Project report generation method and device and electronic equipment

Abstract

The invention discloses a project report generation method and device and electronic equipment. Relates to the field of artificial intelligence. The method comprises: receiving a report generation request of a target project, the report generation request comprising at least one of the following: a time range covered by a to-be-generated project analysis report and a project identifier of the target project; based on the report generation request, obtaining a project management data snapshot meeting the time range, and obtaining target management data; based on the target management data, a model consensus strategy is utilized to generate a project analysis report of the target project, and the model consensus strategy comprises the steps of generating an initial analysis report of the target project based on the N large language models and auditing the initial analysis report. The technical problem that the writing efficiency of the project analysis report is low due to the fact that the writing of the project analysis report depends on artificial experience because the function of the project management tool is solidified in the related technology is solved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and more specifically, to a method, apparatus, and electronic device for generating project reports. Background Technology

[0002] To support the complex activities in project management, a large number of mature project management software tools have emerged in the market. However, they share common and profound limitations in report generation and intelligent analysis, specifically including: 1. The static and descriptive nature of reporting functions: The reporting functions provided by tools in related technologies are essentially static snapshots and descriptive statistical summaries of current data in the project database. They can only show "how many tasks are currently in 'in progress' status," "the completion rate of a certain epic (a unit of project requirements)," or "the number of code submissions by team members this week." However, these reports are passive and descriptive; they answer the "what" question but cannot automatically answer the "why" and "what's next" questions. 2. Lack of dynamic comparison and deep reasoning capabilities: The essence of project management lies in understanding change. A key pain point is that these mainstream tools generally lack native, automated dynamic comparison and analysis capabilities. Complex comparison, analysis, and visualization using external tools are required, which is not only time-consuming and labor-intensive, but also highly dependent on the project manager's personal experience, domain knowledge, and data sensitivity in terms of depth and quality, making it difficult to standardize and scale. 3. Heavy cognitive load and analytical bottlenecks: Due to the lack of analytical capabilities of the tools themselves, a significant amount of time must be spent immersing oneself in massive amounts of discrete project data, manually filtering, correlating, integrating, and interpreting information in order to write a valuable weekly, monthly, or project status review report.

[0003] Therefore, the project management tools in related technologies have fixed functions and lack in-depth and dynamic intelligent analysis capabilities, resulting in low efficiency in report writing and a high dependence on human experience.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This invention provides a project report generation method, apparatus, and electronic device to at least solve the technical problem in related technologies where the functions of project management tools are fixed, making the writing of project analysis reports dependent on human experience and resulting in low efficiency in writing project analysis reports.

[0006] According to one aspect of the present invention, a project report generation method is provided, comprising: receiving a report generation request for a target project, wherein the report generation request includes at least one of the following: a time range covered by the project analysis report to be generated, and a project identifier of the target project; based on the report generation request, obtaining a snapshot of project management data conforming to the time range to obtain target management data, wherein the snapshot of project management data includes: a snapshot of project management data of the target project pushed by a project management tool; and based on the target management data, generating a project analysis report for the target project using a model consensus strategy, wherein the model consensus strategy includes: generating an initial analysis report for the target project based on N large language models and reviewing the initial analysis report, wherein N is a positive integer.

[0007] According to another aspect of the present invention, a project report generation apparatus is also provided, comprising: a request processing module, configured to receive a report generation request for a target project, wherein the report generation request includes at least one of the following: a time range covered by the project analysis report to be generated, and a project identifier of the target project; a data retrieval module, configured to obtain a snapshot of project management data conforming to the time range based on the report generation request, thereby obtaining target management data, wherein the snapshot of project management data includes: a snapshot of project management data of the target project pushed by a project management tool; and a report generation module, configured to generate a project analysis report for the target project based on the target management data and using a model consensus strategy, wherein the model consensus strategy includes: generating an initial analysis report for the target project based on N large language models and reviewing the initial analysis report, wherein N is a positive integer.

[0008] According to another aspect of the present invention, an electronic device is also provided, including: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to perform the project report generation method of any of the above via executing the executable instructions.

[0009] According to another aspect of the present invention, a computer-readable storage medium is also provided, which stores a computer program, wherein the computer program controls the device where the computer-readable storage medium is located to execute the project report generation method described above when it is running.

[0010] In this invention, the following method is adopted: receiving a report generation request for a target project, wherein the report generation request includes at least one of the following: the time range covered by the project analysis report to be generated, and the project identifier of the target project; based on the report generation request, obtaining a snapshot of project management data that conforms to the time range to obtain target management data, wherein the project management data snapshot includes: a snapshot of the project management data of the target project pushed by the project management tool; based on the target management data, generating a project analysis report for the target project using a model consensus strategy, wherein the model consensus strategy includes: generating an initial analysis report for the target project based on N large language models and reviewing the initial analysis report, wherein N is a positive integer. This solves the technical problem in related technologies where the functions of project management tools are fixed, making the writing of project analysis reports dependent on human experience, resulting in low efficiency in writing project analysis reports. In this invention, through a consensus strategy of N large language models, a generation + review method is used to obtain the project analysis report, avoiding the low efficiency and poor reliability of writing project analysis reports based on human experience in related technologies, thereby achieving the technical effect of improving the generation efficiency and reliability of project analysis reports. Attached Figure Description

[0011] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0012] Figure 1 This is a flowchart of an optional project report generation method according to an embodiment of the present invention;

[0013] Figure 2 This is an example diagram of a user interface for an optional project report according to an embodiment of the present invention;

[0014] Figure 3 This is an architecture diagram of an optional project report generation system according to an embodiment of the present invention;

[0015] Figure 4 This is a main flowchart of an optional project report generation method according to an embodiment of the present invention;

[0016] Figure 5 This is an optional dual-model generation-review flowchart according to an embodiment of the present invention;

[0017] Figure 6 This is a schematic diagram of an optional project report generation device according to an embodiment of the present invention. Detailed Implementation

[0018] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0019] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0020] For ease of description, some of the terms or nouns involved in this invention are explained below.

[0021] Project management is a key process to ensure that organizational goals are achieved efficiently. Its core lies in the systematic planning, execution, and monitoring of tasks, milestones, resources, budgets, and risks.

[0022] LLM: Large Language Model. A deep learning-based artificial intelligence model trained on massive amounts of text data, possessing powerful capabilities in natural language understanding, generation, reasoning, and knowledge integration.

[0023] AI: Artificial Intelligence.

[0024] API: Application Programming Interface. A set of definitions, protocols, and tools for building and integrating application software.

[0025] JSON: JavaScript Object Notation, a lightweight data interchange format that is easy for humans to read and write, and also easy for machines to parse and generate.

[0026] Markdown: A lightweight markup language that allows people to write documents in an easy-to-read and easy-to-write plain text format, and then convert them into structured HTML (HyperText Markup Language, the standard markup language for creating web pages).

[0027] Project management data snapshot: Full or incremental data in the project management system collected and fixed at a specific point in time (such as a certain day), usually stored in JSON format, to record the project status at that point in time.

[0028] Multi-model consensus mechanism: refers to using two or more large language models to reach a consensus on the accuracy and rationality of generated content through generation-review or multi-party collaboration, in order to improve content quality and reduce the risk of "illusion".

[0029] Generative Model: In the consensus mechanism, this is the model responsible for initially generating the report chapter content based on the prompt words.

[0030] Review Model: In the consensus mechanism, it is responsible for reviewing whether the content of the chapters generated by the generation model is obviously inappropriate or has reasoning errors.

[0031] Illusion: This refers to the phenomenon where large language models fabricate facts, generate erroneous information that contradicts the input information, or create something out of thin air when generating content. This is one of the main challenges currently facing LLM technology.

[0032] It should be noted that the user information (including but not limited to user device information, user personal information, etc.), the collected information and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with the relevant laws, regulations, and standards of the relevant regions, have taken necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation portals for users to choose to authorize or refuse.

[0033] Example 1

[0034] According to an embodiment of the present invention, an optional project report generation method is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0035] Figure 1 This is a flowchart of an optional project report generation method according to an embodiment of the present invention, such as... Figure 1As shown, the method includes the following steps:

[0036] Step S101: Receive a report generation request for the target project, wherein the report generation request includes at least one of the following: the time range covered by the project analysis report to be generated, and the project identifier of the target project.

[0037] The aforementioned report generation request may include, but is not limited to: the time range covered by the project analysis report to be generated (e.g., the time interval of the project analysis report to be generated (i.e., start_date and end_date)) and the project identifier of the target project (e.g., project ID (proj_id)).

[0038] An optional project report generation system can be used to execute the project report generation method provided in this embodiment. In this embodiment, the project report generation system can receive report generation requests sent from the front-end interface or other business systems by calling the API interface. The request body can explicitly specify the project ID that needs to generate a report and the time interval (i.e., start_date and end_date) covered by the report.

[0039] Step S102: Based on the report generation request, obtain a snapshot of project management data within the specified time range to obtain target management data. The snapshot of project management data includes a snapshot of the project management data of the target project pushed by the project management tool.

[0040] In this embodiment, project management data snapshots can be filtered based on the report generation request, according to the project representation and time range, to obtain the target management data. For example, using proj_id, start_date, and end_date as query conditions, two corresponding data snapshot records can be precisely retrieved from the t_proj_snapshot (project management data snapshot table) database table. If data for any date is missing, an error message can be returned.

[0041] In this embodiment, a method comparing data snapshots at multiple time points (e.g., start_date and end_date) can be introduced (using structured management data snapshots of the start date and end date as input), significantly reducing the shortcomings of open domain data retrieval lacking structured verification mechanisms. By locking in project status differences across specific time dimensions, the generated content is strictly anchored to verifiable business data, thereby achieving a dual improvement in report accuracy and logical consistency in scenarios such as project management.

[0042] In one alternative example, for two time points, an important approach is to process data over a continuous time period. For instance, the input could be a snapshot of project data for each day of the past month, or an operational log stream over a period of time. In this case, a data preprocessing module can be added. This module can utilize traditional time-series analysis algorithms (such as ARIMA) or an LLM specifically designed for time-series data analysis to automatically extract key trends, abrupt changes, and anomalies from the continuous data. Then, these preprocessed and refined "summary information" or "insights," instead of the raw, verbose data stream, are fed into the subsequent report generation model (which can also incorporate a consensus mechanism). This improves the ability to handle different types and dimensions of data input, expanding the application scenario from "two-point static comparison" to "continuous process dynamic analysis," covering a wider range of project management data analysis needs.

[0043] Step S103: Based on the target management data, a project analysis report for the target project is generated using a model consensus strategy. The model consensus strategy includes generating an initial analysis report for the target project based on N large language models and reviewing the initial analysis report, where N is a positive integer.

[0044] The aforementioned consensus strategy can use a generative model from N large language models to generate the initial chapter content for each chapter in the project analysis report. Then, an auditing model from the N large language models audits the accuracy of each initial chapter content (e.g., whether there are hallucination issues). For each chapter whose initial chapter content fails auditing, the initial chapter content can be regenerated. If the regenerated content passes auditing, it can be used as the chapter content. If it still fails auditing, the chapter content can be regenerated for auditing again until a preset number of generation attempts are reached. If the last generated chapter content still fails auditing, a warning flag can be added to the last generated chapter content (e.g., "[Generation result is questionable, please adopt with caution!]"). The chapter content with the added warning flag will then be used as the final chapter content to indicate the chapter's content.

[0045] Through the above steps, in this embodiment, a project analysis report is obtained by using a consensus strategy based on N large language models and a generation + review approach. This avoids the inefficiency and poor reliability of project analysis reports written based on human experience in related technologies, thereby improving the generation efficiency and reliability of project analysis reports. Furthermore, it solves the technical problem in related technologies where the functions of project management tools are fixed, making the writing of project analysis reports dependent on human experience and resulting in low efficiency.

[0046] Optionally, the initial analysis report includes: the chapter content associated with each chapter in the project analysis report to be generated; based on the target management data, using a model consensus strategy, generating the project analysis report for the target project, including: generating the chapter content associated with each chapter through N large language models based on the target management data, obtaining M initial chapter contents; reviewing each initial chapter content through N large language models to obtain the review results; and obtaining the project analysis report based on the M initial chapter contents and the review results.

[0047] The initial analysis report mentioned above may include the content of each of the M chapters in the project analysis report to be generated. In one optional example, the M chapters may include, but are not limited to, four core chapters: "Brief Summary of Project Dynamics," "Detailed Display of Project Dynamics," "Project Issues / Risk Analysis," and "Next Steps." The initial chapter content can be in Markdown format.

[0048] In this embodiment, project management data snapshots can be collected and compared at two different dates (start date and end date). By leveraging the powerful contextual understanding and logical reasoning capabilities of Large Language Models (LLM), an in-depth project report containing multiple analytical dimensions can be automatically generated. The report content includes at least four core chapters: "Brief Summary of Project Dynamics", "Detailed Display of Project Dynamics", "Project Issues / Risk Analysis", and "Next Steps".

[0049] Based on target management data and utilizing model consensus strategies, generating project analysis reports for target projects can include the following optional examples:

[0050] Example 1: A model consensus strategy can be a dual-model consensus mechanism (i.e., using two large oracle models for generation and review to obtain the final project analysis report). This mechanism dynamically assigns the two LLMs as a "generation model" and a "review model" during the generation of each chapter of the report. The generation model is responsible for initially producing the chapter content, while the review model performs rigorous fact verification and logical review of the produced chapter content. Through the established retry and consensus-reaching logic, this mechanism can effectively filter and correct errors in the generation process, thereby significantly improving the accuracy, factual consistency, and overall reliability of the final generated report.

[0051] Example 2 further expands the binary "generative model-review model" structure in Example 1 into a multi-party collaborative structure. Three or more LLM agents can be launched simultaneously. In the first round, all agents receive the same prompt (to guide the LLM agents in generating project analysis reports) and independently generate their own answers (e.g., a risk analysis report, which can also be called the initial project analysis report). In the second round, a "debate" phase begins, where each agent receives answers from all other agents and is instructed to analyze and critique the strengths and weaknesses of these answers, revising and improving its own. This process can iterate multiple times. Ultimately, a voting mechanism (majority rule) or a designated "referee" agent can synthesize all final versions of the answers (i.e., project analysis reports) to form an optimal integrated result. Compared to bi-model consensus, multi-party debate can stimulate deeper, multi-faceted analysis, thereby uncovering more hidden problems or providing more comprehensive recommendations. In other words, in scenarios where the "idea" is generalized to an "N-model" (N≥3), it covers more complex, parallel collaborative modes.

[0052] Example 3 leverages the randomness inherent in the LLM generation process by making multiple (e.g., 3, 5, or more) independent generation calls to the same prompt word. The diversity of the generated results can be controlled by adjusting the "temperature" parameter during these calls. After obtaining multiple versions of the answer, natural language processing techniques (such as text embedding and clustering algorithms) are used to analyze these answers, identifying the most recurring and semantically consistent core viewpoints or conclusions, which are then used as the final and most reliable output to obtain the project analysis report. Example 3 has low implementation costs and does not rely on multiple different LLM APIs. This method of improving reliability through multiple generation and consensus-seeking, even with only one physical model, covers an equivalent implementation that trades computational cost for reliability.

[0053] Example 4 employs an asymmetric collaborative architecture. Multiple low-cost, fast LLMs (e.g., small open-source models) can be used as "contestants" to simultaneously generate report content. Then, a more capable, authoritative, but potentially more costly and slower LLM is invoked as a "judge." The judge model scores and ranks all contestants' answers based on a predefined, detailed scoring standard, such as factual accuracy, logical coherence, analytical depth, and language fluency, and provides detailed comments. Ultimately, the highest-scoring answer can be selected, or the judge model can be instructed to combine the strengths of all contestants' answers to generate a final, higher-quality report. This collaborative model, where a higher-order agent supervises and evaluates the outputs of multiple lower-level agents, is a hierarchical and asymmetric implementation of the "generate-review" concept.

[0054] In this embodiment, the model consensus strategy of large language model is used to improve the generation efficiency of project analysis report and avoid the illusion problem that exists in directly generating project analysis report through a single LLM model, thereby achieving the technical effect of improving the reliability of project analysis report.

[0055] Optionally, based on the target management data, the chapter content associated with each chapter is generated through N large language models to obtain M initial chapter content, including: obtaining the first prompt word template associated with each chapter to obtain M first prompt word templates, wherein the first prompt word template associated with each chapter is used to guide the large language model to generate the specific content of that chapter; filling the target management data into the M first prompt word templates to obtain M first prompt words; and generating the chapter content associated with each chapter through N large language models based on the M first prompt words to obtain M initial chapter content.

[0056] In this embodiment, the system can obtain the pre-set prompt word templates (i.e., the first prompt word template) for report chapters (such as summary, details, risks, and recommendations) in the project report generation system. It can select the appropriate template based on the currently generated chapter and accurately fill the template with dynamic information such as the two retrieved data snapshots (start_date_snapshot and end_date_snapshot) and the project name, thereby constructing a structured and information-rich prompt word.

[0057] In one optional example, a "knowledge retrieval" preprocessing step can be added before the "constructing the first prompt word" step. First, key project elements (such as requirement names, task IDs, personnel names, technical modules, etc.) are identified from the input project management data snapshot (i.e., target management data) using keyword extraction or named entity recognition technology. Then, similarity searches are performed using these keywords in a pre-built vector knowledge base. This knowledge base can contain general project management best practices, a library of successful / failed case studies from past projects within the company, industry risk checklists, etc. The most relevant knowledge fragments (text) retrieved are then combined with the original data snapshot to populate the prompt word template. This provides the LLM with richer and more reliable external knowledge as a reference during generation, thus reducing the illusion from the outset.

[0058] The verification process can be internalized into the generation process by changing the structure of the prompts. This means that instead of directly requiring the LLM to output the final analysis conclusion, it instructs it to follow a "chain-of-thought" or "chain-of-verification" pattern for step-by-step reasoning. For example, when analyzing risks, the prompts could be designed as follows: "Please follow these steps for analysis: First, compare the two JSON snapshots one by one, listing all task items whose states have changed or whose data differs. Second, for each change, analyze its potential positive or negative impacts. Third, summarize all negative impacts and assess their severity to form the final risk list." This structured, step-by-step reasoning process forces the model to self-examine and verify, improving the rigor of the logic. This method of designing structured, multi-step prompts to guide the model in self-verification and step-by-step reasoning improves the reliability of the results.

[0059] Table 1 shows a set of prompt templates for constructing prompts, including prompt templates for the first to fourth paragraphs of the report (corresponding to M first prompt templates) and prompts for the generated review results.

[0060] Table 1

[0061]

[0062]

[0063]

[0064] In this embodiment, the project identifier (e.g., project name), time range (e.g., start and end dates), and project management data snapshot text can be dynamically filled into the placeholders of the first prompt word template to generate a complete and specific prompt word, resulting in M ​​first prompt words.

[0065] In this embodiment, each first prompt word can be input into the generation model of N large language models to obtain the initial chapter content corresponding to each chapter, thus achieving the technical effect of accurately generating the initial chapter content of each chapter in the project analysis report.

[0066] Optionally, the N large language models include: a generation model and a review model. The generation model is used to generate the content of each initial chapter, and the review model is used to review the content of each initial chapter. Based on M first prompt words, the N large language models generate the chapter content associated with each chapter, resulting in M ​​initial chapter contents. This includes: inputting the M first prompt words into the generation model to obtain M initial chapter contents; and reviewing each initial chapter content through the N large language models to obtain a review result. This includes: obtaining second prompt words, and reviewing each initial chapter content through the review model based on the second prompt words to obtain a review result. The second prompt words are used to review whether the content of each initial chapter is accurate.

[0067] To avoid the pitfalls of directly inputting data from project management tools as context into a large language model (LLM) and instructing it to generate a project analysis report, this embodiment compares snapshots of project management data representing the target project's state at two or more different points in time, stored in a database. Using at least two large language models (i.e., a generation model and an auditing model), and through a consensus mechanism that includes generation and auditing steps, an analytical project report containing information on project dynamics, risks, or recommendations can be automatically generated. The pitfalls of directly generating a project analysis report based on a single LLM include, but are not limited to: (1) The "illusion" problem leads to highly unreliable information: Illusion refers to the fact that the model may generate content that appears smooth and reasonable, but is actually completely inconsistent with the input facts, fabricated out of thin air, or logically flawed. In the serious business scenario of project management, which requires a high degree of factual accuracy, decisions based on "illusion" information are disastrous. For example, an incorrect risk assessment may lead to the misallocation of resources, and a fictitious schedule update may cover up the real project delays, which may ultimately lead to budget overruns, delivery failures, and loss of customer trust. Therefore, the low reliability of content generated by a single LLM is the main obstacle to its application in the generation of critical reports. (2) Lack of transparency in the process leads to a lack of trust: The reasoning process of a single LLM is like a "black box," and users find it difficult to understand how it derives the final analytical conclusions from the raw input data. When an LLM report points out that " When "requirement X is at risk of delay", it is impossible to trace the basis for the judgment. Is it based on the stagnation of the task status or the increase of related defects? This lack of interpretability seriously weakens the manager's trust in AI-generated content, making them hesitant to adopt these conclusions as the basis for decision-making. (3) Insufficient stability of complex reasoning tasks: Risks and problems in projects are generally not determined by a single data indicator, but are the result of multiple factors being interconnected and dynamically evolving. For example, identifying an "integration risk" may require analyzing the progress of multiple modules, the status of interface definition documents, and the workload of relevant developers at the same time. Such complex, context-aware reasoning tasks place extremely high demands on the stability of LLM. Without specific guidance and verification mechanisms, it is difficult for a single LLM to guarantee that it can stably and accurately complete such tasks every time it is generated.

[0068] The aforementioned second prompt word can be obtained from the prompt word template of the review result (hereinafter referred to as the second prompt word template). As shown in Table 1, an optional prompt word template for the review result is provided. In this embodiment, for each initial chapter content, the original first prompt word corresponding to the initial chapter content and the initial chapter content can be filled into the prompt word template of the review result to obtain the second prompt word corresponding to the initial chapter content. By inputting the second prompt word corresponding to each initial chapter content into the review model, the review result of the initial chapter content can be obtained.

[0069] The review result can be a JSON string. In this embodiment, the review result can be parsed. If the parsing is successful and the value of the "Review Result" field is "No" (or other predefined synonyms for "approved", such as "Pass" or "No Objection"), then the consensus is confirmed (i.e. the review is passed). If the value of the "Review Result" field is "Yes" (indicating "Disapproved"), or the JSON parsing fails, then the consensus is confirmed to have failed (i.e. the review is not passed), and a retry is initiated (i.e., the initial chapter content is regenerated and reviewed).

[0070] In this embodiment, the initial chapter content of the project analysis report is generated by generating a model, and the initial chapter content is reviewed by an auditing model. This avoids the situation where it is difficult to detect illusion problems when directly generating a project analysis report using a single LLM model, and achieves the technical effect of improving the reliability of the generated project analysis report.

[0071] Optionally, before generating chapter content associated with each chapter through N large language models based on the target management data to obtain M initial chapter contents, the method further includes: obtaining a model list, wherein the model list includes: model identifiers of M large language models, where M is a positive integer greater than N; and determining N large language models based on the model list using a preset selection strategy, wherein the preset selection strategy includes: random selection.

[0072] In this embodiment, a list containing identifiers of LLMs (Large Language Models) (i.e., a model list) can be predefined. In this embodiment, the model list can be obtained, and N large language models can be randomly selected from the model list. For example, a list containing available LLM identifiers can be defined, such as models=['Large Language Model 1 / Version', 'Large Language Model 2 / Version']. At the beginning of each attempt (including the first and retry) to generate chapter content, the models list (i.e., the model list) is randomly shuffled. Then, the first element of the list (i.e., the model identifier) ​​is taken as the generated model A, and the second element is taken as the review model B. This randomization strategy can avoid systematic bias and prevent a particular model from dominating the consensus result for a long time due to its inherent generation or review style, thus achieving the technical effect of improving the reliability of the generated project analysis report.

[0073] Optionally, based on the content of M initial chapters and the review results, a project analysis report is obtained, including: if the review results indicate that all M initial chapter contents have passed the review, integrating the M initial chapter contents and the generation process of each initial chapter content to obtain a project analysis report; if the review results indicate that any initial chapter content has failed the review, regenerating the initial chapter content that failed the review to obtain the target chapter content, and obtaining the project analysis report based on the target chapter content and the initial chapter content that has passed the review in the review results.

[0074] In this embodiment, if all M initial chapter contents have been approved, the M initial chapter contents can be compiled and integrated with the basic information of the target project, the start and end times of the data range included in the project analysis report, and the "reasoning process" of the report generation into a single, aesthetically pleasing HTML page (i.e., rendering the initial chapter contents into a structured HTML page), thus obtaining the project analysis report. Figure 2 This is an example diagram of a user interface for an optional project report according to an embodiment of the present invention, such as... Figure 2 As shown, when a "View Reasoning Process" button can be set below each chapter in the HTML page, when a user clicks the "View Reasoning Process" button below each chapter, the relevant log can be displayed through a pop-up or expanded form to enhance the interpretability of the report.

[0075] For example, after all chapter content has been successfully generated and approved, the initial chapter content (i.e., the individual Markdown format report chapters) can be compiled. This chapter content, along with basic project information, report start and end dates, and the "reasoning process" text recorded during the consensus process, can be integrated into a single, aesthetically pleasing HTML page. In one optional example, front-end interactive logic can also be implemented within the project report generation system. For instance, the "reasoning process" text can be hidden by default, and when the user clicks the "View Reasoning Process" button below each chapter, relevant logs can be displayed via pop-ups or expansions to enhance the report's interpretability.

[0076] In one optional example, for each chapter whose initial content fails review, the initial content can be regenerated. If the regenerated content passes review, it can be used as the chapter content. If it still fails review, the chapter content can be regenerated for review again until a preset number of generation attempts are reached. If the last generated chapter content still fails review, a warning mark can be added to it (e.g., "[Generation result is questionable, please use with caution!]"). The chapter content with the warning mark added will then be used as the final chapter content to indicate the chapter's content. This improves the reliability of the generated project analysis report.

[0077] Optionally, if the review result indicates that any initial chapter content has failed the review, the initial chapter content that failed the review is regenerated to obtain the target chapter content. This includes: obtaining the review reason and the number of retries for the initial chapter content that failed the review, wherein the review reason includes the reason for the failure, and the number of retries includes the number of times the initial chapter content that failed the review is regenerated; for any initial chapter content that failed the review, if the number of retries has not reached a preset threshold, the first prompt word associated with the initial chapter content is adjusted based on the review reason to obtain the adjusted first prompt word; based on the adjusted first prompt word, the initial chapter content is regenerated through N large language models and the regenerated initial chapter content is reviewed; if the regenerated initial chapter content passes the review, the regenerated initial chapter content is determined as the target chapter content; if the number of retries reaches the preset threshold, a warning mark is added to the initial chapter content that failed the review last time, thus obtaining the target chapter content.

[0078] For initial chapters that fail the review, a retry phase can be initiated. The name of the review model and its returned "review reason" are recorded in `reasoning_log` (an interaction log used to record the entire consensus process). The retry counter `retry_count` is incremented by 1. It is then checked whether `retry_count` has reached `MAX_RETRIES` (a preset threshold for the number of retry attempts). If not, the chapter content is regenerated. To improve the success rate of retrying, the original prompt (first prompt word) can be enhanced in the next generation of chapter content by inputting the reason for the previous failure into the generation model. For example, appending "Please note: The previous generation result was rejected due to '[review reason]'. Please avoid the same problem in this generation" to the end of the prompt. If `MAX_RETRIES` has been reached, it means that consensus cannot be reached within the limited number of retry attempts, and the retry loop terminates. If the loop exits due to reaching the maximum number of retry attempts, the system will adopt the last result generated by the generation model, but will append a clear, user-visible warning label before its content, such as "[The generation result is questionable; please adopt with caution!]". The function returns this warning content and the failure status. By combining "post-verification" (consensus mechanism) with "pre-election reinforcement" (RAG), a dual-protection system is formed. This protects any methods that enhance the reliability of the context by introducing external knowledge before generation.

[0079] In this embodiment, chapter-level fact verification and adaptive retry logic are introduced (when the review model detects factual errors or logical contradictions, the content reconstruction process is automatically triggered), which fundamentally solves the "illusion" technical defect caused by relying on a single LLM to generate content (i.e., the deviation between the model output and objective facts or semantic requirements).

[0080] Optionally, before obtaining the target management data by acquiring project management data snapshots within the specified time range based on the report generation request, the method further includes: receiving project management data snapshots of the target project sent by the project management tool through a first data interface, formatting the project management data snapshots to obtain formatted project management data snapshots, and sending the formatted project management data snapshots to the target storage space through a second data interface. The formatted project management data snapshots include: a project identifier and the generation time of the data snapshots. Obtaining the target management data by acquiring project management data snapshots within the specified time range based on the report generation request includes: filtering project management data snapshots within the specified time range from the target storage space based on the project identifier to obtain the target management data.

[0081] In this embodiment, project management data pushed by external systems can be received through a standardized API interface (corresponding to the first data interface, e.g., / insertProjSnapshot). The received data can be formatted as a "project management data snapshot" (a JSON structure) and persistently stored in the system's target storage space (e.g., a database (e.g., the t_proj_snapshot table) along with metadata such as project identifier, project name, and snapshot date). A report generation request from the front-end user interface or other business systems can also be received through an API interface (corresponding to the second data interface, e.g., / generateProjReport). The request body explicitly specifies the project ID for which a report needs to be generated and the time interval covered by the report (i.e., start_date and end_date) (corresponding to the aforementioned time range).

[0082] Interface 1 (corresponding to the first data interface): / insertProjSnapshot (data entry snapshot)

[0083] Request body (Content-Type: application / json):

[0084] JSON:

[0085] {

[0086] "proj_name": "xx project",

[0087] "proj_id": "PROJ_001",

[0088] "snapshot_date": "2025-09-12",

[0089] "snapshot": "{...}" / / This is the complete JSON string containing the project data;

[0090] }

[0091] Processing logic: Upon receiving this request, the backend service first performs validity checks on the input data, such as checking if proj_id is empty and if snapshot_date is in a valid date format. If the checks pass, the data in the request body is assembled into a database record object and inserted into the t_proj_snapshot table.

[0092] Interface 2 (corresponding to the second data interface): / generateProjReport (generates a project report)

[0093] Request body (Content-Type: application / json):

[0094] JSON:

[0095] {

[0096] "proj_id": "PROJ_001",

[0097] "start_date": "2025-09-12",

[0098] "end_date": "2025-09-20";

[0099] }

[0100] Processing Logic: This interface 2 is the entry point driving the entire intelligent report generation process. Upon receiving a request, its controller or view function can invoke the core report generation service, passing in `proj_id`, `start_date`, and `end_date`. The service layer will then orchestrate all subsequent steps, including data retrieval, iterative chapter generation, and report integration.

[0101] In one optional example, the structure of the project management data snapshot table (t_proj_snapshot) is shown in Table 2.

[0102] Table 2

[0103]

[0104] In this embodiment, the design of storing the snapshot field in JSON format offers significant engineering advantages. The data model of a project management tool evolves with business development, potentially adding new data entities (such as "code commit records" or "environment deployment records") or new fields to existing entities. Using the flexible JSON format allows for compatibility with these changes without modifying the backend database table structure, greatly enhancing the system's forward compatibility and scalability.

[0105] In this embodiment, by introducing dynamic comparative analysis of data snapshots at different points in time, the project's progress trend, efficiency changes, emerging problems, and potential risks can be automatically and deeply inferred during this period. This provides project managers with unprecedented, high-value decision support information, achieving a fundamental shift from "data presentation" to "intelligent insight." To avoid the inherent "illusion" problem of large language models, a multi-model consensus mechanism provides an effective engineering solution. By introducing the separation of "generator-reviewer" roles and a multi-round cross-validation game process, factual errors and logical fallacies in the generated content can be greatly filtered and corrected. This transforms AI-generated analytical conclusions from "wild ideas" to "data-driven, trustworthy judgments," fundamentally solving the core "trust" pain point when applying AI to critical business reporting scenarios and significantly improving the reliability and credibility of AI-generated reports.

[0106] This implementation avoids project managers spending significant time on tedious data export, organization, comparative analysis, and report writing. By automating the generation of high-quality analytical reports, project managers are freed from these low-value, repetitive tasks. This allows them to dedicate their valuable energy to more creative and valuable work such as higher-level strategic planning, complex risk management, cross-departmental coordination, and team leadership, thereby improving overall project management efficiency and organizational operational efficiency. It significantly enhances project management efficiency while reducing labor costs. Unlike the "black box" operation of a single LLM model, this implementation, while pursuing automation and intelligence, highly values ​​process transparency. It not only provides the final analysis report but also carefully designs and preserves the complete "reasoning chain" behind each analytical conclusion. By clicking the "View Reasoning Process" function, users can clearly see the original prompts used to generate the conclusion, the model's multiple attempts during the consensus process, and the specific opinions on model review. This greatly enhances user trust in the system's results and provides clear and traceable evidence for auditing, review, and problem localization of abnormal results.

[0107] Example 2

[0108] Embodiment 2 of the present invention provides an optional project report generation system (hereinafter referred to as the system). Figure 3 This is an architecture diagram of an optional project report generation system according to an embodiment of the present invention, such as... Figure 3 As shown, the project report generation system includes: a data acquisition and storage module, a report generation request processing module, a data retrieval and preprocessing module, a core report generation engine, and a report integration and formatting module, as detailed below:

[0109] 1. Data Acquisition and Storage Module: As the system's data input layer, this module is responsible for interfacing with various external project management tools. It receives project management data pushed by external systems via POST through a standardized API interface (e.g., / insertProjSnapshot). The received data is formatted as a "project management data snapshot" (a JSON structure) and, along with metadata such as project identifier, project name, and snapshot date, is persistently stored in the system's database (corresponding to the target storage space in Implementation Example 1, e.g., the t_proj_snapshot table (project management data snapshot table)).

[0110] 2. Report Generation Request Processing Module (hereinafter referred to as the Request Processing Module): As the system's interaction entry point, this module provides an API interface (e.g., / generateProjReport) to receive report generation requests from the front-end user interface or other business systems. The request body explicitly specifies the project ID for which a report needs to be generated and the time range covered by the report (i.e., start_date and end_date).

[0111] 3. Data Retrieval and Preprocessing Module: This module is activated upon receiving a report generation request. Based on the proj_id, start_date, and end_date provided in the request, this module accurately retrieves two corresponding project management data snapshots from the database (corresponding to the target management data in Example 1). These two snapshots form the basis for all subsequent analysis and reasoning.

[0112] 4. Core Report Generation Engine: This is the "brain" of the system, the core responsible for all intelligent analysis and content generation. This engine consists of multiple collaborative units:

[0113] a. Prompt Term Construction Unit: This unit has a built-in set of prompt term templates for different report sections (such as summary, details, risks, and recommendations). It selects the appropriate template based on the currently generated section and accurately fills the template with dynamic information such as the two retrieved data snapshots (start_date_snapshot and end_date_snapshot) and the project name, thereby constructing structured and information-rich prompt terms. Table 1 in Example 1 shows a set of prompt term templates used by the prompt term construction unit.

[0114] b. Multi-model consensus execution unit: The key innovation of this embodiment lies in its orchestration and execution of the "dual-model consensus mechanism" (corresponding to the model consensus strategy in Embodiment 1). It receives prompts generated by the prompt word construction unit, then schedules the generation and review of the model, managing the entire "generation-review-retry" loop logic until consensus is reached or the maximum number of retries is achieved. Its output is validated, high-quality Markdown format report content.

[0115] c. Large Model Interface Unit: As an adapter for interacting with the underlying AI capabilities, this unit encapsulates the calling logic for APIs of different large language models (Model 1, Model 2, etc. in this embodiment). It can handle details such as authentication, request formatting, network communication, and response parsing, providing a unified and concise LLM calling interface for upper-layer units.

[0116] 5. Report Integration and Formatting Module: As the system's output layer, this module is responsible for assembling the four independent Markdown-formatted report sections generated by the core report generation engine. It integrates these section contents, along with basic project information, report start and end dates, and the "reasoning process" text recorded during the consensus process, into a single, aesthetically pleasing HTML page (example user interface shown in the image). Figure 2 (As shown). This module also needs to implement front-end interaction logic. For example, the "reasoning process" text should be hidden by default. When the user clicks the "View Reasoning Process" button below each chapter, the relevant logs should be displayed through a pop-up or expanded window to enhance the interpretability of the report.

[0117] Figure 4 This is a main flowchart of an optional project report generation method according to an embodiment of the present invention, such as... Figure 4 As shown, the process includes: initial generation and review. A report fragment (i.e., initial chapter content) is generated through generator model A (i.e., generation model). The report fragment is reviewed through reviewer model B (i.e. review model). If the review is passed, the final Markdown format content can be output. If the review is failed, the report fragment can be regenerated and reviewed again. If the number of retries is greater than 3, the last generation result can be adopted and a warning mark can be added.

[0118] The following is a detailed explanation of this embodiment:

[0119] 1. The hardware and environment configuration of this system is as follows:

[0120] The system described in this embodiment does not depend on specific, proprietary hardware devices and can be deployed in a general server environment. A typical implementation environment includes:

[0121] Server: One or more cloud servers or local physical servers, preferably with a preset operating system.

[0122] Operating environment: A specified version or higher of the target programming system. The system backend service can be built using mainstream web frameworks to facilitate the rapid development of efficient and stable API interfaces.

[0123] Dependencies: A series of Python libraries need to be installed to support the various functions, including but not limited to: (1) Flask or FastAPI: used to build Web API services. (2) SQLAlchemy: used to interact with relational databases and provide Object-Relational Mapping (ORM) capabilities. (3) Requests or httpx: used to send HTTP requests to API endpoints of third-party Large Language Models (LLMs). (4) Python-dotenv: used to manage environment variables, such as API keys. (5) Database: Any mainstream relational database can be used to store snapshots of project data.

[0124] LLM API Access: It is necessary to ensure that the network of the deployment environment can access the API endpoints of at least two different large language model providers. In this example, access to "Large Language Model 1" and "Large Language Model 2" can be used as an example. Sensitive information such as API keys should be managed securely, such as through environment variables.

[0125] 2. Data Model Design:

[0126] The core data in this embodiment is stored in the project management data snapshot table, and its logical structure is shown in Table 2 of Embodiment 1. The design of storing the snapshot field in JSON format in this embodiment offers significant engineering advantages. The data model of the project management tool will continuously evolve with business development, and new data entities (such as "code commit records" and "environment deployment records") or new fields may be added to existing entities. Using the flexible JSON format, the system can accommodate these changes without modifying the backend database table structure, greatly enhancing the system's forward compatibility and scalability.

[0127] 3. Implementation details of the core modules:

[0128] 3.1 Interface Implementation:

[0129] This system provides two core API interfaces. Interface 1: / insertProjSnapshot (data snapshot) and Interface 2: / generateProjReport (project report). The request bodies of Interface 1 and Interface 2 are shown in Example 1, and will not be described again here.

[0130] 3.2 Detailed implementation of the multi-model consensus mechanism:

[0131] Figure 5 This is an optional dual-model generation-review flowchart according to an embodiment of the present invention, and its detailed workflow is as follows: Figure 5 As shown. The implementation is explained below through textual description, and the process is described as follows:

[0132] 1. STEP 1: Initialization. The function entry point receives a pre-constructed prompt. Define a list containing available LLM identifiers (corresponding to the model list in Example 1), for example, models=['Large Language Model 1 / Version', 'Large Language Model 2 / Version']. Set the maximum number of retries MAX_RETRIES=3 (corresponding to the preset threshold in Example 1), and initialize the current retry counter retry_count to 0. Create a list reasoning_log to record the interaction logs of the entire consensus process.

[0133] 2. STEP 2: Random Role Assignment. At the start of each attempt (including the first and retry) to generate chapter content, the list of models is randomly shuffled. The first element of the list is then taken as the generating model A, and the second element as the reviewing model B. This randomization strategy aims to avoid systematic bias and prevent a particular model from dominating the consensus result in the long term due to its inherent generation or review style.

[0134] 3. STEP 3: Content Generation. Invoke the large model interface unit and send a prompt to model A. Wait for and receive the preliminary generated report subsection (generated_report_subsection) returned by model A. Record the model name used in this generation, the prompt (or its summary), and the generated text in the reasoning_log.

[0135] 4. STEP 4: Construct the review prompt. Based on the template defined in the technical solution, dynamically construct a review prompt (review_prompt) for review. This prompt must contain sufficient information for the review model to make a judgment.

[0136] 5. STEP 5: Content Review. Call the large model interface unit to send review_prompt to review model B and obtain the review result review_result returned by it.

[0137] 6. STEP 6: Consensus Decision. Parse review_result (usually a JSON string).

[0138] If the parsing is successful and the "Review Result" field has a value of "No" (or another predefined synonym for "approved," such as "Passed" or "No Objection"), then consensus is considered reached. The process ends successfully, and the function returns `generated_report_subsection` and a success status.

[0139] If the "Audit Result" field value is "Yes" (meaning "Not Approved"), or if JSON parsing fails, consensus is considered to have failed, and the retry logic is initiated.

[0140] 7. STEP7: Retry logic.

[0141] Record the name of audit model B and the "audit reason" it returns in the reasoning_log.

[0142] Increment the retry counter retry_count by 1.

[0143] Check if retry_count has reached MAX_RETRIES.

[0144] If the desired result is not achieved, return to STEP3 and begin a new round of attempts. To improve the success rate of retrying, the original prompt can be enhanced during the next generation by informing the generation model of the reason for the previous failure. For example, append the following to the end of the prompt: "Please note: The previous generation result was rejected due to '[Reason for review]'. Please avoid the same problem in this generation."

[0145] If the target has been reached, it means that a consensus could not be reached within the limit of attempts, and the retry loop terminates.

[0146] 8. STEP 8: Final Processing. If the loop exits due to reaching the maximum number of retries, the system will adopt the last result generated by model A, but will append a clear, user-visible warning, such as "[The generated result is questionable; please adopt with caution!]". The function returns this warning message and the failure status.

[0147] The following is an explanation of a specific example from the "Project Issues / Risk Analysis" section of the "PROJ_001 Project" document:

[0148] Prerequisites: The report generation request requests the generation of a report for project "PROJ_001" from "2025-09-12" to "2025-09-20". Data snapshots for these two days are already stored in the database. Assume that in `start_date_snapshot`, the progress of request STORY_001 is 30%; while in `end_date_snapshot`, its progress becomes 35%, but the actual end time field remains null.

[0149] Step 1: Construct prompt words.

[0150] The system retrieves two snapshot JSON texts from the database and constructs a complete prompt using the prompt word templates in Table 1. The instruction model analyzes data changes and identifies problems / risks.

[0151] Step 2: First attempt (simulated consensus successful).

[0152] The function `generate_subsection_with_consensus` (used to generate report sections using a bi-model consensus mechanism. It returns a dictionary containing section content, status, and logs) is invoked.

[0153] Assume that Large Language Model 1 is randomly assigned as the generation model and Large Language Model 2 as the review model. After analyzing the data, Large Language Model 1 generates the following content: 1. Risk of schedule delay The progress of Request STORY_001 has only increased by 5% (from 30% to 35%) in 8 days, which is slow. Considering that its expected end date is approaching, but the actual end date is still uncertain, there is a significant risk of delay.

[0154] The system generates review prompts and sends them to the Large Language Model 2. After review, the Large Language Model 2 deems the analysis data-driven and logically sound, and returns a JSON response: {"Review Result": "No", "Reason for Review": "The analysis is reasonable and accurately infers the potential delay risk from data changes (slow progress)."}

[0155] Consensus reached. The function successfully returned the generated content, and this chapter is now complete.

[0156] Step 3: Another attempt (simulating consensus failure and retry).

[0157] Suppose that in a certain call, the generative model (large language model 1) "illusoryly" generates the following content: 1. Human resource risks The conclusion that project member Zhang San's workload is insufficient and his contribution to the project is low is not supported by any data snapshot.

[0158] During the audit, the audit model (Large Language Model 2) will find that this conclusion cannot be supported by the two provided JSON data sets. Therefore, it can return: {"Audit Result": "Yes", "Audit Reason": "The conclusion 'Project member Zhang San's workload is insufficient and his contribution to the project is low' is an unfounded guess (illusion). The input project data snapshot does not contain any information reflecting that Zhang San's workload is insufficient."}

[0159] Consensus failed. Retry_count is incremented by 1.

[0160] The system will then proceed to the next loop. In the new loop, the prompt can be enhanced; for example, it could be appended: "Your previous attempt was rejected for the following reason: 'The conclusion that Zhang San is underemployed is an illusion and is not supported by the provided data.' Please generate a new response based on the provided JSON data." The enhanced prompt will guide the generative model to focus more on factual data in the next generation, thereby increasing the probability of reaching a consensus.

[0161] In this embodiment, by introducing dynamic comparative analysis of data snapshots at different points in time, the project's progress trend, efficiency changes, emerging problems, and potential risks can be automatically and deeply inferred during this period. This provides project managers with unprecedented, high-value decision support information, achieving a fundamental shift from "data presentation" to "intelligent insight." To avoid the inherent "illusion" problem of large language models, a multi-model consensus mechanism provides an effective engineering solution. By introducing the separation of "generator-reviewer" roles and a multi-round cross-validation game process, factual errors and logical fallacies in the generated content can be greatly filtered and corrected. This transforms AI-generated analytical conclusions from "wild ideas" to "data-driven, trustworthy judgments," fundamentally solving the core "trust" pain point when applying AI to critical business reporting scenarios and significantly improving the reliability and credibility of AI-generated reports.

[0162] Example 3

[0163] Embodiment 3 of the present invention provides an optional project report generation device, wherein each implementation unit in the project report generation device corresponds to each implementation step in Embodiment 1.

[0164] Figure 6 This is a schematic diagram of an optional project report generation device according to an embodiment of the present invention, such as... Figure 6As shown, the system includes: a request processing module 61, a data retrieval module 62, and a report generation module 63. The request processing module 61 receives a report generation request for the target project, wherein the report generation request includes at least one of the following: the time range covered by the project analysis report to be generated, and the project identifier of the target project. The data retrieval module 62, based on the report generation request, obtains a snapshot of project management data conforming to the time range to obtain target management data, wherein the project management data snapshot includes a snapshot of the project management data of the target project pushed by the project management tool. The report generation module 63, based on the target management data, generates a project analysis report for the target project using a model consensus strategy, wherein the model consensus strategy includes: generating an initial analysis report for the target project based on N large language models and reviewing the initial analysis report, where N is a positive integer.

[0165] In the project report generation apparatus provided in this embodiment of the invention, a request processing module 61 can receive a report generation request for a target project. The report generation request includes at least one of the following: the time range covered by the project analysis report to be generated, and the project identifier of the target project. A data retrieval module 62, based on the report generation request, obtains a snapshot of project management data conforming to the time range to obtain target management data. The project management data snapshot includes a snapshot of the project management data of the target project pushed by the project management tool. A report generation module 63, based on the target management data and utilizing a model consensus strategy, generates a project analysis report for the target project. The model consensus strategy includes generating an initial analysis report for the target project based on N large language models and reviewing the initial analysis report, where N is a positive integer. This solves the technical problem in related technologies where the functions of project management tools are fixed, making the writing of project analysis reports dependent on human experience, resulting in low efficiency in writing project analysis reports. In this embodiment, a consensus strategy of N large language models is used to obtain the project analysis report through a generation + review method, avoiding the low efficiency and poor reliability of writing project analysis reports based on human experience in related technologies, thereby achieving the technical effect of improving the generation efficiency and reliability of project analysis reports.

[0166] Optionally, in the project report generation apparatus provided in this embodiment of the invention, the initial analysis report includes: the chapter content associated with each chapter in the project analysis report to be generated; the report generation module includes: a generation submodule, used to generate the chapter content associated with each chapter based on the target management data through N large language models to obtain M initial chapter contents; an auditing submodule, used to audit each initial chapter content through N large language models to obtain audit results; and a processing submodule, used to obtain the project analysis report based on the M initial chapter contents and the audit results.

[0167] Optionally, in the project report generation apparatus provided in this embodiment of the invention, the generation submodule includes: an acquisition unit, used to acquire the first prompt word template associated with each chapter to obtain M first prompt word templates, wherein the first prompt word template associated with each chapter is used to guide the large language model to generate the specific content of the chapter; a filling unit, used to fill the target management data into the M first prompt word templates to obtain M first prompt words; and a generation unit, used to generate the chapter content associated with each chapter based on the M first prompt words through N large language models to obtain M initial chapter content.

[0168] Optionally, in the project report generation device provided in this embodiment of the invention, the N large language models include: a generation model and an auditing model. The generation model is used to generate the content of each initial chapter, and the auditing model is used to audit the content of each initial chapter. The first generation unit includes: an input subunit, used to input M first prompt words into the generation model respectively to obtain M initial chapter contents; the auditing submodule includes: a processing unit, used to obtain second prompt words, and based on the second prompt words, audit the content of each initial chapter through the auditing model to obtain an audit result, wherein the second prompt words are used to audit whether the content of each initial chapter is accurate.

[0169] Optionally, in the project report generation device provided in this embodiment of the invention, the report generation module further includes: an acquisition submodule, used to acquire a model list before generating chapter content associated with each chapter through N large language models based on target management data to obtain M initial chapter contents, wherein the model list includes: model identifiers of M large language models, where M is a positive integer greater than N; and a determination submodule, used to determine N large language models based on the model list using a preset selection strategy, wherein the preset selection strategy includes: random selection.

[0170] Optionally, in the project report generation apparatus provided in this embodiment of the invention, the processing submodule includes: an integration unit, used to integrate the M initial chapter contents and the generation process of each initial chapter content when the review result indicates that all M initial chapter contents have passed the review, to obtain a project analysis report; and a second generation unit, used to regenerate the initial chapter contents that failed the review when the review result indicates that any initial chapter content has failed the review, to obtain the target chapter content, and to obtain the project analysis report based on the target chapter content and the initial chapter contents that have passed the review in the review result.

[0171] Optionally, in the project report generation device provided in this embodiment of the invention, the second generation unit includes: an acquisition subunit, used to acquire the review reasons and retry counts of initial chapter content that failed the review, wherein the review reasons include the reasons for the failure to pass the review, and the retry counts include the number of times the initial chapter content that failed the review is regenerated; a processing subunit, used to, for any initial chapter content that failed the review, adjust the first prompt word associated with the initial chapter content based on the review reasons when the number of retry counts has not reached a preset threshold, to obtain the adjusted first prompt word, regenerate the initial chapter content through N large language models based on the adjusted first prompt word, and review the regenerated initial chapter content, and determine the regenerated initial chapter content as the target chapter content when the retry counts pass the review; and an adding subunit, used to, when the number of retry counts reaches a preset threshold, add a warning mark to the initial chapter content that failed the review for the last time, to obtain the target chapter content.

[0172] Optionally, in the project report generation apparatus provided in this embodiment of the invention, the project report generation apparatus further includes: a data storage module, which, before obtaining the target management data by acquiring a project management data snapshot conforming to the time range based on the report generation request, receives the project management data snapshot of the target project sent by the project management tool through a first data interface, formats the project management data snapshot to obtain a formatted project management data snapshot, and sends the formatted project management data snapshot to the target storage space through a second data interface, wherein the formatted project management data snapshot includes: a project identifier and the generation time of the data snapshot; the data retrieval module includes: a filtering submodule, used to filter project management data snapshots conforming to the time range from the target storage space based on the project identifier to obtain the target management data.

[0173] The aforementioned project report generation device may also include a processor and a memory. The aforementioned request processing module 61, data retrieval module 62, and report generation module 63 are all stored in the memory as program units, and the processor executes the aforementioned program units stored in the memory to realize the corresponding functions.

[0174] The aforementioned processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured. By adjusting kernel parameters, a consensus strategy based on N large language models is used to generate and review project analysis reports. This avoids the inefficiency and poor reliability of writing project analysis reports based on human experience in related technologies, thus achieving the technical effect of improving the generation efficiency and reliability of project analysis reports.

[0175] The aforementioned memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0176] According to another aspect of the present invention, an electronic device is also provided, including: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to perform the project report generation method of any of the above via executing the executable instructions.

[0177] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0178] In the above embodiments of the present invention, the descriptions of each embodiment have their own emphasis. Parts not described in detail in a certain embodiment can be referred to in the relevant descriptions of other embodiments. It should be understood that the disclosed technical content can be implemented in other ways in the several embodiments provided in this application. The device embodiments described above are merely illustrative. For example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Furthermore, the displayed or discussed mutual coupling or direct coupling or communication connection can be through some interfaces; the indirect coupling or communication connection of units or modules can be electrical or other forms.

[0179] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units.

[0180] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0181] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A project report generation method characterized by, Comprise: Receiving a report generation request of a target project, wherein the report generation request includes at least one of the following: a time range covered by the project analysis report to be generated, a project identifier of the target project; Based on the report generation request, obtain the project management data snapshot conforming to the time range, and obtain the target management data, wherein the project management data snapshot includes the data snapshot of the project management data of the target project pushed by the project management tool; Based on the target management data, generate the project analysis report of the target project by using a model consensus strategy, wherein the model consensus strategy comprises: generating an initial analysis report of the target project based on N large language models and auditing the initial analysis report, wherein N is a positive integer.

2. The project report generation method of claim 1, wherein, The initial analysis report includes chapter content associated with each chapter of the project analysis report to be generated, and the project analysis report of the target project is generated based on the target management data by using a model consensus strategy, comprising: Based on the target management data, generate chapter content associated with each chapter by N large language models, and obtain M initial chapter contents; Audit each of the initial chapter contents by N large language models to obtain an audit result; Based on M initial chapter contents and the audit result, the project analysis report is obtained.

3. The project report generation method of claim 2, wherein, Based on the target management data, generate chapter content associated with each chapter by N large language models, and obtain M initial chapter contents, comprising: Obtain a first prompt word template associated with each chapter, and obtain M first prompt word templates, wherein the first prompt word template associated with each chapter is used to guide the large language model to generate the specific content of the chapter; Fill the target management data into M first prompt word templates to obtain M first prompts; Based on M first prompts, generate chapter content associated with each chapter by N large language models to obtain M initial chapter contents.

4. The project report generating method according to claim 3, characterized by, N large language models include a generation model and an auditing model, the generation model is used to generate each of the initial chapter contents, and the auditing model is used to audit each of the initial chapter contents, Based on M first prompts, generate chapter content associated with each chapter by N large language models to obtain M initial chapter contents, comprising: inputting M first prompts into the generation model respectively to obtain M initial chapter contents; Audit each of the initial chapter contents by N large language models to obtain an audit result, comprising: obtaining a second prompt word, and auditing each of the initial chapter contents by the auditing model based on the second prompt word to obtain the audit result, wherein the second prompt word is used to audit whether each of the initial chapter contents is accurate.

5. The project report generating method of claim 2, wherein, Before generating chapter content associated with each chapter by N large language models based on the target management data to obtain M initial chapter contents, further comprising: Obtaining a model list, wherein the model list includes model identifiers of M large language models, M being a positive integer greater than N; Determine N large language models based on the model list using a preset selection strategy, wherein the preset selection strategy includes random selection.

6. The project report generating method of claim 2, wherein, Based on the M initial chapter contents and the audit results, the project analysis report is obtained, including: In the case that the audit result indicates that M initial chapter contents are all passed, the M initial chapter contents and the generation process of each initial chapter content are integrated to obtain the project analysis report; In the case that the audit result indicates that any initial chapter content fails the audit, the initial chapter content that fails the audit is regenerated to obtain a target chapter content, and based on the target chapter content and the initial chapter content that passes the audit in the audit result, the project analysis report is obtained.

7. The project report generation method of claim 6, wherein, In the case that the audit result indicates that any initial chapter content fails the audit, the initial chapter content that fails the audit is regenerated to obtain a target chapter content, including: Obtain the audit reason and the number of retries of the initial chapter content that fails the audit, wherein the audit reason includes the reason for failing the audit, and the number of retries includes the number of times the initial chapter content that fails the audit is regenerated; For any initial chapter content that fails the audit, if the number of retries does not reach a preset number threshold, adjust the first prompt word associated with the initial chapter content based on the audit reason to obtain an adjusted first prompt word, regenerate the initial chapter content based on the adjusted first prompt word using N large language models, and audit the regenerated initial chapter content, and if the regenerated initial chapter content passes the audit, determine the regenerated initial chapter content as the target chapter content; If the number of retries reaches the preset number threshold, add a warning mark to the initial chapter content that fails the audit for the last time to obtain the target chapter content.

8. The project report generation method of claim 1, wherein Before obtaining the target management data by acquiring the project management data snapshot conforming to the time range based on the report generation request, further comprising: receiving the project management data snapshot of the target project sent by the project management tool through a first data interface, formatting the project management data snapshot to obtain a formatted project management data snapshot, and sending the formatted project management data snapshot to a target storage space through a second data interface, wherein the formatted project management data snapshot includes the project identifier and the generation time of the data snapshot; Based on the report generation request, obtaining the target management data by acquiring the project management data snapshot conforming to the time range, including: based on the project identifier, filtering the project management data snapshot conforming to the time range from the target storage space to obtain the target management data.

9. A project report generating apparatus characterized by comprising: Including: The request processing module is configured to receive a report generation request of a target project, wherein the report generation request comprises at least one of the following: a time range covered by a project analysis report to be generated, and a project identifier of the target project. The data retrieval module is configured to obtain a project management data snapshot conforming to the time range based on the report generation request, to obtain target management data, wherein the project management data snapshot comprises a data snapshot of project management data of the target project pushed by a project management tool. The report generation module is configured to generate a project analysis report of the target project based on the target management data by using a model consensus strategy, wherein the model consensus strategy comprises generating an initial analysis report of the target project based on N large language models and auditing the initial analysis report, wherein N is a positive integer.

10. An electronic device, comprising: The one or more processors and the memory are configured to implement the project report generation method according to any one of claims 1 to 8 when the one or more programs are executed by the one or more processors.

Images