A conversational tool for generating intelligence products and an interactive method

Abstract

The application discloses a dialogue type tool for generating intelligence products and an interactive method, and belongs to the technical field of large language model application and natural language processing. The application realizes the method as follows: 1, constructing an interactive framework for generating intelligence products; 2, integrating original materials, and initializing the logic state of a process navigator of an interactive presentation layer; 3, generating a structured outline in Markdown format; 4, constructing a chapter task list, generating chaptered intelligence content by batch asynchronous calling of an LLM model, and displaying the generation process in real time through a real-time editing feedback window of the interactive presentation layer; 5, automatically formatting and logically checking the generated intelligence content in sequence; 6, forming a complete intelligence product draft; 7, real-time editing in the whole process, re-generating and performing hot updating fine tuning on corresponding chapter content; compared with the prior art, the application solves the technical problem of realizing interactive and standardized intelligence product generation in intelligence compilation.

Description

Technical Field

[0001] This invention relates to a dialogic tool and interactive method for generating intelligence products, belonging to the field of large language model application and natural language processing technology, and is applicable to intelligence compilation and analysis and intelligence product generation. Background Technology

[0002] The compilation and analysis of intelligence products is a core part of intelligence work. Through the mining and analysis of multi-source heterogeneous data, structured reports with decision-support value are formed. The writing of traditional intelligence products includes discrete steps such as data collection, outline formulation, content filling, format adjustment, logical verification, and final review.

[0003] Current writing aids and large language model applications in the intelligence field suffer from four major limitations: First, fragmented processes and excessive reliance on manual intervention lead to long processing cycles, delayed feedback, and difficulty in integrating heterogeneous data from multiple sources, making it hard to meet real-time requirements. Second, inconsistent product quality and insufficient standardization result in results that depend on human experience, leading to contradictory conclusions; template tools cannot intelligently adjust structures or identify vulnerabilities. Third, limited human-computer interaction; general AI lacks professional understanding of the intelligence field and can only perform shallow generation, requiring users to master advanced prompt engineering skills; the process is uncontrollable and difficult to reproduce. Fourth, difficulties in content iteration and tracing; modifications require full-text manual review, structural adjustments and data supplementation are prone to errors, and there is a lack of intelligent guidance mechanisms throughout the entire process.

[0004] Therefore, how to achieve interactive and standardized intelligence product generation in intelligence integration is a technical problem that urgently needs to be solved. Summary of the Invention

[0005] The purpose of this invention is to achieve standardized interactive intelligence product generation during the intelligence compilation process, proposing a dialogical tool and interactive method for generating intelligence products. This invention pioneers a dialogical task decomposition method with a "navigable process," solidifying intelligence compilation into a five-step visualized process, thereby lowering the barrier to entry for large models. Simultaneously, it designs a "context-accumulated" generation mechanism, supporting local modifications and precise reconstruction at any stage. Furthermore, it integrates domain standards and multi-source citation technology, automatically inserting citation marks and generating references, ultimately achieving end-to-end automated conversion from Markdown to official government document format Word, completing the intelligent compilation from heterogeneous data to standardized intelligence products.

[0006] The objective of this invention is achieved through the following technical solution: This invention discloses a conversational interactive method for generating intelligence products, comprising the following steps: Step 1: Build an interactive framework for generating intelligence products based on the React framework, consisting of an interactive presentation layer, an intelligent processing layer, and a resource output layer; Step 1.1: Construct an interactive presentation layer based on the React framework, consisting of a workflow navigator, a conversational function panel, and a real-time editing feedback window, serving as the interactive entry point for users and tools. Furthermore, the workflow navigator generates and manages the logical states of the five steps: template selection, paragraph recognition, formatting review, logic verification, and product integration. The conversational function panel translates user task operation commands into LLM model call requests. The real-time editing feedback window integrates a Markdown editor, enabling real-time editing, previewing, and feedback of content. Step 1.2: Construct an intelligent processing layer for the core of the tool, consisting of a process-driven engine, a Prompt project library, and a structure parsing state manager. Further, the process-driven engine coordinates the execution order and data object transfer between steps; the Prompt project library encapsulates specialized structured prompt templates for the intelligence field; and the structure parsing state manager maintains and persistently stores intermediate results at each stage in real time, supporting process backtracking and contextual understanding. Step 1.3: Construct a resource output layer consisting of a multi-source data integrator and a standardized output generator for data integration and standardized product output; furthermore, the multi-source data integrator is used to access and parse multi-source heterogeneous intelligence data from upstream intelligence retrieval modules or third-party databases; the standardized output generator is used to convert the processed Markdown content into standardized documents that conform to specifications, realizing multi-format output of intelligence products and data archiving; Step 2: Integrate the multi-source heterogeneous intelligence data accessed through the multi-source data integrator of the resource output layer with the supplementary materials received by the user input through the real-time editing feedback window of the interactive presentation layer to form the original material, and initialize the logical state of the process navigator of the interactive presentation layer; Step 2.1: Call the multi-source data integrator of the resource output layer to access multi-source heterogeneous intelligence data from the upstream intelligence retrieval module or third-party database, parse and extract structured text intelligence and unstructured data tags, and store them in the searchResults state object; Step 2.2: Users manually enter supplementary factual materials through the real-time editing feedback window of the interactive presentation layer, which are synchronously updated to the inputText variable, forming complete original materials with the data loaded by the multi-source data integrator; Step 2.3: Call the process navigator of the interactive presentation layer, initialize the logical states of the five steps of template selection, paragraph recognition, format proofreading, logic verification, and product integration, set the global progress index to 0, establish a Boolean index for each step and mark it as incomplete. Step 2.4: Establish a data dependency verification mechanism to lock the activation permissions of subsequent steps. The state transition can only be triggered after the result of the previous step is generated and the verification is passed. Step 3: The interactive presentation layer receives the user's selected intelligence product template type through a dialog-style function panel. The intelligent processing layer matches and constrains the prompt word templates from the Prompt project library, and generates a Markdown-formatted structured outline that conforms to intelligence compilation standards based on the template type and the original materials. Step 3.1: The user triggers the template selection operation through the dialog-style function panel of the interactive presentation layer. The panel receives the template type selected by the user and converts the operation intention into a preset LLM model call request. The request contains the core parameters of template type identifier and original material association index. Step 3.2: Call the Prompt project library of the intelligent processing layer, match the corresponding structured prompt word template for the intelligence domain according to the template type identifier. This template encapsulates the core elements of outline chapter constraints and content professional specifications; at the same time, use the complete original material in the inputText variable as a semantic seed and concatenate it into the prompt word template to generate standardized LLM calling instructions. Step 3.3: Send the encapsulated standardized LLM calling instructions to the LLM model. The model generates a Markdown-formatted outline that conforms to the intelligence compilation specifications based on the original materials and prompt word constraints. Step 3.4: Call the process navigator of the interactive presentation layer, receive the non-empty verification signal of the templateResult field returned by the structure parsing state manager, update the boolean index of the template selection step to the completed state; at the same time, increment the global progress index to 1, remove the disabled attribute of the interactive entry of the paragraph recognition step, unlock the activation permission of this step, and complete the process loop of the template selection stage. Step 4: The intelligent processing layer's workflow-driven engine uses regular expressions and AST syntax trees to parse the Markdown-formatted structured outline, constructs a list of chapter tasks, generates chapter-specific information content through batch asynchronous calls to the LLM model, and displays the generation process in real time through the interactive presentation layer's real-time editing feedback window; Step 4.1: Receive paragraph recognition execution instructions through the interactive presentation layer's dialog-style function panel, call the intelligent processing layer's process-driven engine to start the handleParagraphRecognition processing program, and retrieve the templateResult outline data from resultMarkdownJson in the structure parsing state manager; use regular expressions and AST syntax trees to parse the Markdown heading levels, filter redundant characters, and group the tasks by using the first-level chapter title as the task index and the corresponding second- and third-level subheadings as logical sub-items to construct a chapter task list; Step 4.2: Call the process-driven engine to traverse the chapter task list, build a dedicated Prompt for each chapter and inject structural context, clarifying the chapter positioning and subheading sequence; perform semantic retrieval through the multi-source data integrator of the resource output layer, embed the matching fact evidence blocks in searchResults into the Prompt, and call the Prompt project library to add writing style constraints such as argument requirements, data citations, and paragraph connections; Step 4.3: The process-driven engine performs batch asynchronous calls to the chapter task queue and sends them to the LLM model, and starts the Stream response listener to receive paragraph text; the real-time editing feedback window dynamically displays the generated content, and the structure parsing state manager fills the paragraph content into the paragraphRecognitionResult state object according to the chapter index; after all chapter content is merged, the currentStep progress index and stepCompleted completion status of the process navigator are updated. Step 5: The intelligent processing layer performs automated format verification and logical checks on the generated intelligence content in sequence; Step 5.1: Receive format review instructions through the interactive presentation layer's dialog-style function panel, call the intelligent processing layer's structure parsing state manager to read the full text of the paragraphRecognitionResult; call the Prompt project library to obtain the intelligence document verification template, and have the process-driven engine encapsulate the text and verification instructions and send them to the LLM model; receive the model's correction data, automatically replace typos, punctuation, and full / half-width errors, unify military and technical terminology, and finally write the optimized text into the formatCheckResult state object; Step 5.2: Trigger the logic verification command through the dialog-based function panel. The structure parsing status manager retrieves the formatCheckResult data as input. Call the logic verification template of the Prompt project library. The process-driven engine drives the LLM model to scan the causal chain and factual evidence of the chapters, locate and correct contradictions in the data conclusions, automatically supplement transition sentences and unify the narrative tone. Save the logic-optimized text to the logicCheckResult status field and increment the currentStep progress index. Step 5.3: Call the structure parsing state manager to persistently store the intermediate versions of format review and logic verification in fragments using resultMarkdownJson; the real-time editing feedback window provides a version snapshot index, supporting non-linear jumps and rollback operations for users; extract the corresponding version text based on the rollback index, overwrite the current active editing state, and realize process backtracking; Step 6: The intelligent processing layer integrates the reviewed intelligence content, automatically establishes a mapping relationship between factual evidence and the main text, and generates a reference section to form a complete draft of the intelligence product; the resource output layer converts the draft of the intelligence product into a standardized editable document and outputs it. Step 6.1: Trigger the product integration command through the dialog-based function panel, call the process-driven engine to start the handleProductIntegration process, recursively scan the entire document and extract the adopted reference evidence blocks from searchResults through the multi-source data integrator; establish the mapping index between factual evidence and the main text, classify the citation sources, and automatically generate the reference chapters according to the specifications by the structure parsing state manager; Step 6.2: Call the global consistency enhancement template of the Prompt project library. The process-driven engine drives the LLM model to refine the logicCheckResult text while retaining the core arguments; automatically insert citation marks in the main text according to the index, merge the optimized text with the references, and generate a complete Markdown draft. Step 6.3: Call the standardized output generator to start the handleExportDocx conversion engine, parse the Markdown draft into an AST abstract syntax tree; map semantic tags to Word document styles according to the selected template, and fill in metadata such as title, security level, and date through the multi-source data integrator to generate an editable .docx document; Step 6.4: The standardized output generator stores the Word document and returns a download link; the process navigator marks the currentStep as complete; the structure parsing status manager packages the initial outline, retrieved evidence, intermediate text, and proofreading notes, and stores them in the historical task library to complete the entire process archiving. Step 7: Use the real-time editing feedback window of the interactive presentation layer to perform global monitoring to support real-time editing throughout the entire process, and trigger the dialog-style function panel to receive the user's partial modification instructions, regenerate the corresponding chapter content and perform hot updates and fine-tuning. Step 7.1: The real-time editing feedback window globally listens for modification instructions in steps 2-6, and invokes the editing dialog box; it calls the structure parsing state manager to extract the current step context from resultMarkdownJson and displays it, and receives the user's natural language modification requirements through the userRequirements interface of the dialogic function panel; Step 7.2: The dialog-based function panel triggers the regeneration command, and the process-driven engine starts the handleRegenerate processing function; the structure parsing state manager locates the current stage through the currentStep index, combines the original constraints of the Prompt project library, the current paragraph data and the user's modification requirements, merges and generates a local regeneration command and sends it to the LLM model to receive the newly generated content; Step 7.3: The structure parsing state manager uses hot updates to overwrite the corresponding field indexes of resultMarkdownJson with new content; it automatically verifies the validity of subsequent steps and locks or unlocks corresponding activation permissions; the process navigator synchronously updates the step status, supporting unlimited local fine-tuning without disrupting the overall architecture.

[0007] The present invention discloses a conversational tool for generating intelligence products, used to implement the above method, comprising an interactive presentation layer consisting of a process navigator component, a conversational function panel component, and a real-time editing feedback window; an intelligent processing layer consisting of a process-driven engine component, a Prompt project library component, and a structure parsing state manager; and a resource output layer consisting of a multi-source data integrator and a standardized output generator. The process navigator component is used to generate and manage the logical states of five major steps: template selection, paragraph recognition, format review, logic verification, and product integration, and respond to user click events to trigger step jumps. The conversational function panel component is used to convert the user's task operation instructions into LLM model call requests, and provides interactive entry points for type selection, step execution, and regeneration. The real-time editing feedback window component is used to enable real-time editing, previewing and feedback of the content generated in each step, and supports users to manually modify and regenerate. The process-driven engine component is used to coordinate the execution order and data object transfer between each step, and to execute core business logic processing. The Prompt project library component is used to encapsulate specialized structured prompt word templates for the intelligence field, ensuring that the output content of the LLM model conforms to the expected format; The structure parsing state manager component is used to maintain and store the intermediate results of each stage in real time, supporting process backtracking and context handling. The multi-source data integrator component is used to access and parse multi-source heterogeneous intelligence data from the upstream intelligence retrieval module or a third-party database, providing raw materials for intelligence generation. The standardized output generator component is used to convert the processed Markdown content into standardized documents that conform to the specifications, so as to realize multi-format output and data archiving of intelligence products; Compared with existing technologies, it has the following beneficial effects: 1. End-to-end intelligent automation significantly improves production efficiency. By integrating all aspects of intelligence gathering and processing through visual process navigation, and relying on the LLM model, days of manual work can be compressed to minutes, solving the problems of long intelligence processing cycles and delayed feedback.

[0008] 2. RAG multi-source data fusion ensures the authenticity and objectivity of the content. It supports semantic processing of text and multimodal data such as remote sensing and radar, extracts high-value evidence by combining re-ranking algorithms, strictly constrains the generation process, suppresses model illusions, and ensures that conclusions are well-founded.

[0009] 3. Iterative conversational interaction lowers the barrier to entry. Abandoning single-command interaction, users do not require advanced Prompt engineering skills; they can dynamically intervene and regenerate parts of the process through natural language, achieving intent-centric human-machine collaboration.

[0010] 4. Automated proofreading and formatting ensure product standardization. Pre-set proofreading rules automatically identify logical inconsistencies, unify professional terminology, and export Word documents according to official document specifications, eliminating quality inconsistencies caused by differences in personnel experience and ensuring that reports are rigorous and consistent.

[0011] 5. Full-chain traceability supports knowledge accumulation. Automatic generation of citation tags and references preserves the complete derivation chain from data to conclusions; end-to-end data archiving provides high-quality corpus for intelligence tracing and model optimization. Attached Figure Description

[0012] Figure 1 is a system architecture block diagram of the intelligence compilation tool described in this invention; Figure 2 This is a schematic diagram of the main interactive interface of the editing tool described in this invention; Figure 3 This is the core business process diagram of the automated intelligence compilation described in this invention; Figure 4 This is a diagram illustrating the interface for modifying content and regenerating the dialog box; Figure 5 This is a schematic diagram illustrating the principle of format conversion from Markdown to standardized Word as described in this invention. Detailed Implementation

[0013] To better illustrate the purpose and advantages of this invention, the invention will be further described below with reference to the accompanying drawings and examples. It should be noted that the implementation of this invention is not limited to the following embodiments, and any modifications or alterations made to this invention will fall within the scope of protection of this invention.

[0014] Example like Figure 1As shown in the figure, the specific implementation steps of the conversational interactive method for generating information products in this embodiment are as follows: Step 1: Build an interactive framework for generating intelligence products based on the React framework, consisting of an interactive presentation layer, an intelligent processing layer, and a resource output layer; Step 1.1: Construct an interactive presentation layer based on the React framework, consisting of a workflow navigator, a conversational function panel, and a real-time editing feedback window, serving as the interactive entry point for users and tools. Furthermore, the workflow navigator generates and manages the logical states of the five steps: template selection, paragraph recognition, formatting review, logic verification, and product integration. The conversational function panel translates user task operation commands into LLM model call requests. The real-time editing feedback window integrates a Markdown editor, enabling real-time editing, previewing, and feedback of content. Step 1.2: Construct an intelligent processing layer for the core of the tool, consisting of a process-driven engine, a Prompt project library, and a structure parsing state manager. Further, the process-driven engine coordinates the execution order and data object transfer between steps; the Prompt project library encapsulates specialized structured prompt templates for the intelligence field; and the structure parsing state manager maintains and persistently stores intermediate results at each stage in real time, supporting process backtracking and contextual understanding. Step 1.3: Construct a resource output layer consisting of a multi-source data integrator and a standardized output generator for data integration and standardized product output; furthermore, the multi-source data integrator is used to access and parse multi-source heterogeneous intelligence data from upstream intelligence retrieval modules or third-party databases; the standardized output generator is used to convert the processed Markdown content into standardized documents that conform to specifications, realizing multi-format output of intelligence products and data archiving; Step 2: Integrate the multi-source heterogeneous intelligence data accessed through the multi-source data integrator of the resource output layer with the supplementary materials received by the user input through the real-time editing feedback window of the interactive presentation layer to form the original material, and initialize the logical state of the process navigator of the interactive presentation layer; Step 2.1: Call the multi-source data integrator of the resource output layer to access multi-source heterogeneous intelligence data from the upstream intelligence retrieval module or third-party database, parse and extract structured text intelligence and unstructured data tags, and store them in the searchResults state object; Step 2.2: Users manually enter supplementary factual materials through the real-time editing feedback window of the interactive presentation layer, which are synchronously updated to the inputText variable, forming complete original materials with the data loaded by the multi-source data integrator; Step 2.3: Call the process navigator of the interactive presentation layer, initialize the logical states of the five steps of template selection, paragraph recognition, format proofreading, logic verification, and product integration, set the global progress index to 0, establish a Boolean index for each step and mark it as incomplete. Step 2.4: Establish a data dependency verification mechanism to lock the activation permissions of subsequent steps. The state transition can only be triggered after the result of the previous step is generated and the verification is passed. Step 3: The interactive presentation layer receives the user's selected intelligence product template type through a dialog-style function panel. The intelligent processing layer matches and constrains the prompt word templates from the Prompt project library, and generates a Markdown-formatted structured outline that conforms to intelligence compilation standards based on the template type and the original materials. Step 3.1: The user triggers the template selection operation through the dialog-style function panel of the interactive presentation layer. The panel receives the template type selected by the user and converts the operation intention into a preset LLM model call request. The request contains the core parameters of template type identifier and original material association index. Step 3.2: Call the Prompt project library of the intelligent processing layer, match the corresponding structured prompt word template for the intelligence domain according to the template type identifier. This template encapsulates the core elements of outline chapter constraints and content professional specifications; at the same time, use the complete original material in the inputText variable as a semantic seed and concatenate it into the prompt word template to generate standardized LLM calling instructions. Step 3.3: Send the encapsulated standardized LLM calling instructions to the LLM model. The model generates a Markdown-formatted outline that conforms to the intelligence compilation specifications based on the original materials and prompt word constraints. Step 3.4: Call the process navigator of the interactive presentation layer, receive the non-empty verification signal of the templateResult field returned by the structure parsing state manager, update the boolean index of the template selection step to the completed state; at the same time, increment the global progress index to 1, remove the disabled attribute of the interactive entry of the paragraph recognition step, unlock the activation permission of this step, and complete the process loop of the template selection stage. Step 4: The intelligent processing layer's workflow-driven engine uses regular expressions and AST syntax trees to parse the Markdown-formatted structured outline, constructs a list of chapter tasks, generates chapter-specific information content through batch asynchronous calls to the LLM model, and displays the generation process in real time through the interactive presentation layer's real-time editing feedback window; Step 4.1: Receive paragraph recognition execution instructions through the interactive presentation layer's dialog-style function panel, call the intelligent processing layer's process-driven engine to start the handleParagraphRecognition processing program, and retrieve the templateResult outline data from resultMarkdownJson in the structure parsing state manager; use regular expressions and AST syntax trees to parse the Markdown heading levels, filter redundant characters, and group the tasks by using the first-level chapter title as the task index and the corresponding second- and third-level subheadings as logical sub-items to construct a chapter task list; Step 4.2: Call the process-driven engine to traverse the chapter task list, build a dedicated Prompt for each chapter and inject structural context, clarifying the chapter positioning and subheading sequence; perform semantic retrieval through the multi-source data integrator of the resource output layer, embed the matching fact evidence blocks in searchResults into the Prompt, and call the Prompt project library to add writing style constraints such as argument requirements, data citations, and paragraph connections; Step 4.3: The process-driven engine performs batch asynchronous calls to the chapter task queue and sends them to the LLM model, and starts the Stream response listener to receive paragraph text; the real-time editing feedback window dynamically displays the generated content, and the structure parsing state manager fills the paragraph content into the paragraphRecognitionResult state object according to the chapter index; after all chapter content is merged, the currentStep progress index and stepCompleted completion status of the process navigator are updated. Step 5: The intelligent processing layer performs automated format verification and logical checks on the generated intelligence content in sequence; Step 5.1: Receive format review instructions through the interactive presentation layer's dialog-style function panel, call the intelligent processing layer's structure parsing state manager to read the full text of the paragraphRecognitionResult; call the Prompt project library to obtain the intelligence document verification template, and have the process-driven engine encapsulate the text and verification instructions and send them to the LLM model; receive the model's correction data, automatically replace typos, punctuation, and full / half-width errors, unify military and technical terminology, and finally write the optimized text into the formatCheckResult state object; Step 5.2: Trigger the logic verification command through the dialog-based function panel. The structure parsing status manager retrieves the formatCheckResult data as input. Call the logic verification template of the Prompt project library. The process-driven engine drives the LLM model to scan the causal chain and factual evidence of the chapters, locate and correct contradictions in the data conclusions, automatically supplement transition sentences and unify the narrative tone. Save the logic-optimized text to the logicCheckResult status field and increment the currentStep progress index. Step 5.3: Call the structure parsing state manager to persistently store the intermediate versions of format review and logic verification in fragments using resultMarkdownJson; the real-time editing feedback window provides a version snapshot index, supporting non-linear jumps and rollback operations for users; extract the corresponding version text based on the rollback index, overwrite the current active editing state, and realize process backtracking; Step 6: The intelligent processing layer integrates the reviewed intelligence content, automatically establishes a mapping relationship between factual evidence and the main text, and generates a reference section to form a complete draft of the intelligence product; the resource output layer converts the draft of the intelligence product into a standardized editable document and outputs it. Step 6.1: Trigger the product integration command through the dialog-based function panel, call the process-driven engine to start the handleProductIntegration process, recursively scan the entire document and extract the adopted reference evidence blocks from searchResults through the multi-source data integrator; establish the mapping index between factual evidence and the main text, classify the citation sources, and automatically generate the reference chapters according to the specifications by the structure parsing state manager; Step 6.2: Call the global consistency enhancement template of the Prompt project library. The process-driven engine drives the LLM model to refine the logicCheckResult text while retaining the core arguments; automatically insert citation marks in the main text according to the index, merge the optimized text with the references, and generate a complete Markdown draft. Step 6.3: Call the standardized output generator to start the handleExportDocx conversion engine, parse the Markdown draft into an AST abstract syntax tree; map semantic tags to Word document styles according to the selected template, and fill in metadata such as title, security level, and date through the multi-source data integrator to generate an editable .docx document; Step 6.4: The standardized output generator stores the Word document and returns a download link; the process navigator marks the currentStep as complete; the structure parsing status manager packages the initial outline, retrieved evidence, intermediate text, and proofreading notes, and stores them in the historical task library to complete the entire process archiving. Step 7: Use the real-time editing feedback window of the interactive presentation layer to perform global monitoring to support real-time editing throughout the entire process, and trigger the dialog-style function panel to receive the user's partial modification instructions, regenerate the corresponding chapter content and perform hot updates and fine-tuning. Step 7.1: The real-time editing feedback window globally listens for modification instructions in steps 2-6, and invokes the editing dialog box; it calls the structure parsing state manager to extract the current step context from resultMarkdownJson and displays it, and receives the user's natural language modification requirements through the userRequirements interface of the dialogic function panel; Step 7.2: The dialog-based function panel triggers the regeneration command, and the process-driven engine starts the handleRegenerate processing function; the structure parsing state manager locates the current stage through the currentStep index, combines the original constraints of the Prompt project library, the current paragraph data and the user's modification requirements, merges and generates a local regeneration command and sends it to the LLM model to receive the newly generated content; Step 7.3: The structure parsing state manager uses hot updates to overwrite the corresponding field indexes of resultMarkdownJson with new content; it automatically verifies the validity of subsequent steps and locks or unlocks corresponding activation permissions; the process navigator synchronously updates the step status, supporting unlimited local fine-tuning without disrupting the overall architecture.

[0015] This embodiment provides a dialogic tool for generating intelligence products, used to implement the above method. It includes an interactive presentation layer consisting of a process navigator component, a dialogic function panel component, and a real-time editing feedback window; an intelligent processing layer consisting of a process-driven engine component, a Prompt project library component, and a structure parsing state manager; and a resource output layer consisting of a multi-source data integrator and a standardized output generator. The process navigator component is used to generate and manage the logical states of five major steps: template selection, paragraph recognition, format review, logic verification, and product integration, and respond to user click events to trigger step jumps. In this embodiment, user clicks are taken as input, and step status update signals are taken as output. The Steps component, based on the React framework, marks the progress of each step using a `stepCompleted` state array. Each step corresponds to a Boolean index, and a global progress index is maintained through the `currentStep` variable. When a step is completed, the index value is automatically incremented, unlocking the interaction entry for the next step. Users can click on completed steps to review them. The step status rendering function renders the visual status of each step as completed or incomplete based on the `stepCompleted` array. The step jump response function responds to user click events and verifies the reachability of the target step, specifying that jumps are only allowed to completed steps or the next unexecuted step. The step navigation button control function renders the previous, next, and completed analysis buttons, dynamically controlling the button display text and disabled state based on the `currentStep` index.

[0016] The conversational function panel component is used to convert the user's task operation instructions into LLM model call requests, and provides interactive entry points for type selection, step execution, and regeneration. In the embodiments, such as Figure 2As shown, the user click event is taken as input, and the constructed Prompt request is taken as output. The type selection button group supports selecting either the domain situation report or the international hot topic analysis template. The step execution button group provides five buttons: execute template selection, execute paragraph recognition, execute format review, execute logic verification, and execute product integration. The regenerated button responds to the user's modification request. The handleTemplateSelection function responds to the template selection button click, constructs different Prompt requirements based on the user's selected type, and outputs a Markdown structure template containing specific chapters. The domain situation report template includes chapters such as introduction, situation overview, key direction analysis, and development trend prediction. The international hot topic analysis template includes chapters such as event overview, background analysis, impact assessment, and response suggestions. The handleParagraphRecognition function responds to the paragraph recognition button click, first calling the parseMarkdownHeadings function to parse the heading structure in the template, dividing the content into chapters according to the first-level headings, and then traversing each chapter to construct a Prompt. This allows the LLM to generate detailed content for all subheadings of that chapter based on the reference text, with a word limit of 700-1000. 0 words, and finally the content generated by all chapters is spliced ​​and saved; the handleFormatReview function responds to the format review button click, sends the paragraph recognition result to LLM, asks it to check the format specifications such as heading level and list nesting, and returns the corrected complete Markdown content; the handleLogicCheck function responds to the logic check button click, sends the format-reviewed content to LLM, asks it to adjust the structure to make its logic smoother, evaluates the coherence, fluency and logic of the content and optimizes it; the handleProductIntegration function responds to the product integration button click, asks LLM to further optimize the expression based on the results of the logic check, and specifically instructs LLM to add reference marks to the searchResults search results in appropriate positions in the text, such as the ^[1]^ format, and automatically generates a reference list at the end of the document to list the source information; the handleRegenerate function responds to the regenerate button click, combines the current editor content and the new requirements userRequirements input by the user, constructs a Prompt to call LLM to regenerate the current step content, and updates the corresponding state variables.

[0017] The real-time editing feedback window component is used to enable real-time editing, previewing and feedback of the content generated in each step, and supports users to manually modify and regenerate. In this embodiment, Markdown formatted text is used as input and rendered HTML content is used as output; the ReactMarkdown component is integrated to render Markdown text, and the #, ##, and ### title identifiers are parsed into corresponding level HTML title elements, paragraph text is parsed into p tags, and the quotation mark ^[1]^ is parsed into superscript elements; the Modal modal box is integrated to realize full-screen editing function, and the content is edited synchronously through the editorContent state variable. After the user modifies, the content is updated to the structure parsing state manager component in real time; the renderStepContent function dynamically renders different operation buttons and result display areas according to the current step index. If loading is in progress, the Spin loading animation is displayed, and if there is no result, the Empty state is displayed. The code demonstrates how, if results are available, the ReactMarkdown component is used to render the Markdown text into visually appealing HTML, with special parsing handling for JSON results; the content rendering function calls ReactMarkdown to convert the Markdown text into HTML and display it; the editing synchronization function listens for changes in the editor content and updates the editorContent state variable in real time; the modal control function responds to full-screen edit button click events, opening or closing the modal and loading the complete content of the current step for the user to edit; the input area rendering function displays the automatically loaded search data, showing searchResults in a table format with expandable details, and if no data is automatically loaded, a text box is provided for the user to manually enter inputText.

[0018] The process-driven engine component is used to coordinate the execution order and data object transfer between each step, and to execute core business logic processing. In this embodiment, user operation instructions and the output of the previous step are used as input, and the output of the current step is used as output. The `callLLM` function encapsulates the general logic for calling the LLM API, receives a prompt word, constructs a request body conforming to the API standard including system and user role messages, sends a POST request to the configured API address, sets a timeout of 60 seconds, and if the API call fails, it captures the error and calls the `getMockResult` function to return the simulation result to ensure the process is not interrupted. The `getLLMConfig` function asynchronously obtains the configuration information of the LLM model, including the API URL, Key, default model, etc. It first checks the local cache; if it is not found, it sends a request to the backend interface ` / v1 / system / llm_config`. If the configuration is obtained successfully, it throws an error and prompts the user if the cache configuration fails. The `parseMarkdownHeadings` function parses the heading structure in the Markdown text, traverses the text lines, uses regular expressions to match lines starting with #, ##, and ###, extracts the heading level and text content, and returns a structured array for the `handleParagraphRecognition` function to generate content by chapter. The `getMockResult` function in LLM... When an API call fails, a simulated result is returned, containing a preset template structure or paragraph content, ensuring that the five-step process can continue to execute even in the event of an API exception, thus avoiding process interruption. The five core business processing functions are handleTemplateSelection, handleParagraphRecognition, handleFormatReview, handleLogicCheck, and handleProductIntegration. Each function is executed in sequence and its output is passed to the next stage. The execution flow is controlled by the currentStep variable.

[0019] The Prompt project library component is used to encapsulate specialized structured prompt word templates for the intelligence field, ensuring that the output content of the LLM model conforms to the expected format; In the embodiment, the template type identifier Template ID and step index are used as inputs, and the completed Prompt is output. Each step prompt template has built-in system roles, chapter structure, word limit, format specifications, citation rules and other constraints. Among them, the template selection template requires LLM to generate a Markdown structure with an introduction and situation overview; the paragraph recognition template requires LLM to generate 700-1000 words of content as subheadings according to the reference text; the format review template checks the title level, list nesting and outputs the corrected Markdown; the logic check template optimizes the content coherence and logic; the product integration template automatically adds ^[1]^ citation marks and generates references; the regenerated template combines userRequirements and the current content to construct a local reconstruction instruction. All templates dynamically construct System and User role messages according to the intelligence type to ensure that the output conforms to the domain specifications.

[0020] The structure parsing state manager component is used to maintain and store the intermediate results of each stage in real time, supporting process backtracking and context handling. In the embodiments, such as Figure 3As shown, the text content generated in each step is used as input, and the persistently stored state object is used as output; it is implemented based on React's useState hook; the templateResult state variable stores the initial template content generated in the template selection step; the paragraphRecognitionResult state variable stores the chapter filling content generated in the paragraph recognition step; the formatReviewResult state variable stores the Markdown content corrected in the format review step; the logicCheckResult state variable stores the optimized content in the logic check step; the productIntegrationResult state variable stores the final report content generated in the product integration step; and the resultMarkdownJson aggregate object is managed uniformly. The above five state variables enable state synchronization: when any state is updated, it is automatically synchronized to the real-time editing feedback window component; the inputText state variable stores the original text material input by the user; the searchResults state variable stores the search result data passed from the upstream search module; the editorContent state variable stores the content in the current editor; the currentStep state variable maintains the global progress index; the stepCompleted state array marks the completion progress of each step; the loading state variable controls the display and hiding of the loading animation; the userRequirements state variable stores the user-inputted regeneration requirements; the llmConfig state variable stores the configuration information of the LLM model; and the importData state variable stores the reconstructed data imported from external sources.

[0021] The multi-source data integrator component is used to access and parse multi-source heterogeneous intelligence data from the upstream intelligence retrieval module or a third-party database, providing raw materials for intelligence generation. In the embodiments, such as Figure 4As shown, the uploaded file or transmitted data is used as input, and the parsed structured data is used as output. The parseExcelData function parses the uploaded Excel file, mapping the worksheet data back to the component state, including the inputText, templateResult, paragraphRecognitionResult, and other results from each step, supporting data import and secondary editing. The customRequest function works with Ant Design's Upload component to handle file reading and parsing logic, listening for file selection events to call FileReader to read file content, calling the corresponding parsing function according to the file type, updating the component state after parsing, and closing the loading prompt. The searchResults data access interface receives the search result data transmitted from the upstream retrieval module, listens for the route state location.state to automatically load data into the searchResults state variable. After uniform formatting, each data source can be used as a reference for intelligence generation and is called by the process-driven engine component in the paragraph recognition and product integration steps to obtain the original factual evidence block.

[0022] The standardized output generator component is used to convert the processed Markdown content into standardized documents that conform to the specifications, so as to realize multi-format output and data archiving of intelligence products; In the embodiments, such as Figure 5 As shown, the code takes Markdown text as input and outputs Word and Excel format files. The `handleExportDocx` function generates a Word-compatible HTML file, builds a dedicated XML namespace and CSS styles, configures SimSun font and A4 page margins to ensure format compliance, and calls the `processMarkdownToHtml` function to complete the format conversion. This function converts Markdown headings to h1-h3 tags, paragraphs to p tags, and quotation marks to superscript sup tags, supporting Word export. The `handleExport` function calls the xlsx library to package the entire process's structured results into a multi-worksheet Excel file, including input text, templates, and processing results at each stage, supporting data backtracking and secondary editing. The download function triggers browser download by hiding the `a` tag and automatically cleans up temporary objects.

[0023] A unidirectional data flow is established between the components. The process navigator receives user operations and triggers function calls to the conversational function panel; the conversational function panel calls the process-driven engine to execute business logic; the process-driven engine retrieves the Prompt project library template and stores the results through the structure parsing state manager; the structure parsing state manager synchronizes the status to the real-time editing feedback window for display; the multi-source data integrator provides raw data to the process-driven engine; and the standardized output generator calls the structure parsing state manager to obtain the final result and complete the document transformation.

[0024] Furthermore, the tool also includes a component lifecycle management module, implemented based on React's useEffect hook; an initialization configuration loading function that calls getLLMConfig to obtain and cache the LLM configuration when the component is mounted; a route data listening function that listens to location.state to automatically load searchResults and inputText from the upstream retrieval module; a state synchronization function that automatically synchronizes currentResult and editorContent when the result of a certain step is updated or when a step is switched to ensure that the view is updated in real time; and an external data import listening function that listens for reconstructed data passed from other pages and populates the corresponding state if so.

[0025] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A conversational interactive method for generating intelligence products, characterized in that: Includes the following steps, Step 1: Build an interactive framework for generating intelligence products based on the React framework, consisting of an interactive presentation layer, an intelligent processing layer, and a resource output layer; Step 2: Integrate the multi-source heterogeneous intelligence data accessed through the multi-source data integrator of the resource output layer with the supplementary materials received by the user input through the real-time editing feedback window of the interactive presentation layer to form the original material, and initialize the logical state of the process navigator of the interactive presentation layer; Step 3: The interactive presentation layer receives the user's selected intelligence product template type through a dialog-style function panel. The intelligent processing layer matches and constrains the prompt word templates from the Prompt project library, and generates a Markdown-formatted structured outline that conforms to intelligence compilation standards based on the template type and the original materials. Step 4: The intelligent processing layer's process-driven engine uses regular expressions and AST syntax trees to parse the Markdown-formatted structured outline, constructs a list of chapter tasks, generates chapter-specific information content through batch asynchronous calls to the LLM model, and displays the generation process in real time through the interactive presentation layer's real-time editing feedback window; Step 5: The intelligent processing layer performs automated format verification and logical checks on the generated intelligence content in sequence; Step 6: The intelligent processing layer integrates the reviewed intelligence content, automatically establishes a mapping relationship between factual evidence and the main text, and generates a reference section to form a complete draft of the intelligence product; the resource output layer converts the draft of the intelligence product into a standardized editable document and outputs it. Step 7: Utilize the real-time editing feedback window of the interactive presentation layer to perform global monitoring, enabling support for real-time editing throughout the entire process, and triggering a dialog-style function panel to receive the user's partial modification instructions, regenerating the corresponding chapter content and performing hot updates and fine-tuning.

2. The interactive method for generating information products in a conversational manner as described in claim 1, characterized in that: Step 1 is implemented as follows: Step 1.1: Construct an interactive presentation layer based on the React framework, consisting of a process navigator, a conversational function panel, and a real-time editing feedback window, as the interactive operation entry point for users and tools; furthermore, the process navigator is used to generate and manage the logical states of the five major steps: template selection, paragraph recognition, format review, logic verification, and product integration; the conversational function panel is used to convert the user's task operation instructions into LLM model call requests; The real-time editing feedback window integrates a Markdown editor, enabling real-time editing, previewing, and feedback of content; Step 1.2: Construct an intelligent processing layer for the core of the tool, consisting of a process-driven engine, a Prompt project library, and a structure parsing state manager. Further, the process-driven engine coordinates the execution order and data object transfer between steps; the Prompt project library encapsulates specialized structured prompt templates for the intelligence field; and the structure parsing state manager maintains and persistently stores intermediate results at each stage in real time, supporting process backtracking and contextual understanding. Step 1.3: Construct a resource output layer consisting of a multi-source data integrator and a standardized output generator for data integration and standardized product output; furthermore, the multi-source data integrator is used to access and parse multi-source heterogeneous intelligence data from upstream intelligence retrieval modules or third-party databases; The standardized output generator is used to convert processed Markdown content into standardized documents that conform to specifications, enabling multi-format output and data archiving of intelligence products.

3. The interactive method for generating information products in a conversational manner as described in claim 1, characterized in that: Step 2 is implemented as follows: Step 2.1: Call the multi-source data integrator of the resource output layer to access multi-source heterogeneous intelligence data from the upstream intelligence retrieval module or third-party database, parse and extract structured text intelligence and unstructured data tags, and store them in the searchResults state object; Step 2.2: Users manually enter supplementary factual materials through the real-time editing feedback window of the interactive presentation layer, which are synchronously updated to the inputText variable, forming complete original materials with the data loaded by the multi-source data integrator; Step 2.3: Call the process navigator of the interactive presentation layer, initialize the logical states of the five steps of template selection, paragraph recognition, format proofreading, logic verification, and product integration, set the global progress index to 0, establish a Boolean index for each step and mark it as incomplete. Step 2.4: Establish a data dependency verification mechanism to lock the activation permissions of subsequent steps. The state transition can only be triggered after the result of the previous step is generated and the verification is passed.

4. The interactive method for generating information products in a conversational manner as described in claim 1, characterized in that: Step 3 is implemented as follows: Step 3.1: The user triggers the template selection operation through the dialog-style function panel of the interactive presentation layer. The panel receives the template type selected by the user and converts the operation intention into a preset LLM model call request. The request contains the core parameters of template type identifier and original material association index. Step 3.2: Call the Prompt project library of the intelligent processing layer, match the corresponding structured prompt word template for the intelligence domain according to the template type identifier. This template encapsulates the core elements of outline chapter constraints and content professional specifications; at the same time, use the complete original material in the inputText variable as a semantic seed and concatenate it into the prompt word template to generate standardized LLM calling instructions. Step 3.3: Send the encapsulated standardized LLM calling instructions to the LLM model. The model generates a Markdown-formatted outline that conforms to the intelligence compilation specifications based on the original materials and prompt word constraints. Step 3.4: Invoke the process navigator of the interactive presentation layer, receive the non-empty verification signal of the templateResult field returned by the structure parsing state manager, update the boolean index of the template selection step to the completed state; at the same time, increment the global progress index to 1, remove the disabled attribute of the interactive entry of the paragraph recognition step, unlock the activation permission of this step, and complete the process loop of the template selection stage.

5. The interactive method for generating information products in a conversational manner as described in claim 1, characterized in that: Step 4 is implemented as follows: Step 4.1: Receive paragraph recognition execution instructions through the interactive presentation layer's dialog-style function panel, call the intelligent processing layer's process-driven engine to start the handleParagraphRecognition processing program, and retrieve the templateResult outline data from resultMarkdownJson in the structure parsing state manager; use regular expressions and AST syntax trees to parse the Markdown heading levels, filter redundant characters, and group the tasks by using the first-level chapter title as the task index and the corresponding second- and third-level subheadings as logical sub-items to construct a chapter task list; Step 4.2: Call the process-driven engine to traverse the chapter task list, build a dedicated Prompt for each chapter and inject structural context, clarifying the chapter positioning and subheading sequence; perform semantic retrieval through the multi-source data integrator of the resource output layer, embed the matching fact evidence blocks in searchResults into the Prompt, and call the Prompt project library to add writing style constraints such as argument requirements, data citations, and paragraph connections; Step 4.3: The process-driven engine performs batch asynchronous calls to the chapter task queue and sends them to the LLM model, enabling the streaming response listener to receive paragraph text; The real-time editing feedback window dynamically displays the generated content, and the structure parsing status manager fills the paragraph content into the paragraphRecognitionResult status object according to the chapter index; after all chapter content is merged, the currentStep progress index and stepCompleted completion status of the process navigator are updated.

6. The interactive method for generating information products in a conversational manner as described in claim 1, characterized in that: Step 5 is implemented as follows: Step 5.1: Receive format review instructions through the interactive presentation layer's dialog-style function panel, call the intelligent processing layer's structure parsing state manager to read the full text of the paragraphRecognitionResult; call the Prompt project library to obtain the intelligence document verification template, and have the process-driven engine encapsulate the text and verification instructions and send them to the LLM model; receive the model's correction data, automatically replace typos, punctuation, and full / half-width errors, unify military and technical terminology, and finally write the optimized text into the formatCheckResult state object; Step 5.2: Trigger the logic verification command through the dialog-based function panel. The structure parsing status manager retrieves the formatCheckResult data as input. Call the logic verification template of the Prompt project library. The process-driven engine drives the LLM model to scan the causal chain and factual evidence of the chapters, locate and correct contradictions in the data conclusions, automatically supplement transition sentences and unify the narrative tone. Save the logic-optimized text to the logicCheckResult status field and increment the currentStep progress index. Step 5.3: Call the structure parsing state manager to persist the intermediate version of the format check and logic verification in fragments using resultMarkdownJson; The real-time editing feedback window provides a version snapshot index, supporting non-linear jumps and rollback operations for users; it extracts the corresponding version text based on the rollback index, covering the current active editing state, and realizes process retracing.

7. The interactive method for generating information products in a conversational manner as described in claim 1, characterized in that: Step 6 is implemented as follows: Step 6.1: Trigger the product integration command through the dialog-based function panel, call the process-driven engine to start the handleProductIntegration process, recursively scan the entire document and extract the adopted reference evidence blocks from searchResults through the multi-source data integrator; establish the mapping index between factual evidence and the main text, classify the citation sources, and automatically generate the reference chapters according to the specifications by the structure parsing state manager; Step 6.2: Call the global consistency enhancement template of the Prompt project library. The process-driven engine drives the LLM model to refine the logicCheckResult text while retaining the core arguments; automatically insert citation marks in the main text according to the index, merge the optimized text with the references, and generate a complete Markdown draft. Step 6.3: Call the standardized output generator to start the handleExportDocx conversion engine, parse the Markdown draft into an AST abstract syntax tree; map semantic tags to Word document styles according to the selected template, and fill in metadata such as title, security level, and date through the multi-source data integrator to generate an editable .docx document; Step 6.4: The standardized output generator stores the Word document and returns a download link; the process navigator marks the currentStep as complete; the structure parsing state manager packages the initial outline, retrieved evidence, intermediate text, and proofreading notes, and stores them in the historical task library to complete the entire process archiving.

8. The interactive method for generating information products in a conversational manner as described in claim 1, characterized in that: Step 7 is implemented as follows: Step 7.1: The real-time editing feedback window globally listens for modification instructions in steps 2-6, and invokes the editing dialog box; it calls the structure parsing state manager to extract the current step context from resultMarkdownJson and displays it, and receives the user's natural language modification requirements through the userRequirements interface of the dialogic function panel; Step 7.2: The dialog-based function panel triggers the regeneration instruction, and the process-driven engine starts the handleRegenerate function; The structural parsing state manager locates the current stage through the currentStep index, combines the original constraints of the Prompt project library, the current paragraph data and the user's modification requirements, merges and generates a local regeneration instruction and sends it to the LLM model to receive the newly generated content; Step 7.3: The structure parsing state manager uses hot updates to overwrite the corresponding field indexes of resultMarkdownJson with new content; it automatically verifies the validity of subsequent steps and locks or unlocks corresponding activation permissions; the process navigator synchronously updates the step status, supporting unlimited local fine-tuning without disrupting the overall architecture.

9. A tool for generating conversational information products according to the method of claim 1, characterized in that: It includes an interactive presentation layer consisting of a process navigator component, a conversational function panel component, and a real-time editing feedback window; an intelligent processing layer consisting of a process-driven engine component, a Prompt project library component, and a structure parsing state manager; and a resource output layer consisting of a multi-source data integrator and a standardized output generator. The process navigator component is used to generate and manage the logical states of five major steps: template selection, paragraph recognition, format review, logic verification, and product integration, and respond to user click events to trigger step jumps. The conversational function panel component is used to convert the user's task operation instructions into LLM model call requests, and provides interactive entry points for type selection, step execution, and regeneration. The real-time editing feedback window component is used to enable real-time editing, previewing and feedback of the content generated in each step, and supports users to manually modify and regenerate. The process-driven engine component is used to coordinate the execution order and data object transfer between each step, and to execute core business logic processing. The Prompt project library component is used to encapsulate specialized structured prompt word templates for the intelligence field, ensuring that the output content of the LLM model conforms to the expected format; The structure parsing state manager component is used to maintain and store the intermediate results of each stage in real time, supporting process backtracking and context handling. The multi-source data integrator component is used to access and parse multi-source heterogeneous intelligence data from the upstream intelligence retrieval module or a third-party database, providing raw materials for intelligence generation. The standardized output generator component is used to convert the processed Markdown content into standardized documents that conform to the specifications, enabling multi-format output and data archiving of intelligence products.

Images