agent's port and the question category of the current session. The matching degree can be calculated using a tag matching algorithm. For example, if the session category is "salary dispute", agent A's skill tags include "salary calculation" and "tax policy", and agent B's skill tags include "attendance management", then agent A's matching degree is significantly higher than agent B's.

[0140] Matching degree can be quantified as a value between 0 and 1, with 1 for a perfect match, 0 for a complete mismatch, and a partial match calculated based on the intermediate value of tag overlap.

[0141] Step S303: Determine the routing score for each agent port based on the user priority, the matching degree between skill tags and the current session's question category, and the current waiting time.

[0142] Subsequently, the system determines the routing score for each agent's port based on the user's priority, the matching degree between skill tags and the current session's question category, and the current waiting time.

[0143] The formula for calculating route scores can be designed as a weighted summation, for example:

[0144] in, Normalized value for user priority. The normalized current waiting time for human agents. The degree of match between the skill tag and the question category in the current session; , , These are preset weighting coefficients, and the sum of the three is 1.

[0145] User priority normalization values ​​map different user levels to numerical values. For example, VIP users are mapped to 1.0, Premium users to 0.7, and Standard users to 0.3.

[0146] The normalized waiting time value maps the waiting time to a value between 0 and 1. The longer the waiting time, the larger the normalized value. (1 - normalized waiting time value) represents the level of idle time. The higher the level of idle time, the higher the score for this item.

[0147] Through this weighted calculation, the system can comprehensively balance user importance, agent professionalism, and agent availability, and select the agent with the highest total score as the target human agent port.

[0148] It should be noted that the above weighting coefficients can be dynamically adjusted according to actual business needs. For example, during peak business periods, in order to quickly alleviate queuing pressure, the weight of waiting time w3 can be appropriately increased and prioritized for allocation to available agents; while during VIP service guarantee periods, the weight of user priority w1 can be significantly increased to ensure that VIP users have priority access.

[0149] Step S304: Determine the agent with the highest routing score as the target agent port.

[0150] For example, continuing from the previous example, after employee Li Ming's salary objection work order is generated, the system obtains Li Ming's user priority as Standard (normalized value 0.3) and the session category as "salary objection > performance deduction appeal".

[0151] There are currently three available human agent queues in the system: Queue A is "Payroll Specialist Group - Wang Fang", whose skill tag includes "Payroll Appeal", with a match rate of 0.95 and a relatively short current waiting time (normalized idle rate 0.9); Queue B is "General HR Group - Zhang Lei", whose skill tag includes "General Consulting", with a match rate of 0.60 and a medium current waiting time (normalized idle rate 0.7); Queue C is "Attendance Specialist Group - Liu Yan", whose skill tag includes "Attendance Management", with a match rate of 0.20 and a relatively long current waiting time (normalized idle rate 0.5).

[0152] The comparison shows that queue A has the highest score, and the system will route Li Ming's transfer work order to "Payroll Special Group - Wang Fang".

[0153] This result aligns with business logic. As a compensation expert, Wang Fang is best positioned to handle salary disputes professionally, and she is currently available and able to respond quickly.

[0154] Through this flexible routing scoring mechanism, this embodiment achieves precise allocation of transfers, which not only improves problem-solving efficiency but also optimizes the utilization rate of human agent resources, significantly improving user satisfaction.

[0155] To more intuitively verify the practical effect of the technical solution of the present invention, this embodiment takes the handling of employee salary objections in an HRSaaS scenario as an example to illustrate the technical features of the above embodiments in a series. In this scenario, the customer service application platform in the performance system is equipped with an AI assistant to handle employees' daily inquiries and appeals.

[0156] Specifically, the process for this application instance is as follows: The first step is metadata injection and initialization.

[0157] Just before employee Li Ming initiates an interaction via WeChat Work, a pre-population operation for all metadata parameters is performed. Based on the pre-configured key-value pair template, static personal information such as Li Ming's employee ID "EMP-20341", name "Li Ming", department "Manufacturing Department - Production Line 3", and job level "P3" are read from the employee profile database of the performance system and populated into the initialization object to form basic metadata.

[0158] Simultaneously, the system identifies salary data as a sensitive field and automatically encrypts the relevant fields. Finally, the basic metadata and the encrypted dynamic information are encapsulated into a sessionMeta object and stored.

[0159] This process enables the implicit preparation of user background information, allowing the AI ​​assistant to directly access the core business without asking for basic information in subsequent interactions, significantly improving interaction efficiency.

[0160] Part Two: AI Interaction and Multi-Dimensional Triggered Judgments.

[0161] Li Ming sent a message: "My salary this month seems to be less than expected, can you help me check?" The AI ​​assistant read the sessionMeta object, knew Li Ming's identity, and directly replied with salary details: "Hello Li Ming! I'm here to check your January salary. Your payslip has been issued, your basic salary is 3500 yuan, and your performance coefficient for this month is 0.8. Would you like me to explain the deductions in detail?"

[0162] When Li Ming saw the performance coefficient was 0.8 and questioned, "This is not right. I stopped working those days because of equipment failure, and I have a record of it. Can you help me solve this?", the semantic feature vector of the current dialogue, the user sentiment index, and the historical interaction frequency features were extracted.

[0163] Specifically, the semantic feature vector identifies questioning and request signals such as "This is not right" and "Can you help me solve this?"; the user sentiment index analyzes the emotions of "anxiety" and "dissatisfaction", and the scores increase; the historical interaction frequency shows that this is the 3rd round of interaction and the problem is still unresolved.

[0164] Based on the above characteristics, a pre-built multi-dimensional trigger decision algorithm model is used to determine the probability that the current session needs to be transferred to a human agent. For example, if the calculated probability value is 0.92, which exceeds the preset transfer threshold of 0.85, the human agent transfer event will be automatically triggered.

[0165] This mechanism enables the AI ​​to proactively identify its capabilities even when the user does not explicitly request a human operator, thus allowing for precise control over the timing of the transfer.

[0166] The third step is parameter extraction and dual quality verification.

[0167] In response to a detected human intervention event, the system extracts conversation parameters from the current AI interaction process. Utilizing a structured parameter extraction algorithm based on a large language model, it extracts structured parameters from the three-turn dialogue history, generating an initial parameter set and a parameter summary.

[0168] The initial parameter set includes fields such as "User Intent: Performance Coefficient Appeal" and "Issue Category: Salary Dispute". The weighted coverage of each parameter in the initial parameter set is calculated. Assuming that the "User Intent" field is not empty and has the highest weight, the calculation result is 0.95, which is greater than the preset first threshold of 0.8, and the parameter extraction is deemed qualified.

[0169] Meanwhile, the parameter summary was compared with the set of key events in the historical conversation. The summary mentioned events such as "inquiring about salary", "questioning performance coefficient" and "mentioning equipment failure". The intersection ratio with the set of key events in the conversation history was 1.0, which is greater than the preset second threshold of 0.8, so the summary quality was judged to be qualified.

[0170] The dual verification mechanism ensures that the extracted session parameters are both complete and accurate, avoiding the omission of key information.

[0171] The fourth step is parameter fusion and integrity verification.

[0172] Query the full set of metadata parameters associated with the currently interacting user, i.e., the previously stored sessionMeta object.

[0173] The session parameters and full metadata parameters are merged. During the merging process, the system iterates through the parameters to be merged and identifies the "department" field as a parameter with the same name. Assuming that Li Ming verbally states "Production Line Group 4" in the session parameters with a confidence level of 0.6, and the full metadata parameters record it as "Production Line Group 3" with a confidence level of 0.95, the parameter with the higher confidence level, "Production Line Group 3", is retained.

[0174] Subsequently, the fusion integrity score is calculated. Assuming that all key parameters are not empty, the score is 1.0, which is greater than the preset threshold of 0.9. The parameter fusion is deemed qualified, and the transfer operation is allowed.

[0175] This step effectively resolves multi-source data conflicts and ensures the accuracy of the transfer parameter package.

[0176] Step 5: Intelligent routing and work order sending.

[0177] Based on the transfer parameter package, a transfer work order is generated using a large language model. The work order content covers Li Ming's core demands, the solutions already tried, and his emotional state.

[0178] Before sending, obtain the current session's session category "salary objection", user priority "Standard", and status data for all available agent ports.

[0179] The matching degree is determined based on the skill tags of each human agent's portal and the question category of the current session. For example, Wang Fang from the payroll team has the skill tag "payroll appeal" and the matching degree is 0.95; Zhang Lei from the general HR team has a matching degree of 0.60.

[0180] Based on user priority, matching degree, and current waiting time, the routing score for each human agent port is calculated. Assuming Wang Fang is currently idle and has a short waiting time, her routing score is the highest. The system identifies Wang Fang, with the highest routing score, as the target human agent port and sends the generated transfer work order to Wang Fang's workstation. Upon receiving the work order, Wang Fang, without needing Li Ming to repeat the problem description, directly verifies and processes the issue based on the complete context.

[0181] Through the above process, this embodiment fully demonstrates the entire process from metadata injection, AI interaction, multi-dimensional trigger judgment, parameter extraction and fusion to final route transfer. Compared with traditional transfer methods, the technical solution of this embodiment achieves near 100% context transmission integrity, improves transfer judgment accuracy to over 89%, and reduces the number of times users repeatedly describe problems by 78%, significantly improving service efficiency and user experience. This fully verifies the effectiveness and superiority of the present invention in solving technical problems such as context loss and inaccurate transfer judgment.

[0182] Based on the same inventive concept, an artificial intelligence transfer device is provided for application in a customer service platform within a performance system, such as... Figure 4 As shown, the device includes: Extraction unit 401 is used to respond to the detected transfer to human intervention event, extract the session parameters in the current AI interaction process and query the full amount of metadata parameters associated with the current interactive user; The fusion unit 402 is used to fuse the session parameters with the full metadata parameters to obtain a complete transfer parameter package; The generation unit 403 is used to generate a transfer work order based on the transfer parameter package and using a large language model. The sending unit 404 is used to send the generated transfer work order to the target human agent port.

[0183] Based on the same technical concept, embodiments of the present invention also provide an electronic device, such as... Figure 5 As shown, it includes a processor 501, a communication interface 502, a memory 503, and a communication bus 504, wherein the processor 501, the communication interface 502, and the memory 503 communicate with each other through the communication bus 504.

[0184] Memory 503 is used to store computer programs; The processor 501 is used to implement the steps of the artificial intelligence switching method when executing the program stored in the memory 503.

[0185] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0186] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0187] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0188] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0189] The computer program product for artificial intelligence switching provided in this embodiment of the invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation details, please refer to the method embodiments, which will not be repeated here.

[0190] The artificial intelligence switching device provided in this embodiment of the invention can be specific hardware on the device or software or firmware installed on the device. The implementation principle and technical effects of the device provided in this embodiment of the invention are the same as those in the foregoing method embodiments. For the sake of brevity, any parts not mentioned in the device embodiments can be referred to the corresponding content in the foregoing method embodiments. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can all be referred to the corresponding processes in the above method embodiments, and will not be repeated here.

[0191] In the embodiments provided by this invention, it should be understood that the disclosed apparatus and method can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0192] 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 units can be selected to achieve the purpose of this embodiment according to actual needs.

[0193] In addition, the functional units in the embodiments provided by the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0194] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0195] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In addition, the terms "first", "second", "third", etc. are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0196] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. All should be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An artificial intelligence switching method, characterized in that, The method, applied to a customer service application platform within a performance system, includes: In response to the detection of a human intervention event, extract the session parameters of the current AI interaction process and query the full metadata parameters associated with the current interactive user; The session parameters are fused with the full metadata parameters to obtain a complete transfer parameter package; Based on the aforementioned transfer parameter package, a transfer work order is generated using a large language model. The generated transfer order is sent to the target human agent port.

2. The method according to claim 1, characterized in that, Extracting session parameters during the current AI interaction process includes: The structured parameter extraction algorithm based on a large language model is used to extract the conversation parameters in the current AI interaction process, and generate an initial parameter set and parameter summary; Calculate the weighted coverage ratio of each parameter in the initial parameter set; the weighted coverage ratio is the ratio obtained by dividing the sum of the products of the preset parameter importance weights and the parameter non-empty states by the total weight value of all parameters. If the weighted coverage rate is greater than or equal to the preset first threshold, the parameter extraction is deemed qualified; otherwise, a secondary extraction mechanism is triggered or the missing key parameters are filled with the system default value. The parameter summary is compared with the set of key events in the historical conversation, and the percentage of intersection between the set of events mentioned in the summary and the set of key events in the conversation history is calculated. If the intersection ratio is greater than or equal to the preset second threshold, the summary quality is deemed acceptable; otherwise, the large language model is called again for extraction by adding prompt words until the requirements are met.

3. The method according to claim 1, characterized in that, The storage procedure for the full set of metadata parameters associated with the current interactive user includes: Before a user initiates an interaction, the static personal information of employees stored in the performance system is populated into the initialization object based on the pre-configured key-value pair template to form basic metadata; During the session, dynamic information input by the user is extracted in real time and mapped to empty or unlocked areas of the key-value pairs; Identify the sensitive fields of the dynamic information and encrypt the sensitive fields; The basic metadata and the encrypted dynamic information are encapsulated into a sessionMeta object and stored for later use.

4. The method according to claim 1, characterized in that, The process of fusing the session parameters with the full metadata parameters to obtain the complete transfer parameter package includes: Traverse the session parameters and full metadata parameters to be merged, and identify parameter items with the same name and identifier. For the identified parameter items with the same name, calculate their confidence level in the session parameters and full metadata parameters, and retain the values ​​of the parameter items with the same name that have high confidence. The processed, conflict-free session parameters are merged with the full metadata parameters to generate a complete transfer parameter package.

5. The method according to claim 4, characterized in that, The process of merging the processed, conflict-free session parameters with the full metadata parameters to generate a complete transfer parameter package includes: Iterate through all parameters to be processed, filter out parameters whose final value is not empty, and sum up the preset importance weights corresponding to each of the filtered non-empty final values ​​to obtain the first value; The second value is obtained by summing the preset importance weights of all parameters; The fusion integrity score is determined based on the relative ratio between the first value and the second value. If the fusion integrity score is greater than or equal to the preset threshold, the parameter fusion is deemed qualified and the transfer operation is allowed. If the fusion integrity score is lower than the preset threshold, the parameter fusion is deemed unqualified, the transfer operation is prevented, and the completion process is triggered.

6. The method according to claim 1, characterized in that, Prior to responding to a detected human intervention event, the method further includes: After each round of AI dialogue, the semantic feature vector, user sentiment index, and current interaction frequency features of the current dialogue are extracted in real time. Based on the semantic feature vector, user sentiment index, and current interaction frequency features, a pre-built multi-dimensional trigger decision algorithm model is used to determine the probability that the current session needs to be transferred to a human operator. The probability that the current session needs to be transferred to a human operator is compared with a preset transfer threshold; When the probability that the current session needs to be transferred to a human operator exceeds the preset transfer threshold, a human operator transfer event is triggered.

7. The method according to claim 1, characterized in that, Before sending the generated transfer order to the target human agent port, the process also includes: Obtain the session category, user priority, and status data of all available human agent ports for the current session. The status data includes at least the skill tag and the current waiting time. The matching degree between the skill tags of each human agent port and the question category of the current session is determined; Based on the user priority, the matching degree between skill tags and the question category of the current session, and the current waiting time, the routing score of each human agent port is determined; The agent with the highest routing score is selected as the target agent port.

8. An artificial intelligence switching device, characterized in that, A customer service application platform used in a performance system, the device comprising: The extraction unit is used to respond to the detected human intervention event, extract the session parameters in the current AI interaction process, and query the full metadata parameters associated with the current interactive user. The fusion unit is used to fuse the session parameters with the full metadata parameters to obtain a complete transfer parameter package; The generation unit is used to generate a transfer work order based on the transfer parameter package using a large language model; The sending unit is used to send the generated transfer work order to the target human agent port.

9. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; The memory is used to store computer programs; When the processor executes the program stored in the memory, it implements the method described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method described in any one of claims 1-7.

Images