A deep research dynamic interactive achievement intelligent agent interaction system generation method

Abstract

The application provides a deep research dynamic interactive achievement intelligent agent interaction system generation method and device and a storage medium. The method comprises the following steps: providing an interface template containing placeholders and a component library containing a plurality of components, each component having a corresponding filling function; an intelligent agent performs information retrieval and analysis according to a user input research question, and generates a structured analysis result; the intelligent agent calls the corresponding filling function to fill data into the component according to the structured analysis result and configuration parameter information of the component to be filled, and generates filled component code; and the filled component code is replaced with the corresponding placeholder in the template to generate a complete interactive interface. The application realizes full-process automation from intelligent agent information retrieval and analysis to interactive interface generation, and significantly improves the flexibility, expandability and interface generation efficiency of the system through decoupled design of the template library and the component library.

Description

Technical Field

[0001] This invention relates to the field of system generation technology, and more specifically to a method for generating intelligent agent interaction systems that deeply study dynamic interaction results. Background Technology

[0002] With the rapid development of artificial intelligence (AI) technology, agent AI systems have demonstrated enormous potential in multiple fields, including internet retrieval and research tasks. These systems, leveraging the powerful reasoning capabilities of large language models (LLMs) and multimodal interaction technologies, are able to perceive their environment, plan tasks, and execute complex operations. Among these, research-oriented agents, which take user research needs as input, retrieve data from a range of data sources such as the internet, and ultimately produce research reports, have attracted significant attention. However, existing technologies still have significant limitations in transforming the research results generated by agents into structured, visual, and interactive interfaces. Furthermore, while some systems can generate dynamic pages online, the interactive data within these pages often relies on the original system environment, databases, or remote API interfaces. When users need to save, archive, or review the generated results offline, if the external data source is inaccessible, interactive functions such as charts, tables, filters, and linked displays may not function properly, making it difficult to deliver complete results and achieve long-term reproducibility.

[0003] Traditional intelligent agent systems primarily focus on backend data processing and decision generation, but their integration with the frontend user interface often relies on predefined static templates or hard-coded rendering logic. This approach results in a severe lack of system flexibility: when users want to customize the display format of results according to different application scenarios (such as academic research, market analysis, content creation, etc.), developers need to manually adjust the code, which is not only inefficient but also makes it difficult to quickly respond to changing needs. Furthermore, existing interface generation technologies generally lack the ability to dynamically combine component-level elements. While intelligent agents can acquire and process information, they cannot intelligently call and assemble basic components (such as charts, tables, and navigation bars) from the frontend component library, leading to highly homogenized final pages that fail to reflect the inherent characteristics of the data and the user's personalized preferences in layout and interaction logic.

[0004] Regarding agent collaboration frameworks, while existing technologies emphasize role division and collaborative processes, their collaborative results (such as research reports and data analysis conclusions) are typically output in unstructured text or simple data formats, rather than user interfaces rich in interactive elements that can be directly rendered. This means that additional, often cumbersome, human intervention is required to transform the agent's output into an interface that end users can intuitively understand and operate, hindering the realization of end-to-end automated experiences. Summary of the Invention

[0005] This invention provides a method for generating an intelligent agent interactive system that deeply integrates the dynamic reasoning capabilities of the intelligent agent with the standardization and reusability of a component library. It allows users to select or define templates, driving the intelligent agent to automatically call component library functions after completing information retrieval and analysis. This accurately fills the resulting data into appropriate components and ultimately assembles a complete, interactive dynamic page, significantly improving the expressiveness, interactivity, and generation efficiency of the intelligent agent system's output.

[0006] In a first aspect of the present invention, a method for generating an intelligent agent interaction system for in-depth research on dynamic interaction results is provided, comprising: Step 1: Provide at least one interface template. Each interface template contains layout structure information and at least one placeholder. Each placeholder is associated with a component identifier. Step 2: Provide a component library containing multiple components corresponding to component identifiers. Each component has a corresponding population function, which is used to receive structured data and generate renderable component code. Step 3: The agent performs information retrieval and analysis based on the research questions input by the user, and generates structured analysis results; Step 4: Based on the structured analysis results and the configuration parameter information of the component to be filled, the agent calls the filling function corresponding to the component to be filled, processes the structured analysis results and fills them into the component to be filled, generating the filled component code; Step 5: Replace the placeholders for the corresponding component identifiers in the interface template with the filled component code to generate a complete interactive interface.

[0007] In some optional embodiments of the present invention, step 3 includes: The intelligent agent acquires the research question input by the user and the description information of the selected interface template; The agent executes at least one round of the "think-execute-observe" cycle, in which the agent generates the search query text during the thinking phase, calls the search tool to obtain the search results based on the search query text during the execution phase, and receives the search results during the observation phase and uses them as input for the next round of thinking. The agent outputs structured analysis results after completing the loop.

[0008] In some optional embodiments of the present invention, step 4 includes: The agent acquires the structured analysis results and the configuration parameter information of all components to be populated. The configuration parameter information includes the component identifier of each component to be populated and the configuration parameters that the component can accept and their data types. For the current component to be filled, the agent generates filling parameters that conform to the input format of the component's filling function, based on the structured analysis results and the component's configuration parameter information. The agent calls the fill function, passes the fill parameters to the fill function, and generates the filled component code; The agent repeats the filling operation for the next component to be filled until all components to be filled are filled.

[0009] In some optional embodiments of the present invention, step 4 further includes: When the agent fails to call the fill function, the agent receives a failure message; The agent adjusts the filling parameters based on the failure information and calls the filling function again to fill the data.

[0010] In some optional embodiments of the present invention, the description information of the interface template includes a description of the application scenario to which the interface template is intended and / or a description of the data type to be analyzed; the agent determines the retrieval strategy and the analysis strategy based on the description information.

[0011] In some optional embodiments of the present invention, the intelligent agent includes a first intelligent agent and a second intelligent agent, step 3 is executed by the first intelligent agent and step 4 is executed by the second intelligent agent; the first intelligent agent and the second intelligent agent are each independently implemented based on a large language model.

[0012] In some optional embodiments of the present invention, the large language model of the first agent and the large language model of the second agent are respectively trained by reinforcement learning; Among them, the reinforcement learning reward function of the large language model of the first intelligent agent is determined based on whether the output conforms to the preset label format; The reinforcement learning reward function of the large language model of the second agent is determined based on whether the output conforms to the preset label format and whether the component filling function call is successful.

[0013] In some optional embodiments of the present invention, the retrieval tool is encapsulated with a unified query interface, and the retrieval tool retrieves information from configured information sources based on the query text, including the Internet and / or local databases.

[0014] In one embodiment of the present invention, a method for generating an intelligent agent interaction system based on a component library is provided. By introducing an external template library as a bridge between the intelligent agent and the component library, a dynamic and customizable interface generation framework is constructed. The core of this system lies in the fact that the predefined templates in the template library not only contain layout structures, but more importantly, they are bound to components and their interaction rules in a specific component library. After the intelligent agent completes an internet search and research task, it can automatically parse the type and structure of the result data according to the template selected by the user, intelligently call the corresponding component rendering function, and dynamically populate the data into the component instance. Finally, the system assembles the component code with complete interactive logic and data population according to the template layout to generate a final interactive interface that can be directly delivered and used.

[0015] This invention proposes a component-based dynamic interaction system generation technology for intelligent agents. It establishes a connection between the generated intelligent agent results and the front-end component rendering through an external template library. The interface templates in the template library are pre-configured with layout structures, placeholders, component IDs, and component parameter constraints, enabling the templates to simultaneously constrain the intelligent agent's research direction, component filling format, and interface assembly process. During system runtime, the intelligent agent completes retrieval and analysis based on the user's research question and the selected template, generating structured result data for component filling. Subsequently, the filling function of the corresponding component is called according to the component ID, writing the result data into the component instance. The filled component code is assembled according to the template placeholder positions, generating a directly displayable and interactive interface. Through this method, this invention achieves a collaborative connection between research questions, intelligent agent analysis, component filling, and interface generation, reducing manual coding conversion and improving interface generation efficiency, component reusability, and scene adaptability. Finally, this invention can also uniformly encapsulate the page structure, filled components, component interaction logic, and data results obtained during page generation after the interactive interface is generated, forming a savable HTML page file or page resource package.

[0016] The above method has the following advantages: Flexible Adaptability and Standardized Control: Through an external, extensible template and component library, the content generation capabilities of large models are combined with enterprise business specifications. Users can flexibly select or extend templates for different application scenarios such as data analysis reports, market comparison dashboards, and literature review pages, generating interactive interfaces with different styles and functions without modifying the core system code. Simultaneously, the structure, component attributes, and parameter constraints in the templates can standardize and limit the generated results, reducing issues such as randomness, inconsistent formats, and difficulty in meeting enterprise requirements in the generation of general large models. Large models only need to complete template parsing, content filling, and component parameter generation, without needing to directly write complete front-end code. This retains the flexibility of AI generation while improving the stability, controllability, and success rate of system deployment.

[0017] True end-to-end automation: This achieves full automation from information retrieval and analysis by the agent to the final generation of an interactive user interface. The agent can understand data semantics and decide on appropriate visualization and interactive components, automatically completing data binding and interface assembly. This eliminates the tedious steps required in traditional processes where front-end developers manually convert data results into interfaces, significantly improving efficiency. It also provides a solution for non-technical personnel to create their own data analysis interfaces and achieve their analytical goals.

[0018] Excellent decoupling and scalability: The system design decouples the agent, template library, and component library. Both the component and template libraries can be updated and extended independently of the agent's core logic. When new UI components or interface styles need to be introduced, they can be added to the component and template libraries without altering the agent's core decision-making logic, making system maintenance and upgrades much simpler.

[0019] Excellent results retention and reproducibility: This invention can not only generate dynamic interactive interfaces online, but also encapsulate the generated page structure, component code, interaction logic and related data results into HTML page files or page resource packages.

[0020] Specifically, the method includes the following steps: 1. The user provides a research question and selects a template from the template display.

[0021] 2. After receiving the user's research question and selected template, the system obtains the template's description, placeholder information, component ID, and corresponding component type, and inputs this information into the agent. Based on the research question and the data requirements of the template's components, the agent uses tools such as internet searches and local database queries to obtain relevant information. It then analyzes, summarizes, and organizes the search results to generate structured analysis results that match the template component filling requirements, preparing for subsequent component filling.

[0022] 3. The system displays information about the corresponding component population functions (e.g., renderLineChart(data), populateComparisonTable(data)) from the component library to the agent. Based on the structured data generated in the previous step, and combined with the unique identifier of each component to be populated in the template, the agent processes the data into the precise format required by the corresponding population function. Then, the agent actively calls these component population functions, passing the processed data as parameters, thereby completing the data binding and content generation for each specific component instance.

[0023] 4. The system receives all component instances that have been populated with data. Next, based on the layout configuration information of the selected template, the system automatically assembles the code of these component instances in a predetermined order and position, creating a fully logically connected interactive interface. Finally, the system renders and displays the ready-to-run interactive interface to the user.

[0024] 5. After assembling and displaying the interactive interface, the system encapsulates the generated page. Based on the component IDs and placeholder mappings in the template, the system extracts the page structure, populated component code, component interaction logic, component parameters, and data results that the page display depends on, and encapsulates them into HTML page files or page resource packages.

[0025] In a second aspect of the present invention, an apparatus for generating an intelligent agent interaction system for in-depth research on dynamic interaction results is provided, comprising: The template library storage module is used to store at least one interface template. Each interface template contains layout structure information and at least one placeholder. Each placeholder is associated with a component identifier. The component library storage module is used to store multiple components corresponding to component identifiers and the corresponding population function for each component. The population function is used to receive structured data and generate renderable component code. The intelligent agent reasoning module is used to perform information retrieval and analysis based on the research questions input by the user, and generate structured analysis results; The intelligent agent filling module is used to call the filling function corresponding to the component to be filled based on the structured analysis results and the configuration parameter information of the component to be filled, process the structured analysis results and fill them into the component to be filled, and generate the filled component code. The interface building module is used to replace the placeholders of the corresponding component identifiers in the interface template with the populated component code to generate a complete interactive interface.

[0026] In a third aspect of the present invention, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method for generating an intelligent agent interaction system for deep research dynamic interaction results as described in any of the embodiments of the first aspect.

[0027] The beneficial effects provided by this invention are: This invention establishes an external template library that can be dynamically parsed and invoked, enabling direct integration between the agent's generated results and the front-end component rendering process. This achieves end-to-end processing from user research question input to interactive interface generation. After the user selects an interface template, the system automatically reads the structured description information from the template, identifies placeholder markers and their corresponding component IDs, and obtains component attribute description information based on the component IDs. This generates a template parsing result containing placeholders, component IDs, and component parameter constraints. This parsing result is simultaneously applied to agent survey analysis and automatic component filling, ensuring that survey data is adapted to component requirements. During the filling phase, data is written to the corresponding components according to component IDs and parameter types, reducing manual conversion and secondary coding. Furthermore, this invention allows for changing or expanding interface templates based on different research scenarios and attaching text, charts, tables, and other component resources to the templates, enabling the same agent's survey capabilities to adapt to different interface organization forms and display needs. The template library, component library, and agent are linked through component IDs, placeholders, and attribute description information. Each module can collaboratively generate interfaces and also expand independently, thereby improving the system's generation efficiency and subsequent maintenance efficiency. Finally, after the interactive interface is generated, the page structure, populated component code, component interaction logic, component parameters, and data results on which the page display depends can be uniformly encapsulated to generate a savable HTML page file or page resource package. This invention not only achieves automatic generation of interactive interfaces but also further improves the interactivity and reproducibility of the generated results.

[0028] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0029] Figure 1 A detailed flowchart of the intelligent agent deep thinking execution technology provided in the embodiments of the present invention; Figure 2 An overall framework diagram of the intelligent agent interaction system for in-depth research on dynamic interaction results provided in this embodiment of the invention; Figure 3 A detailed flowchart of the intelligent agent component filling technology provided in the embodiments of the present invention; Figure 4 Detailed flowchart of page encapsulation technology provided for embodiments of the present invention; Figure 5 A detailed flowchart of the method for generating an intelligent agent interaction system for in-depth research dynamic interaction results provided in this embodiment of the invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0031] I. Style Library Module This module provides the basic resources needed to build dynamic interactive systems, including independently maintainable and expandable template and component libraries. After a user selects an interface template, the system generates template parsing results for constraint surveys, filling, and interface construction based on placeholder information in the template library and component attribute descriptions in the component library.

[0032] 1.1 Template Library Module This module primarily stores a series of predefined interface template resources for different scenarios required for building dynamic interactive systems. Each template resource is mainly a complete UI code, where component-related parts are replaced by placeholders, and each placeholder is associated with a unique component ID. When a template is selected, the system extracts the template description text, placeholder context, and the corresponding component ID from the template resource as components of the template parsing result.

[0033] 1.2 Component Library Module This module is primarily used for unified management of predefined component resources required during the construction of dynamic interactive systems. Each component establishes a mounting relationship with its corresponding interface template through a unique component identification ID, enabling the system to quickly locate, call, and instantiate the target component during template parsing. Each component resource includes a unique component ID, a core rendering function, and component attribute description information. The core rendering function receives structured data input and generates UI code that can be rendered by the front-end environment; the component attribute description information defines the configuration parameters and data types that the component can receive. This design allows the component library to provide the corresponding component's attribute description information and related parameters based on the component ID obtained from the template parsing results. This ensures that components are decoupled from the interface template while providing a unified standard for subsequent data generation and automatic component population.

[0034] II. In-depth Research Module for Intelligent Agents This module is used to conduct targeted research based on user-inputted research questions and template and component constraint information obtained from the parsing of selected templates. It then invokes search tools to generate structured data that meets component population requirements. During the research phase, this module incorporates template description text, placeholder context, component IDs, and component attribute descriptions from the template parsing results. This constrains the search direction, data type, and return format, transforming the research process from a generalized search to a precise data generation process oriented towards template display and component rendering. This improves the accuracy and consistency of subsequent interface population.

[0035] 2.1 Agent Reasoning Module This module is used to perform intelligent agent reasoning for research questions. Its core is a large language model trained through deep reinforcement learning. Upon receiving the research question input by the user, and information about the selected template, template structure, and attached components obtained from the template library, the model conducts research and analysis according to a cyclical mechanism of "thinking—execution—observation." Specifically, the agent first... <think>< / think> Identify the current inference content and generate a list based on the inference result. <search>< / search> The query text for the package is used to invoke a search tool to retrieve information. The results returned by the search tool are then processed through... <observation>< / observation> Feedback is sent to the agent, which then adjusts its research direction, supplements its analysis, and continues to advance the reasoning process. After multiple iterations, the agent finally... <result>< / result> Output research and analysis results that match the selected template structure, placeholder positions, and component filling requirements.

[0036] 2.2 Search Tool Module This module primarily encapsulates a unified query tool, providing external information retrieval capabilities for the agent reasoning module. When the agent generates... <search>< / search> After receiving the query text for the package, the retrieval tool invokes its built-in query function based on the query text and retrieves relevant search results from preset information sources. These sources include one or more of the following: the internet, local databases, or dedicated knowledge bases. The search results are then processed and fed back to the agent as the basis for its subsequent reasoning and result generation.

[0037] III. Intelligent Agent Component Filling Module This module automatically populates the structured survey data generated by the intelligent agent deep survey module into the mounted components of the selected template. During the population process, the system locates the target component based on the component ID in the template parsing results, and performs parameter matching and format conversion on the survey data in conjunction with the component attribute description information, so that the survey results are transformed into populated content that can be directly used for component rendering.

[0038] 3.1 Intelligent Agent Filling Module This module intelligently processes the data required for component population and calls the component population tool to write the processed data into the corresponding parts, thus achieving automatic component population. Its core is a large language model trained through deep reinforcement learning. After receiving survey analysis data output from the agent's deep survey module and the component information to be populated from the template parsing results, it filters the survey data based on the component ID, configurable parameters, and data type. During the population process, the agent... <think>< / think> Identify the current fill judgment process and generate a result from... <act>< / act> The package contains component call information, including the component ID to be populated and fill parameters that conform to the required format. Then, according to... <observation>< / observation> The feedback results determine whether to continue filling in the next component or correct any abnormal results. By repeatedly executing the above "think-execute-observe" process, the automatic filling of each mounted component in the template is completed.

[0039] 3.2 Fill Tool Module This module encapsulates a unified component population interface for the agent population module to call. The system obtains a list of components to be populated based on the template parsing results. This list includes component IDs, placeholder contexts, and receivable parameter information. The agent determines the target component ID based on the matching relationship between survey data and the list of components to be populated, and then... <act>< / act> Output the component ID and corresponding population parameters. The component population tool calls the corresponding component's population function based on the component ID, writes the population parameters into the component structure, generates the populated component code, and returns the population status. If population fails, it returns the reason for the failure so that the agent can correct and repopulate.

[0040] IV. Dynamic Interactive Interface Construction Module This module is used to write back the pre-filled component code generated by the intelligent agent component filling module to the corresponding positions according to the mapping relationship between placeholders and component IDs in the interface template, forming a dynamic interactive interface that can be displayed to the user. This method unifies the assembly of component resources, survey data, and template structure, realizing the automatic generation from research questions to interactive interfaces.

[0041] V. Dynamic Interactive Page Encapsulation Module This module is primarily used to encapsulate the generated interactive dynamic page structure, pre-populated component code, component interaction logic, and their dependent data into a savable HTML page resource package. Based on the mapping relationship between placeholders and component IDs output by the dynamic interactive interface construction module, the system extracts the rendering code, population parameters, data binding relationships, and interaction event information for each component, and organizes them uniformly.

[0042] 5.1 Page Dependency Resolution Module This module is used to parse the runtime dependencies of the generated page. Based on the component IDs in the template parsing results, the system retrieves the code resources, style resources, population parameters, and other data information of the corresponding components one by one, forming a page encapsulation dependency list. This dependency list includes not only the resources required for static page display but also the data fields and state parameters required for interactive behaviors such as filtering, switching, expanding, sorting, and chart linkage. This module clearly defines the smallest operational unit that each component needs to retain when exporting the HTML page, avoiding the loss of component interaction logic.

[0043] 5.2 Dynamic Data Snapshot Module This module is used to snapshot data from databases, API interfaces, or dedicated knowledge bases on a webpage. Based on the component ID, component attribute descriptions, and data binding relationships, the system extracts the data results required for the current page rendering and interaction, and converts them into a structured data snapshot that matches the component parameters. For database or API interface data, the system saves the data results obtained when the page was generated, data source identifiers, request parameters, and field mapping relationships. When the saved HTML page is reopened, it can prioritize reading the local data snapshot to complete component rendering and interactive display.

[0044] 5.3 Page Encapsulation and Reproduction Module This module unifies and encapsulates page structure, component code, interaction logic, style resources, and dynamic data snapshots, generating HTML page files or page resource packages that can be saved and reopened. During the encapsulation process, the system establishes a mapping between "component ID—data snapshot—interaction state," enabling the page to automatically restore component content, default states, and interaction rules based on the component ID upon loading. When a user opens the exported HTML page, the page first loads the encapsulated template structure and component resources, then reads the corresponding data snapshot based on the component ID, and re-injects the data into the component rendering logic. For interactive operations such as filtering, switching, sorting, expanding, and chart linkage, the page can continue to respond to user actions within the encapsulated data scope, thus achieving interactive page results.

[0045] Figure 1 A detailed flowchart of the intelligent agent deep thinking execution technology provided in the embodiments of the present invention.

[0046] Preliminary work steps Step 1: Style Data Preparation The system pre-constructs the style data required for generating dynamic interactive interfaces, including interface template resources, component resources, and component population functions. The interface template resources include template UI code, template description text, and placeholder information; the component resources include component IDs, core rendering functions, and component attribute descriptions. After the user selects a template, the system generates the template parsing result based on the placeholders, component IDs, and corresponding component attribute descriptions in the template.

[0047] Step 2: Training Data Preparation Training data was constructed separately for the agent inference module and the agent completion module. The training samples were all constrained by template parsing results: the former included the research question, template parsing results, and a complete thought execution trajectory, used to train the model to generate structured research data oriented towards component requirements; the latter included research analysis data, template parsing results, and a complete thought execution trajectory, used to train the model to complete component selection, parameter organization, completion execution, and anomaly correction. Seed samples for the training data were constructed by human experts, and a high-performance teacher model (using the Deepseek official API) was guided to generate candidate thought execution trajectories through prompt templates. During the generation process, pre-written retrieval and component completion tools were made available to the teacher model, ensuring that candidate trajectories included real tool call results. Subsequently, through rule-based filtering and manual review, high-quality trajectories with standardized label formats, valid tool calls, and output results matching the template parsing results were added to the training set.

[0048] Step 3: Model Training After the training data was prepared, the large language model (using Qwen3-A3B-30B-no-think) used by the agent reasoning module and the agent filling module was trained separately. Both models were trained using a combination of supervised fine-tuning and deep reinforcement learning. First, supervised fine-tuning was performed using a portion of the training data to allow the model to learn the standardized labeled thinking and execution trajectories. Then, the remaining training data was used to train the model using the GRPO deep reinforcement learning method to further optimize the model in terms of tool invocation, result matching, and anomaly correction. For the agent reasoning module, the reward rule mainly considers... <think>< / think> , <search>< / search> , <observation>< / observation> and <result>< / result> The evaluation criteria include whether the tags are used correctly, whether the search requests are valid, and whether the final survey results meet the component data requirements in the template parsing results. For the agent filling module, the reward rules, in addition to examining the tag format, also simultaneously determine... <act>< / act>The system checks whether the component ID comes from the template parsing result, whether the filling parameters match the component attribute description information, and whether the component filling tool returns a success status. If the filling fails, the corresponding reward is reduced; if the model can correct the parameters based on the reason for the failure and successfully refill, the corresponding reward is increased. The above work further enhances the agent's understanding and execution capabilities of the template parsing result, thereby enabling the survey, filling, and interface construction to operate collaboratively around the same constraint object.

[0049] System execution phase, such as Figure 2 As shown.

[0050] Step 1: User Input A single execution of the system begins with the user inputting a research question and selecting an interface template. After receiving the user's research question, the system reads the selected interface template and parses the template description text, placeholder context, component ID, and component attribute description information corresponding to the template from the style library module to generate the template parsing result.

[0051] Step Two: Deep Thinking and Execution by the Intelligent Agent The system inputs the user's research questions and template parsing results into the agent's reasoning module. Guided by the template description text, placeholder context, and component attribute constraints, the agent executes a cyclical reasoning process of "thinking—retrieval—observation" to generate research and analysis results that match the template display objectives and component filling requirements.

[0052] Figure 3 Detailed flowchart of intelligent agent deep thinking execution technology Step 2.1: Agent Thinking and Execution After receiving the user's research question, template parsing results, and inference and observation information from previous rounds, the agent outputs... <think>< / think> The current reasoning content of the package, and generate a function based on the reasoning result. <search>< / search> The query text for the package is used to invoke the search tool. When the agent determines that the survey information meets the requirements for template display and component population, the output is... <result>< / result> The final survey and analysis results of the package.

[0053] Step 2.2: Use the search tool to search. After receiving the thinking execution result output by the agent in step 2.1, if the system recognizes that it is from... <search>< / search> The query text for the package is extracted using rule parsing and input into the retrieval tool function. The retrieval tool determines the retrieval target and method based on the query text and obtains relevant information from preset sources. Internet sources can be retrieved using intelligent crawling, and SQL database sources can be retrieved using natural language to SQL query conversion. After the retrieval is completed, the system summarizes and organizes the results returned from different sources and presents them in a...<observation>< / observation> The information is fed back to the agent, enabling it to continue with step 2.1 based on the search results, and to revise the research direction or generate the final result.

[0054] Step 2.3: Final Result Output After the system receives the thinking and execution results of the intelligent agent, if it recognizes that it is caused by... <result>< / result> Once the package outputs its contents, the current research process is considered complete. The system extracts the final research analysis results through rule parsing and uses these results as the data source for subsequent component population stages, thus completing the "thinking-execution-observation" cycle in step two.

[0055] Step 3: Populate the intelligent agent components After obtaining the final survey and analysis results, the system retrieves information on all components to be populated in the user-selected interface template. This information includes the component ID, placeholder context, acceptable configuration parameters, and their data types. The system inputs the final survey and analysis results along with the component information into the agent-filling module, enabling the agent to filter the survey results based on component requirements and complete component filling through a "think-execute-observe" loop.

[0056] Figure 4 Detailed flowchart of intelligent agent component filling technology Step 3.1: The agent calls the tool to populate the component. After receiving the final survey and analysis results, the information on the components to be filled, and the execution feedback information from previous rounds, the agent performs a filling judgment and outputs the result. <think>< / think> The package contains thought processes and generates data from it. <act>< / act> The component population instruction includes the target component ID and the processed population parameters, which match the parameter names, data types, and format requirements that the target component can accept.

[0057] Step 3.2: Use the component population tool to populate the component. After receiving the "think-execute" result output by the agent, the system extracts it through rule parsing. <act>< / act>The system retrieves the component ID and fill parameters, and then calls the component fill tool to fill the corresponding component. The component fill tool locates the target component based on the component ID, calls the component's fill function, writes the fill parameters into the component structure, and generates the corresponding component code. If the component fill is successful, the system records the filled component code and removes the component from the list of components to be filled. If there are still unfilled components, the final survey analysis results, the updated list of components to be filled, and the filled status are input into the agent, returning to step 3.1 to execute the next round of filling. If all components have been filled, the system outputs all filled component code, ending step three. If component filling fails, the system records the reason for the failure and sends the failure information to the agent. <observation>< / observation> The form is fed back to the agent. The agent combines the reasons for failure, the final survey and analysis results, and the information of the components to be filled, corrects the component ID selection, parameter content, or parameter format, and re-executes step 3.1.

[0058] Step 4: Final Interactive Results Presentation After the system completes the population of all components, it replaces the populated component code in the corresponding positions of the template according to the correspondence between placeholders and component IDs in the selected template, generating complete interactive result display interface code. Subsequently, the system renders this interface code and displays the directly interactive result interface to the user. Through this step, the aforementioned survey results, component population results, and template structure are uniformly assembled, achieving automatic generation from research questions to a dynamic interactive interface.

[0059] Page encapsulation stage After generating the interactive results display interface code, the system further encapsulates the interface to generate a savable HTML page file or page resource package. The system then organizes the complete interface code, populated component code, component interaction logic, component population parameters, and data results on which the page display depends, generated in step four. This ensures that the exported page can still display the survey results at the time of generation when reopened, and retains dynamic interactive capabilities such as filtering, switching, expanding, sorting, and chart linkage within the scope of the saved data.

[0060] Figure 5 Detailed flowchart of page encapsulation technology Page Dependency Resolution The system parses the interactive result display interface code generated in step four. Based on the mapping relationship between placeholders and component IDs in the template, it determines the code resources, style resources, fill parameters, data binding fields, and interactive event information corresponding to each component on the page. For components with dynamic interactive behavior, the system further extracts their interactive state parameters and data access relationships. These interactive state parameters include default display items, filter conditions, tab states, pagination states, chart dimensions, etc. This step clarifies the component runtime units and data dependencies that need to be retained when exporting the HTML page, avoiding the loss of component interactivity due to only saving the static page structure.

[0061] Dynamic data snapshot generation The system snapshots data from survey results, databases, API interfaces, or dedicated knowledge bases on the page based on the page's dependency list. For data already generated by the agent deep survey module and agent component population module, the system stores it in a structured format according to component ID, component parameter name, and data type. For data from databases or API interfaces, the system saves the data results already obtained when the page was generated, and records the corresponding data source identifier, request parameters, returned fields, and field mapping relationships. Furthermore, the system does not package the original database service or remote API service itself into the HTML page; instead, it saves the current data results required for page rendering and interaction as a local data snapshot.

[0062] Interaction Logic and Data Adaptation The system establishes a mapping between "component ID - data snapshot - interaction state" based on the component ID and generates data adaptation logic for page runtime. This data adaptation logic is used to read the corresponding data snapshot based on the component ID when the exported page is opened, and convert it into a parameter format acceptable to the component rendering function. When a user triggers operations such as filtering, switching, expanding, sorting, or chart linkage in the exported HTML page, the page no longer relies on the original system's agent execution flow. Instead, it completes data reading, state updates, and component re-rendering locally using the encapsulated data snapshot and interaction state parameters.

[0063] HTML Page Generation and Export The system unifies and encapsulates the interactive result display interface code, component code, style resources, interaction scripts, dynamic data snapshots, and other logical information to generate an HTML page resource package. For pages with a small number of components and a small data scale, the system can write the page structure, scripts, styles, and data snapshots into a single HTML file. For pages containing a large amount of image, chart data, or other media resources, the system can generate a page resource package consisting of an HTML main file, resource files, and data files, and establish reference relationships through relative paths. After the page encapsulation is completed, the system performs integrity verification on the exported results, checking whether the component ID corresponds to the data snapshot, whether the component parameter format conforms to the component attribute description information, whether interactive events can be triggered normally, and whether page resources can be loaded correctly. If the verification passes, a saveable HTML page file or page resource package is output to the user; if the verification fails, the reason for the failure is recorded, and the system returns to the page dependency resolution or component population stage for correction.

[0064] In the above embodiments, the template description information is mainly used to help the agent understand the application scenario and required data type of the template. In other embodiments, the template may also contain richer metadata, such as template version information, template author information, template dependency component list, template style theme configuration, etc., to support more complex application scenarios.

[0065] In the above embodiments, the component population function is called synchronously. In other embodiments, for complex components that require a long rendering time, asynchronous calls can be used, allowing subsequent steps to continue through callbacks or event notifications after the component rendering is complete.

[0066] In the above embodiment, the reinforcement learning training uses the GRPO algorithm. In other embodiments, other reinforcement learning algorithms such as PPO (Proximal Policy Optimization) and DPO (Direct Preference Optimization) can also be used, as long as they can optimize the model output format and call correctness.

[0067] It should be understood that in the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples" refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this description, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate different embodiments or examples described in this specification, as well as some features of different embodiments or examples.

[0068] Of course, those skilled in the art can make various corresponding changes and modifications based on the present invention without departing from its spirit and essence, but such changes and modifications should all fall within the protection scope of the claims of the present invention.

Claims

1. A method for generating an intelligent agent interaction system based on in-depth research on dynamic interaction results, characterized in that, include: Step 1: Provide at least one interface template, each interface template containing layout structure information and at least one placeholder, each placeholder being associated with a component identifier; Step 2: Provide a component library containing multiple components corresponding to the component identifier. Each component has a corresponding population function, which is used to receive structured data and generate renderable component code. Step 3: The agent performs information retrieval and analysis based on the research questions input by the user, and generates structured analysis results; Step 4: The agent calls the filling function corresponding to the component to be filled according to the structured analysis results and the configuration parameter information of the component to be filled, processes the structured analysis results and fills them into the component to be filled, and generates the filled component code; Step 5: Replace the placeholders of the corresponding component identifiers in the interface template with the filled component code to generate a complete interactive interface.

2. The method for generating an intelligent agent interaction system based on deep research dynamic interaction results according to claim 1, characterized in that, Step 3 includes: The intelligent agent acquires the research question input by the user and the description information of the selected interface template; The agent executes at least one round of "think-execute-observe" cycle, wherein the agent generates search query text during the thinking phase, calls a search tool to obtain search results based on the search query text during the execution phase, and receives the search results during the observation phase as input for the next round of thinking; After completing the loop, the agent outputs the structured analysis results.

3. The method for generating an intelligent agent interaction system based on deep research dynamic interaction results according to claim 1, characterized in that, Step 4 includes: The agent acquires the structured analysis results and the configuration parameter information of all components to be populated. The configuration parameter information includes the component identifier of each component to be populated and the configuration parameters that the component can accept and their data types. The agent generates filling parameters that conform to the input format of the filling function for the current component to be filled, based on the structured analysis results and the configuration parameter information of the component. The agent calls the fill function, passes the fill parameters to the fill function, and generates the filled component code; The agent repeats the filling operation for the next component to be filled until all components to be filled are filled.

4. The method for generating an intelligent agent interaction system based on deep research dynamic interaction results according to claim 3, characterized in that, Step 4 also includes: When the agent fails to call the fill function, the agent receives a failure message; The agent adjusts the filling parameters based on the failure information and calls the filling function again to fill the data.

5. The method for generating an intelligent agent interaction system based on deep research dynamic interaction results according to claim 1, characterized in that, The description information of the interface template includes a description of the application scenario to which the interface template is intended and / or a description of the data type to be analyzed; the agent determines the retrieval strategy and analysis strategy based on the description information.

6. The method for generating an intelligent agent interaction system based on deep research dynamic interaction results according to claim 1, characterized in that, The intelligent agent includes a first intelligent agent and a second intelligent agent. Step 3 is executed by the first intelligent agent, and step 4 is executed by the second intelligent agent. The first intelligent agent and the second intelligent agent are each implemented independently based on a large language model.

7. The method for generating an intelligent agent interaction system based on deep research dynamic interaction results according to claim 6, characterized in that, The large language models of the first and second intelligent agents are trained through reinforcement learning. Among them, the reinforcement learning reward function of the large language model of the first intelligent agent is determined based on whether the output conforms to the preset label format; The reinforcement learning reward function of the second agent's large language model is determined based on whether the output conforms to the preset label format and whether the component filling function call is successful.

8. The method for generating an intelligent agent interaction system based on deep research dynamic interaction results according to claim 1, characterized in that, The retrieval tool encapsulates a unified query interface. The retrieval tool retrieves information from configured information sources based on the query text. The information sources include the Internet and / or local databases.

9. A device for generating intelligent agent interaction systems based on in-depth research on dynamic interaction results, characterized in that, include: A template library storage module is used to store at least one interface template, each interface template containing layout structure information and at least one placeholder, and each placeholder is associated with a component identifier; The component library storage module is used to store multiple components corresponding to the component identifier and the population function corresponding to each component. The population function is used to receive structured data and generate renderable component code. The intelligent agent reasoning module is used to perform information retrieval and analysis based on the research questions input by the user, and generate structured analysis results; The intelligent agent filling module is used to call the filling function corresponding to the component to be filled according to the structured analysis results and the configuration parameter information of the component to be filled, process the structured analysis results and fill them into the component to be filled, and generate the filled component code; The interface building module is used to replace the placeholders of the corresponding component identifiers in the interface template with the filled component code to generate a complete interactive interface.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method for generating an intelligent agent interaction system based on the deep research dynamic interaction results as described in any one of claims 1-8.

Images