Customized service information generation method and system, storage medium, computer device

Abstract

The application relates to the technical field of information dynamic generation, can be applied to the financial technical field, and discloses a customized service information generation method and system, a storage medium and computer equipment.The method comprises the following steps: an intelligent agent calls each skill unit in sequence according to a customized service information generation task, receives aggregated level data of feedback, determines different user behavior characteristics based on the aggregated level data, generates customized initial service information according to the different user behavior characteristics, and when the customized initial service information passes the compliance verification, the customized initial service information is determined as customized service information and stored in a temporary storage area; and the intelligent agent reads the customized service information matched with the user behavior characteristics corresponding to a user behavior event from the temporary storage area and sends the customized service information to a preset terminal of the user in response to the user behavior event. The full-process automation from intelligent analysis to personalized information content generation can be realized without exposing original sensitive data, and the technical contradiction between data security and the efficiency of the intelligent agent is solved.

Description

Technical Field

[0001] This application relates to the field of information dynamic generation technology, and in particular to a customized service information generation method and system, storage medium, and computer equipment. Background Technology

[0002] With the rapid advancement of digital financial services, financial institutions' service information delivery to users has shifted from traditional homogeneous notifications to a personalized push model tailored to each individual. How to generate highly relevant service information while adhering to compliance has become a core technological requirement in the current financial services sector.

[0003] The current mainstream customized service information generation solutions in the industry mostly adopt the mode of directly calling the original user individual data for real-time calculation. That is, after the push task is triggered, the system directly pulls the user's transaction details, behavior trajectory, personal attributes and other full individual data, and performs feature analysis and content generation on the server side. After completion, it is directly sent to the user, which can easily cause the risk of individual data leakage and is difficult to meet the security requirements of the financial industry's strong supervision. Summary of the Invention

[0004] In view of this, this application provides a customized service information generation method and system, storage medium, and computer equipment, which can automate the entire process from intelligent analysis to personalized information content generation without exposing the original sensitive data, thus resolving the technical contradiction between data security and intelligent agent performance.

[0005] According to one aspect of this application, a method for generating customized service information is provided, the method comprising: The intelligent agent receives a customized service information generation task, determines multiple skill units required to complete the customized service information generation task, and generates a directed acyclic graph representing the execution order and dependencies of the skill units. The skill units are preset modular functional components. The customized service information generation task is used to generate customized service information separately for different user behavior characteristics. The agent sequentially calls each skill unit based on the directed acyclic graph, so that each skill unit processes the execution result into aggregated data and feeds it back to the agent. The aggregated data is the statistical result of the user group dimension, which includes the aggregated statistical result of historical behavior events, the aggregated value of user behavior characteristics, and the group size. The intelligent agent determines different user behavior characteristics based on aggregated data fed back by multi-skill units, and generates customized initial service information adapted to each user behavior characteristic. When any customized initial service information passes the compliance verification, it is identified as customized service information and stored in the temporary storage area. Different user behavior characteristics correspond to different customized service information. In response to a user behavior event, the intelligent agent determines the user behavior characteristics corresponding to the user behavior event, reads the customized service information corresponding to the user behavior characteristics from the temporary storage area, and sends the customized service information to the user's preset terminal.

[0006] According to another aspect of this application, a customized service information generation system is provided, the system comprising: The intelligent agent decision-making center module is used to receive customized service information generation tasks, determine multiple skill units required to complete the customized service information generation tasks, and generate a directed acyclic graph representing the execution order and dependencies of the skill units. The skill units are preset modular functional components. The customized service information generation tasks are used to generate customized service information separately for different user behavior characteristics. The data security isolation module is used to sequentially call each skill unit based on a directed acyclic graph, so that each skill unit processes the execution results into aggregated data and then feeds them back. The aggregated data is the statistical results of the user group dimension, including the aggregated statistical results of historical behavioral events, the aggregated values ​​of user behavior characteristics, and the group size. The personalized content generation module is used to determine different user behavior characteristics based on aggregated data from multi-skill units and generate customized initial service information adapted to each user behavior characteristic. The information is verified by the compliance module. When any customized initial service information passes the compliance verification, it is identified as customized service information and stored in the temporary storage area. Different user behavior characteristics correspond to different customized service information. The real-time triggering and caching module is used to respond to user behavior events, determine the user behavior characteristics corresponding to the user behavior events, read the customized service information corresponding to the user behavior characteristics from the temporary storage area, and send the customized service information to the user's preset terminal.

[0007] According to another aspect of this application, a storage medium is provided that stores a computer program thereon, which, when executed by a processor, implements the above-described customized service information generation method.

[0008] According to another aspect of this application, a computer device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the program to implement the above-described customized service information generation method.

[0009] By employing the above technical solutions, the customized service information generation method and system, storage medium, and computer equipment provided in this application can achieve full-process automation from intelligent analysis to personalized information content generation without exposing original sensitive data, thus resolving the technical contradiction between data security and intelligent agent performance.

[0010] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0011] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A schematic diagram of the application environment of an embodiment of this application is shown; Figure 2 This illustration shows a flowchart of a customized service information generation method according to an embodiment of this application. Figure 3 This illustration shows a schematic diagram of a directed acyclic graph provided in an embodiment of this application; Figure 4 This illustration shows a schematic diagram of the structure of a customized service information generation system provided in an embodiment of this application; Figure 5 A schematic diagram of a computer device structure according to an embodiment of this application is shown; Figure 6 This illustration shows another schematic diagram of a computer device structure according to an embodiment of this application. Detailed Implementation

[0012] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0013] The method provided in this application embodiment can be applied to, for example, Figure 1 In the application environment, it consists of three levels of collaboration: terminal, network, and cloud. More specifically, it can be applied to customized financial service push systems. The aforementioned customized financial service push system is used to form a closed loop of service push, from collecting user behavior, generating customized services in batches, to precise delivery. Figure 1In this architecture, the client is the user's interaction entry point, responsible for collecting various user behavior events such as clicks, browsing, and transactions. It is also the final terminal for receiving customized service information. The network, as the end-to-end data transmission hub, carries bidirectional communication between the client and server. It is responsible for uploading user behavior events and interaction requests to the server and pushing customized service information generated by the server back to the client. The server is a cluster node deployed in the cloud, serving as the core computing and processing carrier, and carrying core business logic such as task scheduling, data aggregation, information generation, compliance verification, and temporary caching.

[0014] In this embodiment, the executing entity is an intelligent agent, which can be directly embedded and run within... Figure 1 Within the server-side nodes, the core computing logic module relies on the server's computing power and storage resources to complete all complex processing tasks. The agent and client interact indirectly via the network, receiving real-time user behavior events reported by the client and sending matched customized service information to the client via the network. The agent and network layer have a pure channel dependency relationship; the network is the communication link between the agent and client. The agent relies on the network to receive uplink events and deliver downlink service information, without directly controlling the network nodes. The agent and server's local resources are deeply integrated. The agent can directly call various pre-set modular skill units in the server to complete aggregate-level data calculations, and can also schedule the server's temporary storage area to store pre-generated customized service information that has passed compliance verification, ensuring millisecond-level fast reading and matching when user behavior is triggered subsequently.

[0015] This embodiment provides a method for generating customized service information, such as... Figure 2 As shown, the method includes: Step 101: The intelligent agent receives the customized service information generation task, determines the multiple skill units required to complete the customized service information generation task, and generates a directed acyclic graph representing the execution order and dependencies of the skill units. The skill units are preset modular functional components. The customized service information generation task is used to generate customized service information separately for different user behavior characteristics.

[0016] Currently, in the financial sector, there are three main technical solutions for generating customized service information: The first category is financial service systems based on traditional rule engines. These systems match service coverage targets based on fixed business rules and use standardized fixed notification templates to send notifications in batches. Representative products include standardized notification SMS mass sending systems. Due to the long rule iteration chain and lack of flexibility, these systems cannot adapt to rapidly changing business requirements. At the same time, the notification content is highly homogenized, resulting in a low user service experience and low response rate.

[0017] The second category is AI (Artificial Intelligence)-assisted customized service information generation tools, which rely on large language models (LLM) to generate customized service notification content. However, they can only focus on a single information generation stage. Representative products include various general AI content generation tools and intelligent copywriting generation SaaS (Software as a Service) services. Because they are completely separated from the front-end user feature analysis stage, they cannot dynamically generate differentiated and adapted content based on the characteristics of different service objects.

[0018] The third type is the raw data direct input large model analysis solution, which directly inputs all raw user details into the large model to complete the analysis and processing. However, it is limited by the context token capacity of the large model. For example, a typical 128K context is equivalent to about 100,000 words of text, which makes it impossible to efficiently process the massive details of millions of users. At the same time, it poses extremely high data security and compliance risks, as core sensitive information such as user mobile phone numbers and account transaction records are directly exposed to the large model, which does not meet the data security control requirements of the financial industry.

[0019] In summary, in the current scenario of customized service information generation, the lack of a unified intelligent agent to coordinate the entire process leads to a disconnect between data analysis and content generation. The volume of raw detailed data far exceeds the model's processing capacity, resulting in a contradiction between the large model's token limitations and the massive amount of data. The technical conflict between data availability and data confidentiality makes it impossible to achieve both data security and AI efficiency. The fact that the analysis logic and copywriting strategies are scattered among personal experience prevents the accumulation and reuse of expert experience. Furthermore, the manual connection of multiple links results in a cycle of several days, leading to poor service response timeliness.

[0020] In the above embodiments of this application, the entire process from intelligent user analysis to personalized SMS content (customized service information) generation can be automated without exposing original sensitive data, thus resolving the technical contradiction between data security and AI performance. Specifically, a three-layer isolation, intelligent orchestration, and secure closed-loop overall concept can be used, including the following: The first layer is the AI ​​Agent (intelligent agent) decision-making center, which acts as the intelligent brain. It is used to understand business needs, autonomously plan tasks, apply reasoning to make decisions, dynamically adjust strategies, and finally coordinate the execution of various Skill modules (skill units). At the same time, it does not touch the raw data throughout the process, but only processes the aggregated results and generated content.

[0021] The second layer is the Skill encapsulation layer, which is used to abstract and encapsulate capabilities for standardized invocation. In other words, capabilities such as data analysis, insight generation, and content creation are encapsulated into standardized interfaces. Each Skill performs specific data operations internally and only exposes the aggregated results to the outside world. Skills can be reused, combined, iteratively upgraded, and form a security barrier where data is available but not visible.

[0022] The third layer is the data platform, which is used to physically isolate the raw data and perform internal execution. Specifically, it is used to store raw user details (phone numbers, transaction records, behavior logs, etc.) and only provides aggregated query services to the outside world through standardized APIs. At the same time, the AI ​​Agent and the raw data are in different network domains and are physically inaccessible.

[0023] In one specific implementation, for example, a bank's operations staff submits a customized service information generation task in the background. The customized service information generation task analyzes user behavior characteristics through user behavior events triggered by the user, and then generates customized service information based on the user behavior characteristics and sends it to the user's preset terminal. For example, it requires pushing a first-swipe benefit reminder adapted to the consumption behavior of new users who have just activated their credit cards within the past 7 days, so as to achieve customized push. Specifically, when the agent receives the task of generating customized service information, it matches multiple required skill units and generates a corresponding directed acyclic graph (DAG). For example, it identifies three skill units, their execution order, and dependencies. The resulting DAG covers the following: First, the new user selection skill unit is executed, outputting the aggregated results of user groups that meet the activation conditions of the past 7 days. This result is then used as input to the consumption behavior aggregation and statistics skill unit. After the consumption behavior statistics are completed and different user behavior characteristics are output, the rights and benefits adaptation rule matching skill unit is activated. Finally, customized rights and benefits content (customized service information) corresponding to different user behavior characteristics is generated. In the entire chain, each subsequent skill unit must wait for the aggregated data of the previous one to be returned before it can be activated, completely eliminating the problem of missing parameters due to premature calls.

[0024] Specifically, after setting up a customized service information generation task, trigger conditions and a validity period can be set. The intelligent agent will start the customized service information generation task when it determines that the task meets the trigger conditions and is within the validity period. For example, the trigger conditions might require it to occur during a non-peak business period at 2 AM daily, and the validity period could be 30 days from the date of submission. After receiving the task, the intelligent agent first verifies that the current time is exactly the trigger time set for the task and is within the 30-day validity period; if so, it determines that the task can be started and executed.

[0025] Specifically, skill units can also correspond to input parameters, and the input parameters between skill units have dependencies. The steps for the agent to generate a directed acyclic graph representing the execution order and dependencies of skill units are as follows: The agent determines the dependency links between skill units based on the input parameter dependencies of each skill unit, marks skill units without dependencies as skill units that can be executed in parallel, and marks skill units with sequential dependencies as skill units that can be executed sequentially. The agent integrates all skill units, their execution order, and dependencies into a loop-free topology, generating a directed acyclic graph that represents the execution order and dependencies of the skill units.

[0026] Specifically, dependencies can include prerequisite dependencies. Skill units with prerequisite dependencies must wait for the prerequisite skill unit to complete execution and return aggregate-level data before starting execution. Furthermore, the agent first reads the input parameter requirements of all selected skill units one by one, checking whether the input of each skill unit depends on the output results of other skill units. For example, the input of the new user aggregation and selection skill unit is only the time range parameter provided by the task, which is the starting node of the entire chain; the input of the behavior aggregation and statistics skill unit must depend on the new user group aggregation result output by the new user selection skill unit, which is a serial dependency node; the input of the points aggregation and statistics skill unit also only needs the new user group aggregation result, and there is no parameter dependency between it and the behavior statistics skill unit, so it will be marked as a skill unit that can be executed in parallel. The rights and interests rules adapt to the input of the skill unit, and can only be activated if the output results of both the usage behavior statistics and the points statistics skill units are obtained simultaneously. This is a subsequent sequential node. After the agent sorts out all parameter dependencies, it forms a complete skill unit dependency chain, clearly distinguishing between parallel and sequential execution units.

[0027] Next, the agent treats all skill units as independent nodes in the topology graph, transforming the identified parameter dependencies into directed connections between nodes. It automatically verifies that the entire topology structure does not contain any closed loops, ultimately generating a complete directed acyclic graph (DAG). The graph clearly marks the rules that parallel nodes can start simultaneously, while sequential nodes must wait for the preceding node to return aggregated data before running. During subsequent agent scheduling, all skill units can be called in an orderly manner according to this topology structure, without issues such as missing parameters or process bottlenecks. An example of a directed acyclic graph is... Figure 3 As shown.

[0028] Step 102: The agent sequentially calls each skill unit based on the directed acyclic graph, so that each skill unit processes the execution result into aggregated data and feeds it back to the agent. The aggregated data is the statistical result of the user group dimension, which includes the aggregated statistical result of historical behavior events, the aggregated value of user behavior features, and the group size.

[0029] Next, a skill-based data isolation mechanism is implemented. Currently, Large Language Models (LLMs) face dual constraints when processing massive amounts of user data: Capacity constraint: The model context window is limited (usually 128K tokens), which cannot accommodate millions of user details; Security constraints: Sensitive industries such as finance require raw data to remain within their domains, and directly inputting it into the model poses a risk of data leakage.

[0030] The existing technical solutions are as follows: Option A: Inputting data in batches by slicing it up disrupts the integrity of the data and distorts the analysis results.

[0031] Option B: Inputting data after anonymization may result in the model still remembering the anonymized data, posing an uncontrollable risk.

[0032] Option C: Abandoning AI (agent) analysis and continuing to use traditional rule engines will result in the loss of AI's intelligent advantages.

[0033] In the embodiments described above in this application, a method is constructed in which AI directs the analysis, rather than performing the analysis itself. Specifically, instead of inputting raw data into the AI, analytical logic is input into the AI, allowing the AI ​​to understand the analytical requirements and then call the encapsulated analytical interface (Skill) to perform specific calculations. The Skill performs data operations in a controlled environment and only returns the aggregated results to the AI. The AI ​​then performs higher-level reasoning and decision-making based on the aggregated results.

[0034] Specifically, based on previous scenarios of bank credit card new customer benefit outreach, aggregated data, such as the full aggregated statistics generated by banks for users who newly activated their credit cards in the past 7 days, does not involve any detailed information that can pinpoint individuals, fully complying with data security requirements. Historical behavioral event aggregation statistics, for example, show that among the newly activated users in the past 7 days, a total of 85,700 users triggered in-app benefit browsing behavior, generating a total of 192,000 benefit page access records. User behavior characteristic aggregation values, such as the average monthly points redemption frequency of the benefit browsing group being 2.1 times, the proportion of users preferring physical benefits reaching 72%, and the average time spent on a single benefit page being 48 seconds. Group size, for example, the total number of newly activated credit card users covered in this statistic in the past 7 days is 126,000, of which the total size of the benefit browsing subgroup is 85,700.

[0035] Optionally, in step 102, the agent sequentially calls each skill unit based on the directed acyclic graph, so that each skill unit processes the execution result into aggregated data and then feeds it back to the agent, including: Step 1021: The agent determines multiple skill units to be scheduled, as well as the execution order and dependencies between the skill units, based on the directed acyclic graph. Step 1022: The agent schedules skill units based on the execution order and dependency relationship. When any scheduled skill unit completes its execution and obtains initial data, the skill unit checks whether the initial data conforms to the aggregation-level data rules and feeds back the aggregation-level data that conforms to the aggregation-level data rules to the agent.

[0036] In the above embodiments of this application, the agent can determine multiple skill units to be scheduled, as well as the execution order and dependencies between skill units, based on a directed acyclic graph. When a skill unit completes execution and obtains initial data, the skill unit checks whether the initial data conforms to the aggregation-level data rules to ensure that the data sent to the agent is aggregation-level data. Specifically, the skill unit can be represented by a Skill abstract model, which is as follows: Skill={Metadata,InputSchema,OutputSchema,ExecutionPolicy,SecurityPolicy} Among them: Skill, Metadata, including Skill identifier, version, author, and function description; InputSchema, Input parameter specification (JSON Schema), used to define the parameters that the AI ​​(agent) can pass in; OutputSchema, Output result specification (JSON Schema), used to define the data format returned to the AI; ExecutionPolicy, Execution policy, such as timeout, retry, concurrency, and degradation; SecurityPolicy, Security policy, such as data access scope, sensitive field control, and audit level.

[0037] The key design of the Skill abstract model is reflected in the mandatory aggregation of the output schema (data schema / data blueprint). This clearly defines the red lines for data output, allowing only aggregated statistical results at the user group level to be returned. It completely prohibits the outflow of any detailed data containing unique user identifiers, phone numbers, or other details that can pinpoint specific individuals, thus fundamentally mitigating the risk of sensitive user information leakage in financial scenarios. Within the compliant aggregated data scope, core information supporting business decisions is retained. This covers basic statistical items such as total scale and number of people in each stratum, while also supplementing high-value analytical content such as average group attributes and business insights, fully meeting the needs for generating differentiated and customized service information. Through preset format rules, it automatically intercepts individual-level data that does not meet requirements, completing data compliance verification without additional manual review, thus adapting to the high-efficiency requirements of processing large volumes of user data in financial scenarios.

[0038] In particular, regarding skill units, skill accumulation can be achieved through expert experience. Currently, financial institutions rely heavily on a few senior experts when generating service information. Senior analysts possess unique user screening logic, such as cross-analysis of points, activity, and preferences. Senior copywriters master the skills of creating high-conversion copy, such as a sense of urgency, specific numbers, and clear actions. These experiences exist in individuals' minds but are lost with staff turnover, resulting in a long training cycle for new employees.

[0039] The core concept of the above embodiments of this application is to encapsulate expert experience into reusable software modules, that is, to "skillify" it. Specific methods include: Analyzing experience into skills: Encapsulate the parameter tuning techniques and cross-analysis logic of the RFM (Recency, Frequency, Monetary) model into data modeling skills. RFM is a classic analytical tool for measuring user value and profitability. It segments users through three core indicators: recent consumption time (R), consumption frequency (F), and consumption amount (M), supporting refined operations.

[0040] Skill-based copywriting: Encapsulate service information matching strategies and tone adaptation rules into copywriting generation skills; Compliance experience skillification: Encapsulate the disabled term library and review rules into compliance review skills.

[0041] Furthermore, this can be achieved through the following technologies: Skill registration and discovery mechanism: Skill registration center, similar to an app store, each skill has a unique ID, version number, and function tags. It supports searching by business scenario, such as points operation, and supports version management, such as retaining compatibility with the old version when upgrading.

[0042] Skill reuse methods: direct invocation, i.e., input parameters, execute, and get results; combination and arrangement: multiple skills are combined according to DAG (directed acyclic graph) to form complex capabilities; inheritance and extension: based on existing skills, customize and reuse core logic.

[0043] Specifically, examples of accumulated experience are shown in Table 1.

[0044] Table 1

[0045] Optionally, each skill unit corresponds to an aggregation validator and a query rewriter. The skill unit pre-stores a field blacklist, which includes unauthorized sensitive fields. In step 1022, the skill unit checks whether the returned data conforms to the aggregation-level data rules, including: Step 10221: The skill unit checks whether the number of rows in the initial data exceeds a preset grouping threshold using an aggregation validator. The initial data with more than the preset grouping threshold represents individual-level data disguised as aggregated data. Step 10222: The skill unit checks whether the initial data includes unauthorized sensitive fields through a field blacklist; Step 10223: The skill unit checks whether the initial data comes from a valid query containing grouping criteria and aggregate functions through the query rewriter; Step 10224: When the number of rows in the initial data does not exceed the preset grouping number threshold, does not contain unauthorized sensitive fields, and comes from a legitimate query, the initial data is determined to conform to the aggregation-level data rules. The initial data that conforms to the aggregation-level data rules is determined to be aggregation-level data.

[0046] In the above embodiments of this application, the skill unit may use a security verification mechanism, specifically including: Aggregation validator: Automatically checks whether the number of returned rows exceeds the expected number of groups, preventing the leakage of individual data in disguised aggregations; Field blacklist: Ensures that Skill cannot return unauthorized sensitive fields to the agent.

[0047] Query rewriter: It can be used to automatically add GROUP BY and aggregate functions to SQL, force aggregation, and can also be used to check whether the initial data comes from a valid query that includes grouping criteria and aggregate functions.

[0048] Specifically, the preset grouping threshold can be set to, for example, 10. When the initial number of data rows exceeds 10, it can be identified as individual detail data disguised as aggregation results, triggering the interception rule. When the number of returned rows exceeds the preset grouping threshold, GROUP BY and aggregate functions can be automatically added to the SQL. That is, GROUP BY grouping rules and aggregate functions such as COUNT and SUM are automatically added to the original SQL (Structured Query Language) statement, forcing the detail data to be gathered into the aggregation result according to the specified dimensions, avoiding the direct output of all individual data.

[0049] GROUP BY: A grouping statement in SQL used to divide query results into several groups based on specified columns. It must be used in conjunction with aggregate functions to group rows with the same value together for statistical calculations.

[0050] COUNT: An aggregate function in SQL used to count the number of rows or the number of non-null values ​​in a specified column.

[0051] SUM: An aggregate function in SQL that calculates the sum of all non-null values ​​in a specified column for numeric fields only. It is commonly used for summarizing and statistically analyzing monetary or quantity data.

[0052] Unauthorized sensitive fields, also known as privacy fields that have not been properly authorized, may include sensitive information that can directly pinpoint an individual, such as a user's ID number, mobile phone number, bank card number, or transaction password.

[0053] Step 103: The intelligent agent determines different user behavior characteristics based on the aggregated data fed back by the multi-skill units, and generates customized initial service information adapted to each user behavior characteristic. Step 104: When any customized initial service information passes the compliance verification, it is identified as customized service information and stored in the temporary storage area. Different user behavior characteristics correspond to different customized service information.

[0054] Next, the intelligent agent can determine different user behavior characteristics based on the aggregated data returned by the multi-skill units, and then generate customized initial service information. When any customized initial service information passes the compliance verification, it is identified as customized service information and stored in the temporary storage area to prepare for subsequent targeted push.

[0055] Optionally, the agent corresponds to an inference decision-maker. In step 103, the agent determines different user behavior characteristics based on aggregated data fed back by multi-skill units, including: Step 1031: The agent sequentially acquires the aggregated data fed back by each skill unit; Step 1032: Based on all aggregated data, the agent calculates cross-dimensional user group behavior correlation statistics. The dimensions include the historical behavior event aggregation statistics dimension, the user behavior feature aggregation value dimension, and the group size dimension. The user group behavior correlation statistics are a set of correlation mappings between different dimensions.

[0056] Step 1033: The intelligent agent extracts differentiated user behavior features by using the inference decision-maker and cross-dimensional user group behavior correlation statistics.

[0057] In the above embodiments of this application, autonomous closed-loop decision-making by the AI ​​Agent (intelligent agent) is achieved. Existing intelligent information generation tools are mostly single-point functions, such as only performing analysis or only generating information. This requires manual coordination of multiple tools to complete the entire process; for example, data analysts use SQL to extract data, operations personnel formulate strategies, copywriters write content, and technical personnel configure delivery. This approach leads to several problems, such as fragmented processes: information is lost during transmission, significant misunderstandings of requirements, lengthy cycles: the entire process takes 2-5 days, missing the optimal push window, and labor-intensive: each project requires 3-5 people, making scalability impossible.

[0058] The concept of the above embodiments in this application is an intelligent agent that autonomously completes the entire process, drawing inspiration from the working methods of human experts. That is, after receiving a task, an expert autonomously breaks down the steps, calls upon tools, and integrates the results. This application simulates this process by having an AI Agent autonomously plan and execute, achieving a key technological paradigm such as ReAct (Reasoning + Acting), which enables cyclical execution until completion. Thought: What is the current state, and what should be done next? Action: Invokes a specific Skill to perform an operation; Observation: What result did the Skill return?

[0059] Furthermore, this can be achieved through a task planner. Input: Business rule configuration, in JSON format.

[0060] Output: A DAG (Directed Acyclic Graph) task graph, used to describe the execution order and dependencies of Skills. A DAG for a points-based user acquisition and activation scenario would look like this... Figure 3 As shown, Figure 3 middle: The data reading skill reads the core data of users with points, such as the user's points balance, acquisition / consumption records, behavioral patterns (such as the most recent login time and the frequency of points usage), and user attributes (such as new user / old user tags), providing a data foundation for subsequent segmentation and copywriting customization.

[0061] The data cleaning skill preprocesses the read data: Handle missing values, such as records where the user's points acquisition time is empty; Filter outliers, such as unreasonably large or negative points; Standardize data formats, such as timestamps and numerical units, to ensure data accuracy and reliability, and support subsequent modeling.

[0062] The data modeling skill uses a variant of the RFM model, which is based on the points-based scenario optimized by Recency (recent points-based behavior), Frequency (points usage frequency), and Monetary (total points value). Users are divided into 12 subgroups, for example: Highly active and high-value groups are characterized by frequent recent use of points and high total points. Dormant user group, representing users who have not used their points for a long time; The new user group represents those who have just earned points but have not yet spent them, providing a basis for differentiated activation strategies.

[0063] The compliance review skill performs compliance checks on the subsequently generated copy content to ensure it complies with advertising laws and platform points rules, such as the description of points validity period and the transparency of activity rules, to prevent illegal content from reaching users.

[0064] The copywriting generation skill generates 3 personalized copy versions for each of the 12 subgroups, totaling 36 copy pieces. For example: For dormant users: Emphasize that "points are about to expire, redeem benefits immediately"; For new users: Highlight "Extra discounts for first-time users with points"; For highly active users: Recommend "exclusive points redemption benefits" to enhance the appeal of the copy.

[0065] The Skill Insights feature analysis of each subgroup, such as the dormancy duration of inactive users and the redemption preferences of high-value users, to generate operational insights, such as "80% of inactive users have not logged in for 3 months," helping to further adjust activation strategies.

[0066] The Version Selection Skill predicts the CTR (click-through rate) of each copy and selects the optimal copy version for each subgroup to ensure maximum reach.

[0067] The file encapsulation skill maps the optimal copy to the corresponding user group into an execution table in JSON format, which is easy for the operation system to call directly, so as to accurately reach users and achieve the goal of acquiring new users and increasing user activity.

[0068] The entire process is automated through the skills of the AI ​​Agent, from data to strategy implementation, which efficiently promotes user acquisition and activity in the points-based system.

[0069] A dynamic adjustment mechanism can also be added. If the data read returns an empty set, skip the subsequent Skill and directly return no target user. In particular, if the insight finds that the characteristics of a certain subgroup are not significant, merge it into the adjacent subgroup and remodel. If the copy compliance review fails continuously, reduce the generation temperature parameter and retry the generation.

[0070] For the reasoning decision-maker (Reasoner), which utilizes reasoning capabilities based on large models, it performs high-level analysis on the Skill's returned results, for example: Based on the first group of 52,300 people with an average score of 680, it can be inferred that this group is sensitive to quality, and limited-edition exclusivity should be emphasized. Based on the 156,200 inactive users, it can be inferred that a reactivation mechanism needs to be designed to prevent points from expiring and being lost.

[0071] Specifically, when the agent sequentially acquires aggregate-level data, it strictly follows the execution order of the directed acyclic graph, serially receiving the aggregate-level data returned by each skill unit after verification. Each received data is temporarily stored in a dedicated cache. If a skill unit does not return aggregate-level data, the subsequent process is terminated directly and a flag indicating that no aggregate-level data was returned for that skill unit is returned.

[0072] The agent aligns the statistical scope of all aggregated data in the dedicated cache area, cross-binds the three dimensions of historical behavior event aggregation statistics, user behavior feature aggregation values, and group size, and generates a set of association mappings that only stores the corresponding relationships between dimensions. This set serves as the user group behavior association statistics result, which can be used to extract differentiated user behavior features. No individual-level data is generated throughout the process.

[0073] Optionally, the association mapping set corresponds to association distribution features. In step 1033, the agent extracts differentiated user behavior features through the inference decision-maker and cross-dimensional user group behavior association statistics, including: Step 10331: The agent performs unsupervised clustering on the cross-dimensional user group behavior association statistics through the inference decision-maker, and automatically groups user groups with similar association distribution characteristics into the same user group cluster. Step 10332: For any user group cluster, the agent extracts user behavior features by using the inference decision-maker and the aggregated statistical results of new historical behavioral events in the user group cluster, the aggregated values ​​of new user behavior features, and the new group size, until the differentiated user behavior features of each user group cluster are obtained.

[0074] In the above embodiments of this application, the inference decision-maker performs unsupervised clustering on the association mapping set, classifies aggregated entries with similar association distributions into the same user group cluster, and automatically generates differentiated user behavior features that match the sub-group attributes based on the sub-dimensional aggregated statistical features corresponding to the user group cluster.

[0075] Furthermore, after the intelligent agent's reasoning decision-making unit obtains the cross-dimensional user group behavior correlation statistics, it extracts the distribution characteristics of all related items and uses the K-Means unsupervised clustering algorithm to automatically group 85,700 new users with similar distribution characteristics, such as "high proportion of privilege browsing" and "high frequency of points redemption", into the same user group cluster, without involving any individual detailed data.

[0076] Specifically, the associated entries are the basic binding units in the cross-dimensional user group behavior correlation statistics results. The entire process is solely an aggregation mapping of the user group dimension and does not involve any individual detailed data. For example: Among newly activated credit card users in the past 7 days, the subgroup that browsed benefits ≥ 3 times corresponds to an average monthly points redemption frequency of 2.1 times, with a subgroup size of 85,700 people; Among newly activated credit card users in the past 7 days, the subgroup with fewer than 1 access to benefits has an average monthly points redemption frequency of 0.3 times, and the subgroup size is 40,300 people. Related entries also carry the binding correspondence between different aggregation dimensions.

[0077] For the user group cluster obtained by this clustering, the intelligent agent extracts exclusive new aggregated data: total number of browsing of sub-group benefits is 192,000 times, average monthly points redemption is 2.1 times, and group size is 85,700 people. Based on the aggregated statistical characteristics of these three sub-dimensions, the differentiated user behavior characteristics of "high frequency preference for physical benefits" are extracted.

[0078] Optionally, in step 104, before any customized initial service information is determined to be customized service information after passing compliance verification, the method further includes: Step 1041: The intelligent agent performs a two-layer compliance check on the customized initial service information, wherein the two-layer compliance check includes a rule engine layer check and an audit layer check. Step 1042: When the agent performs rule engine layer verification on the customized initial service information, it intercepts preset prohibited keywords in the customized initial service information by keyword matching, and intercepts content that is a violation in the customized initial service information by regular expression matching. Step 1043: When the intelligent agent performs review layer verification on the customized initial service information, it calls the large model to perform semantic understanding on the customized initial service information, and determines whether the customized initial service information contains implicit profit promises, whether the expression is compliant, and whether the tone is inflammatory. Step 1044: When both the rule engine layer verification and the review layer verification pass, the agent will determine the customized initial service information as customized service information. Step 1045: If any verification fails, the agent automatically generates a repair suggestion and feeds it back to the customized initial service information generation stage for regeneration. If it fails for a preset number of consecutive times, it will be downgraded to backup service information.

[0079] In the above embodiments of this application, a compliance embedded generation mechanism can be introduced. The content of financial service information must comply with the following: it must not promise principal or return, it must not use exaggerated or misleading statements, and it must clearly indicate the identity of the sender.

[0080] Existing review methods are mostly post-event reviews, that is, content is generated first and then reviewed manually or by tools. This may result in the violation content already being published. In addition, the use of rule filtering, that is, based on keyword matching, makes it impossible to identify semantic violations, such as variations of the phrase "guaranteed profit".

[0081] The core concept of the above embodiments of this application is to embed compliance during the generation process, rather than remediation afterward. Specifically, this is achieved through a two-layer compliance mechanism, including: Rule engine layer: hard rules, based on keywords and regular expressions, to block known violations.

[0082] AI review layer: soft rules, based on large model semantic understanding, to identify variant expressions and contextual violations.

[0083] Furthermore, compliance embeddings are generated within the pipeline, and the copywriting generation skill execution flow is implemented, for example: The large model generates the initial text, and the rule engine performs hard filtering: Disable keyword matching: Directly block entries containing terms like "guaranteed principal," "guaranteed interest," or "guaranteed profit." Regular expression pattern matching: Automatically intercepts when the expected return is found; In particular, short link validity checks can be added to ensure accessibility.

[0084] Next, AI semantic review will be conducted: Large-scale model assessment: Does this copy imply a profit promise? Contextual analysis: Is this statement compliant with regulations? Sentiment analysis: Is the tone overly inflammatory? Next, the review results are processed: Access: Enter version preference; Failed: Automatic repair suggestion, feedback to the generation stage for retry; Consecutive (e.g., 5 consecutive failures): Degrade to a safe template.

[0085] The key design of the compliance-embedded generation mechanism is the integration of review and generation. Unlike reviewing separately after generation, this application embeds the review logic into the generation skill. The generation parameters include compliance constraints, such as prohibited words and tone requirements. The generation process performs real-time self-checks, similar to self-censorship when writing. It automatically repairs itself when the review fails, rather than simply rejecting it.

[0086] Step 105: In response to a user behavior event, the intelligent agent determines the user behavior characteristics corresponding to the user behavior event, reads the customized service information corresponding to the user behavior characteristics from the temporary storage area, and sends the customized service information to the user's preset terminal.

[0087] Next, user behavior events, such as actions triggered by users within the bank's app that can be captured by the system in real time, such as a newly activated credit card user opening the app to enter the "My Benefits" page, or a user clicking to view the points redemption portal.

[0088] The extraction of user behavior features can be achieved by collecting user operation behavior data in real time through APP tracking points, and combining it with pre-generated user group aggregation tags. From the dimensions of operation time, page type visited, dwell time, and historical operation preferences, corresponding user behavior features can be extracted. Throughout the process, sensitive detailed data that can be located to individuals will not be collected, and only compliant aggregation feature results will be output.

[0089] The extracted user behavior features, such as card activation in the past 7 days and intention to redeem points, are non-sensitive tags extracted from aggregated data that are adapted to user behavior and fully comply with the security requirements of aggregated data.

[0090] For example, when a user triggers a behavior event to enter the "My Benefits" page, the intelligent agent first matches the user's pre-stored user behavior characteristics in real time, and then accurately retrieves customized service information that is fully adapted to the user's behavior characteristics from the temporary storage area. This information is then directly pushed to the user's bank APP as a message pop-up or to a preset terminal (such as a mobile phone), without the need to generate content in real time, thus significantly reducing response latency.

[0091] By applying the technical solution of this embodiment, a skill-based data isolation mechanism is established, encapsulating data analysis logic into standardized interfaces. The AI ​​Agent only calls these interfaces to obtain aggregated results without accessing the original data. Autonomous closed-loop decision-making by the AI ​​Agent allows it to autonomously complete the entire process from requirement analysis to task planning, skill scheduling, and result integration, eliminating the need for manual intervention. This achieves the skill-based accumulation of expert experience, encapsulating senior analysts' user selection logic and copywriting skills into reusable skills. Coupled with a compliance-embedded generation mechanism, compliance review is embedded during the copy (service information) generation process, rather than post-event review.

[0092] In one specific embodiment, the complete process for attracting and activating new users with points is as follows: A financial institution operates a points system and hopes to reach its points users via SMS, with messages including a request to follow their official WeChat account, to increase points activity. The target user base is approximately 500,000.

[0093] System deployment examples include: hardware environment: AI Agent service: K8s cluster, 4 cores, 8G x 3 nodes; Skill gateway: independently deployed; data platform accessible via intranet; data platform: existing cluster, no need to add new one; caching service: Redis Cluster, 6 nodes; large model: general-purpose large language model.

[0094] Network environment: The AI ​​Agent is located in the DMZ zone (10.1.0.0 / 24) and can only access the Skill Gateway. The data platform is located in the core zone (10.2.0.0 / 24) and can only be accessed by the Skill Gateway. There is no direct route between the two zones and the route must be relayed through the Skill Gateway.

[0095] Execution process example: Rule configuration: The operations staff configured the following rules on the service information generation platform: Rule Name: Points User Acquisition and Activation Activity; Validity Period: 2026-05-01 to 2026-05-31; Target Users: Points users, active users in the past 30 days, who have not followed the official account; Triggering Conditions: Login to the points page / 7 days of inactivity / points about to expire; Differentiation Dimensions: Points balance, activity frequency, redemption preferences; Generation Constraints: ≤70 characters, including a short link to the official account, and a friendly and caring tone.

[0096] AI Agent Analysis and Planning: After receiving the rules, the AI ​​Agent analyzes the business intent: identifying it as a user acquisition and activation scenario, requiring layering and generation, generating a DAG task graph: Node 1: Data Reading Skill, reads basic data of users with points; Node 2: Data Cleaning Skill, handles missing values ​​and outliers; Node 3: Data Modeling Skill, RFM variant layering, 12 subgroups; Node 4: Insight Generation Skill, analyzes the characteristics of each subgroup; Node 5: Copywriting Generation Skill, 12 subgroups × 3 versions = 36 copywritings; Node 6: Compliance Review Skill, reviews all 36; Node 7: Version Optimization Skill, predicts CTR and selects the optimal version; Node 8: File Encapsulation Skill, assembles into a JSON mapping table.

[0097] Specifically, CTR, or Click-Through Rate, is a core metric in internet operations. It refers to the ratio of actual clicks to total impressions of content / ads. The formula is: CTR = Actual Clicks ÷ Impressions × 100%. A higher CTR indicates stronger appeal to users. In user acquisition and engagement scenarios, it is used to predict the click-through rate of different copywriting and to select the optimal reach copy.

[0098] Next, Skill execution and data isolation are performed. Taking the data reading Skill as an example, after the Skill receives the input, it performs internal execution: connects to the data platform, executes an SQL query (not visible to AI), and then the Skill returns the output: aggregated results, without individual details.

[0099] Next, personalized content generation is performed: taking the first group of people - active - with a strong preference for physical goods as an example: Input profile: 680 points, logged in 5 times in the last 7 days, prefers 3C redemption; Strategy matching: urgency (limited time offer) + quality assurance; Version A: Mr. Zhang, you have 680 points that are about to expire. Redeem them for a smart bracelet for a limited time. Click the link >> [short link].

[0100] Compliance review: Passed. No prohibited words or profit promises.

[0101] Predicted performance: CTR 3.8%, highest, selected as the version to send.

[0102] Next, cache loading and real-time triggering are performed: Personalized content for all 498,000 users is written to the Redis cache. When a user logs into the app to view their points page, the behavior event is reported, the platform queries the cache, retrieves the personalized content, and calls the SMS delivery system, so that the user receives a differentiated SMS message.

[0103] Specifically, AI Agent refers to an intelligent program that can autonomously think, plan, and execute tasks; Skill refers to a standardized software module that encapsulates specific capabilities and can be invoked and executed; ReAct refers to a technological paradigm that allows AI to alternate between thinking and acting, such as a reasoning-action loop mechanism; DAG refers to a directed acyclic graph that describes the order of task execution and dependencies, such as a task execution plan graph; Token refers to the smallest unit of text processing in a large model, with a quantity limit, which is also a unit of context processing capacity; Few-shot represents giving AI several examples to imitate and generate, which is also an example-guided generation strategy; RFM refers to a model that analyzes user value based on recent consumption, consumption frequency, and consumption amount, which can also be called a user value analysis model.

[0104] Furthermore, as Figure 2 In terms of specific implementation, this application provides a customized service information generation system, such as... Figure 4 As shown, the system includes: The intelligent agent decision-making center module 201 is used to receive customized service information generation tasks, determine multiple skill units required to complete the customized service information generation tasks, and generate a directed acyclic graph representing the execution order and dependency relationship of the skill units. The skill units are preset modular functional components. The customized service information generation tasks are used to generate customized service information separately for different user behavior characteristics. The data security isolation module 202 is used to sequentially call each skill unit based on the directed acyclic graph, so that each skill unit processes the execution result into aggregated data and then feeds it back. The aggregated data is the statistical result of the user group dimension, which includes the aggregated statistical result of historical behavior events, the aggregated value of user behavior characteristics, and the group size. The personalized content generation module 203 is used to determine different user behavior characteristics based on the aggregated data fed back by the multi-skill units, and generate customized initial service information adapted to each user behavior characteristic. The information is verified by the compliance module 204, which determines that any customized initial service information is customized and stores it in the temporary storage area when the compliance verification is passed. Different user behavior characteristics correspond to different customized service information. The real-time triggering and caching module 205 is used to respond to user behavior events by reading customized service information that matches the user behavior characteristics corresponding to the user behavior event from the temporary storage area and sending it to the user's preset terminal.

[0105] In the above embodiments of this application, the AI ​​Agent decision-making central module 201 can be used to receive business rule configurations, autonomously parse business intents, generate task execution plans (i.e., DAGs, directed acyclic graphs), and be responsible for scheduling the collaborative execution of each Skill module. The AI ​​Agent decision-making central module can adopt the ReAct inference paradigm to perform dynamic DAG planning and context memory management.

[0106] The data security isolation module 202 ensures that the AI ​​Agent never accesses the original detailed data, only obtaining aggregation results through the Skill interface. This can be achieved through network isolation, field-level access control, aggregation verification, and end-to-end auditing.

[0107] The personalized content generation module 203 can be used to generate personalized SMS content tailored to each user based on user insight results and by calling a large model. This can be achieved through Few-shot prompt engineering, selling point matching strategy, compliance embedded review, and effect prediction optimization.

[0108] The real-time triggering and caching module 204 can be used to respond to user behavior events, read personalized content from the cache, and deliver it to the user in real time. This can be achieved through streaming event processing, distributed caching, preheating, and consistency assurance.

[0109] Specifically, the customized service information generation system may also include a Skill invocation orchestration module 205, used to manage Skill registration, discovery, invocation, and version control; supporting parallel execution of multiple Skills and result assembly. It has a corresponding standardized interface protocol, enabling topology sorting scheduling and automatic fault recovery.

[0110] It should be noted that other corresponding descriptions of the functional units involved in the customized service information generation system provided in this application embodiment can be found in the following references. Figure 2 The corresponding descriptions in the method will not be repeated here.

[0111] Based on the above, Figure 2 Accordingly, this application also provides a storage medium storing a computer program, which, when executed by a processor, implements the above-described method. Figure 2 The method for generating customized service information is shown.

[0112] Based on this understanding, the technical solution of this application can be embodied in the form of a software product. The software product can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, or portable hard drive), and includes several instructions to cause a computer device (such as a personal computer, server, or network device) to execute the methods described in the various implementation scenarios of this application.

[0113] Based on the above, Figure 2 The method shown, and Figure 4 To achieve the above objectives, the present application also provides a computer device, such as the virtual system embodiment shown. Figure 5 As shown, specifically, it can be a personal computer, server, network device, etc. This computer device includes a non-volatile storage medium, a processor, a system bus, internal memory, and a network interface. The non-volatile storage medium includes an operating system, computer programs, and a database; the non-volatile storage medium is used to store the computer programs; the processor is used to execute the computer programs to achieve the above-mentioned functions. Figure 2 The method for generating customized service information is shown.

[0114] Optionally, embodiments of this application also provide another computer device, such as... Figure 6 As shown, the computer device includes a non-volatile storage medium, a processor, a system bus, internal memory, a network interface, an input device, and a display screen.

[0115] Optionally, the computer device may also include a user interface, camera, radio frequency (RF) circuitry, sensors, audio circuitry, Wi-Fi module, etc. The user interface may include input units such as a keyboard, and optional user interfaces may also include USB interfaces, card reader interfaces, etc. Network interfaces may optionally include standard wired interfaces, wireless interfaces (such as Bluetooth interfaces, Wi-Fi interfaces), etc.

[0116] Those skilled in the art will understand that the computer device structure provided in this embodiment does not constitute a limitation on the computer device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0117] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages and stores the hardware and software resources of a computer device, supporting the operation of information processing programs and other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software within the physical device.

[0118] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented using software plus necessary general-purpose hardware platforms, or it can be implemented using hardware to determine multiple skill units required to complete the customized service information generation task, generate a directed acyclic graph, sequentially call each skill unit, and receive the returned aggregated data for fusion calculation to generate customized initial service information for different user behavior characteristics. When any customized initial service information passes compliance verification, it is determined as customized service information and stored in a temporary storage area. The intelligent agent responds to user behavior events, reads customized service information matching the user behavior characteristics corresponding to the user behavior event from the temporary storage area, and sends it to the user's preset terminal. It can achieve full-process automation from intelligent analysis to personalized information content generation without exposing original sensitive data, and solve the technical contradiction between data security and intelligent agent performance.

[0119] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application. Those skilled in the art will understand that the modules in the system of the embodiment scenario can be distributed throughout the system of the embodiment scenario as described, or they can be modified to reside in one or more systems different from this embodiment scenario. The modules of the above-described embodiment scenario can be combined into one module, or further divided into multiple sub-modules.

[0120] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0121] The serial numbers in this application are for descriptive purposes only and do not represent the superiority or inferiority of any particular implementation scenario. The above disclosures are merely a few specific implementation scenarios of this application; however, this application is not limited thereto, and any modifications that can be made by those skilled in the art should fall within the protection scope of this application.

Claims

1. A method for generating customized service information, characterized in that, The method includes: The intelligent agent receives a customized service information generation task, determines multiple skill units required to complete the customized service information generation task, and generates a directed acyclic graph representing the execution order and dependencies of the skill units. The skill units are preset modular functional components. The customized service information generation task is used to generate customized service information separately for different user behavior characteristics. The agent sequentially calls each skill unit based on the directed acyclic graph, so that each skill unit processes the execution result into aggregated data and feeds it back to the agent. The aggregated data is the statistical result of the user group dimension, which includes the aggregated statistical result of historical behavior events, the aggregated value of user behavior characteristics, and the group size. The intelligent agent determines different user behavior characteristics based on aggregated data fed back by multi-skill units, and generates customized initial service information adapted to each user behavior characteristic. When any customized initial service information passes the compliance verification, it is identified as customized service information and stored in the temporary storage area. Different user behavior characteristics correspond to different customized service information. In response to a user behavior event, the intelligent agent determines the user behavior characteristics corresponding to the user behavior event, reads the customized service information corresponding to the user behavior characteristics from the temporary storage area, and sends the customized service information to the user's preset terminal.

2. The method according to claim 1, characterized in that, The agent sequentially calls each skill unit based on a directed acyclic graph, so that each skill unit processes the execution result into aggregated data and feeds it back to the agent, including: The agent determines multiple skill units to be scheduled, as well as the execution order and dependencies between the skill units, based on the directed acyclic graph. The agent schedules skill units based on the execution order and dependencies. When any scheduled skill unit completes its execution and obtains initial data, the skill unit checks whether the initial data conforms to the aggregation-level data rules and feeds back the aggregation-level data that conforms to the aggregation-level data rules to the agent.

3. The method according to claim 2, characterized in that, Each skill unit corresponds to an aggregation validator and a query rewriter. Each skill unit pre-stores a field blacklist, which includes unauthorized sensitive fields. The process of the skill unit checking whether the initial data conforms to the aggregation-level data rules includes: The skill unit checks whether the number of rows in the initial data exceeds a preset grouping threshold using an aggregation validator. The initial data with more rows than the preset grouping threshold represents individual-level data disguised as aggregate-level data. The skill unit checks whether the initial data includes unauthorized sensitive fields by using a field blacklist. The skill unit checks whether the initial data comes from a valid query containing grouping criteria and aggregate functions through the query rewriter; If the number of rows in the initial data does not exceed the preset grouping threshold, does not contain unauthorized sensitive fields, and comes from a legitimate query, the initial data is determined to conform to the aggregation-level data rules. The initial data that conforms to the aggregation-level data rules is determined to be aggregation-level data.

4. The method according to claim 1, characterized in that, The intelligent agent corresponds to an inference decision-maker, which determines different user behavior characteristics based on aggregated data fed back from multiple skill units, including: The agent sequentially acquires aggregated data from each skill unit. Based on all aggregated data, the intelligent agent integrates and calculates cross-dimensional user group behavior correlation statistics. The dimensions include the historical behavior event aggregate statistics dimension, the user behavior feature aggregate value dimension, and the group size dimension. The user group behavior correlation statistics are a set of correlation mappings between different dimensions. The intelligent agent extracts differentiated user behavior characteristics by using inference decision-makers and cross-dimensional user group behavior correlation statistics.

5. The method according to claim 4, characterized in that, The association mapping set corresponds to association distribution characteristics. The intelligent agent extracts differentiated user behavior characteristics through the inference decision-maker and cross-dimensional user group behavior association statistics, including: The intelligent agent performs unsupervised clustering on the cross-dimensional user group behavior correlation statistics through the inference decision-maker, and automatically groups user groups with similar correlation distribution characteristics into the same user group cluster. For any user group cluster, the agent extracts user behavior features by using the inference decision-making unit and the aggregated statistical results of new historical behavioral events in the user group cluster, the aggregated values ​​of new user behavior features, and the new group size, until the differentiated user behavior features of each user group cluster are obtained.

6. The method according to claim 1, characterized in that, Before any customized initial service information is determined to be customized service information and stored in the temporary storage area after passing compliance verification, the method further includes: The intelligent agent performs a two-layer compliance check on the customized initial service information, wherein the two-layer compliance check includes a rule engine layer check and an audit layer check; When the agent performs rule engine layer verification on the customized initial service information, it intercepts preset prohibited keywords in the customized initial service information through keyword matching, and intercepts content that is a violation in the customized initial service information through regular expression matching. When the intelligent agent performs review layer verification on the customized initial service information, it calls the large model to perform semantic understanding on the customized initial service information, and judges whether the customized initial service information contains implicit profit promises, whether the expression is compliant, and whether the tone is inflammatory. When both the rule engine layer verification and the review layer verification pass, the agent will determine the customized initial service information as the customized service information. If any verification fails, the agent automatically generates a repair suggestion and feeds it back to the customized initial service information generation stage for regeneration. If it fails for a preset number of consecutive times, it will be downgraded to backup service information.

7. The method according to claim 1, characterized in that, Each skill unit corresponds to an input parameter, and the input parameters of each skill unit have dependencies. The agent generates a directed acyclic graph representing the execution order and dependencies of the skill units, including: The agent determines the dependency links between skill units based on the input parameter dependencies of each skill unit, marks skill units without dependencies as skill units that can be executed in parallel, and marks skill units with sequential dependencies as skill units that can be executed sequentially. The agent integrates all skill units, their execution order, and dependencies into a loop-free topology, generating a directed acyclic graph that represents the execution order and dependencies of the skill units.

8. A customized service information generation system, characterized in that, The system includes: The intelligent agent decision-making center module is used to receive customized service information generation tasks, determine multiple skill units required to complete the customized service information generation tasks, and generate a directed acyclic graph representing the execution order and dependencies of the skill units. The skill units are preset modular functional components. The customized service information generation tasks are used to generate customized service information separately for different user behavior characteristics. The data security isolation module is used to sequentially call each skill unit based on a directed acyclic graph, so that each skill unit processes the execution results into aggregated data and then feeds them back. The aggregated data is the statistical results of the user group dimension, including the aggregated statistical results of historical behavioral events, the aggregated values ​​of user behavior characteristics, and the group size. The personalized content generation module is used to determine different user behavior characteristics based on aggregated data from multi-skill units and generate customized initial service information adapted to each user behavior characteristic. The information is verified by the compliance module. When any customized initial service information passes the compliance verification, it is identified as customized service information and stored in the temporary storage area. Different user behavior characteristics correspond to different customized service information. The real-time triggering and caching module is used to respond to user behavior events, determine the user behavior characteristics corresponding to the user behavior events, read the customized service information corresponding to the user behavior characteristics from the temporary storage area, and send the customized service information to the user's preset terminal.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method for generating customized service information as described in any one of claims 1 to 7.

10. A computer device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method for generating customized service information as described in any one of claims 1 to 7.

Images