An adaptive budget report generation method based on multi-agent

Abstract

The present invention discloses an adaptive budget report generation method based on multiple agents, which belongs to the technical field of report generation, and comprises: according to the budget sample collected, using a sample parsing agent, obtaining the sample structure based on image recognition, identifying the basic indicator fields, performing semantic matching based on the basic indicator fields, and obtaining the parsing logic; according to the parsing logic, collecting basic data through a data collection agent, and assigning parsing tasks to each professional agent through a task assignment agent to parse the basic data; generating a budget report through a large model based on the basic data and the parsing results. The present invention realizes template-free dynamic sample parsing, automatically extracts structures and fields, and reduces manual configuration costs through the image recognition technology and semantic matching of the sample parsing agent. The data collection agent integrates multi-source heterogeneous data, and the task assignment agent dynamically schedules professional agents to divide the work and cooperate, thereby improving parsing accuracy and efficiency.

Description

Technical Field

[0001] The present invention belongs to the technical field of report generation, and in particular relates to a multi-agent based adaptive budget report generation method. Background Art

[0002] Against the backdrop of the rapid development of the digital economy, businesses are placing higher demands on the accuracy and efficiency of budget reports. As a crucial basis for corporate strategic decision-making, budget report generation technology has evolved from manual operations to automated tools and finally to intelligent systems driven by artificial intelligence.

[0003] In the early days, budget report generation relied primarily on manual data collection, organization, and analysis. This cumbersome and error-prone process struggled to meet rapidly evolving market demands. With the development of information technology, some automated tools have emerged, but these tools are often based on pre-set templates, populating them with fixed-format data to generate reports.

[0004] In recent years, artificial intelligence technologies (such as machine learning and natural language processing) have gradually been introduced into this field. Some systems have attempted to use algorithms to conduct preliminary analysis and integration of budget data, but they are still limited to the framework of fixed templates and lack the ability to intelligently recognize dynamically changing budget forms and indicators.

[0005] Specifically, existing budget report generation tools generally rely on pre-set templates, requiring users to manually adjust them to accommodate customized forms (e.g., adding new fields, merging cells, and cross-page associations). Whenever the form structure or indicators change, templates must be repeatedly modified and numerous rules manually configured, resulting in cumbersome and inefficient operations.

[0006] Furthermore, enterprise budget data is often scattered across heterogeneous sources such as ERP systems, databases, and unstructured files (such as PDFs and Excel images). Traditional tools lack intelligent parsing capabilities, and data formats and calibers vary across departments, making it difficult to effectively integrate multi-source data. This makes budget reports difficult to generate accurately or slow to generate.

[0007] Therefore, existing technologies generally have problems such as reliance on fixed templates, difficulty in cross-source data integration, and delayed timeliness. It is urgent to develop an adaptive budget report generation method based on multi-agents that can perform template-free dynamic analysis and intelligent integration of multimodal data, so as to achieve the purpose of adaptive generation of budget reports. Summary of the Invention

[0008] The present invention provides an adaptive budget report generation method based on multi-agents. Through dynamic analysis of budget forms and intelligent integration of multimodal data, it solves the problems of existing technologies such as reliance on fixed templates, difficulty in cross-source data integration, and delayed timeliness, and constructs a fully automated and intelligent budget report generation method.

[0009] The technical solution adopted in the present invention is:

[0010] A multi-agent based adaptive budget report generation method comprising:

[0011] Based on the collected budget form, the form parsing agent obtains the form structure based on image recognition, identifies the basic indicator fields, performs semantic matching based on the basic indicator fields, and obtains the parsing logic;

[0012] According to the analysis logic, the data collection agent collects basic data, and the task assignment agent assigns analysis tasks to each professional agent to analyze the basic data;

[0013] Based on the basic data and analysis results, a budget report is generated through the large model.

[0014] The multi-agent-based adaptive budget report generation method provided in the present invention also has the following additional technical features:

[0015] The table structure is obtained based on image recognition, specifically:

[0016] According to the budget table sample, obtaining the row and column structure of the budget table sample through image recognition;

[0017] Obtaining a header field according to the field information filled in the row and column structure;

[0018] According to the header structure, the topological relationship between the header fields is analyzed and a header hierarchy tree is constructed to obtain the mapping relationship between the header fields.

[0019] Identify the basic indicator fields, specifically:

[0020] In response to the budget report indicator input, identifying the basic indicator field based on the correlation between the budget report indicator and the header field through semantic recognition; and / or,

[0021] According to the budget report history database, the basic indicator fields are identified by searching the header fields and according to the frequency of occurrence of the header fields.

[0022] Perform semantic matching based on the basic indicator fields to obtain the parsing logic, specifically:

[0023] Based on the basic indicator fields, natural language processing technology is used to analyze the business meaning of the basic indicator fields;

[0024] Based on the business meaning, the budget indicator knowledge base is called to obtain multiple parsing logics;

[0025] The parsing logic is screened and determined according to the matching degree between the parsing logic and the header field.

[0026] Basic data is collected through the data collection agent, specifically:

[0027] According to the header hierarchy tree, basic data is collected sequentially from bottom to top according to the levels of the header fields in the header hierarchy tree.

[0028] The task assignment agent assigns parsing tasks to each professional agent, specifically:

[0029] The analysis task includes at least one of a basic calculation task and a data comparison and analysis task;

[0030] Based on the basic calculation tasks, the financial agent performs basic data calculations to obtain basic indicators.

[0031] According to the data comparison and analysis task, the basic indicators are compared through the operation intelligent agent to obtain a comparison chart.

[0032] Assigning parsing tasks to professional agents through the task assignment agent also includes:

[0033] By allocating tasks to agents, we monitor the task execution status, load balance, and resource consumption of each professional agent to obtain a state vector.

[0034] According to the state vector, abnormal events are identified. When the task execution state is greater than 5 seconds, it is judged as a processing timeout, triggering computing power expansion or task reallocation.

[0035] When the load balance degree is greater than 0.3, it is judged as load imbalance and task redistribution is triggered.

[0036] When resource consumption exceeds 90%, it is judged as insufficient resources and computing power expansion is triggered.

[0037] The training of the large model is specifically as follows:

[0038] Based on the generated budget report, a quality score is obtained by evaluating the logical coherence, analytical depth, and indicator calculation accuracy;

[0039] Obtain user feedback scores through user explicit evaluation scores and user implicit behavior scores;

[0040] Obtaining a processing efficiency score based on the budget report generation speed and resource consumption;

[0041] A reward function is set based on the weighted quality score, user feedback score, and processing efficiency score to guide the training direction of the large model.

[0042] The present invention also provides a storage medium,

[0043] The storage medium stores a computer program, which, when executed, implements the steps of the multi-agent-based adaptive budget report generation method.

[0044] The present invention again provides a processing device, comprising:

[0045] Memory for storing computer programs;

[0046] A processor is used to implement the steps of the multi-agent based adaptive budget report generation method when executing the computer program.

[0047] Due to the adoption of the above technical solution, the beneficial effects achieved by the present invention are as follows:

[0048] 1. This invention utilizes image recognition technology within a table parsing agent to automatically extract the row and column structure and basic indicator fields of budget tables. It then generates parsing logic through semantic matching, eliminating the need for pre-set templates. This intelligent parsing of custom tables addresses the inability of traditional fixed templates to adapt to dynamic table structures. Automatically identifying new table fields and structures eliminates the need for manual template adjustments, adapting to changing business scenarios and reducing the time and effort required to manually configure rules. This results in automated and intelligent table parsing.

[0049] Furthermore, semantic matching automatically generates parsing logic, reducing manual rule configuration. Users no longer need to master complex rules; they can simply upload a sample form to generate a report, lowering the barrier to entry.

[0050] The data collection agent extracts fundamental data from heterogeneous, multi-source data. The task allocation agent dynamically assigns parsing tasks to specialized agents, who then perform specialized parsing for different task types. Cross-source data integration overcomes the traditional tools' reliance on a single data source and enables standardized integration of multi-source data. The division of labor and collaboration among specialized agents not only enables specialized processing and improves data parsing accuracy, but also increases task processing efficiency through task division, thereby improving the efficiency of budget report generation.

[0051] The large model generates logically rigorous text content based on underlying data and analysis results. The entire process, from table parsing to report generation, is collaboratively completed by intelligent agents, eliminating the need for human intervention. This automated process reduces manual effort, shortens the budget report generation cycle, and improves efficiency. This method can quickly respond to changes in table structure or adjustments to business needs, dynamically updating report content to ensure timeliness. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0053] Figure 1 The figure is a flow chart of the multi-agent-based adaptive budget report generation method according to one embodiment of the present invention. DETAILED DESCRIPTION

[0054] In order to more clearly illustrate the overall concept of the present invention, a detailed description is given below in an exemplary manner in conjunction with the accompanying drawings.

[0055] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0056] like Figure 1 As shown, a multi-agent-based adaptive budget report generation method includes:

[0057] S100: Based on the collected budget sample, a sample parsing agent obtains the sample structure based on image recognition, identifies basic indicator fields, performs semantic matching based on the basic indicator fields, and obtains parsing logic.

[0058] The core purpose of this step is to achieve template-free dynamic parsing of the budget table structure, automatically identify the row and column structure and basic indicator fields of the customized budget table through intelligent means, and generate parsing logic based on semantic matching, thereby solving the problem of traditional fixed template dependence and inability to adapt to dynamic table structures.

[0059] This step uses OCR (optical character recognition) technology combined with computer vision algorithms (such as deep learning models) to scan the budget table sample and extract features such as its row and column structure, field names, merge cell rules, and the topological relationship of multi-level headers (such as the nested relationship between the main header and sub-headers).

[0060] Image recognition technology can accurately extract complex table structures (such as cross-page associations and multi-level table headers), reduce manual intervention, and improve parsing efficiency and accuracy.

[0061] It's understandable that a table template typically includes a header and content section, which can be distinguished by field names. Furthermore, header fields may contain redundant information. To ensure a more accurate budget report, at least some header fields are selected as foundational indicator fields. This facilitates subsequent parsing and report generation based on these foundational indicator fields. This also reduces the amount of parsing required, thereby speeding up report generation.

[0062] Identify basic indicator fields and perform semantic matching. Based on natural language processing (NLP) technology, perform semantic analysis on table header fields, extracting their business meaning (e.g., "budget amount" corresponds to the "financial budget" category) based on field names, context, and historical data (such as field frequency). The budget indicator knowledge base is then invoked to match the identified fields with fields in the predefined indicator library to generate parsing logic.

[0063] Semantic matching, combined with a budget indicator knowledge base, improves field parsing accuracy and enhances the efficiency of generating analysis logic. Furthermore, this method responds in real time to changes in table structure (such as added fields and row and column adjustments), dynamically updating parsing logic and avoiding duplication caused by template changes.

[0064] Moreover, users do not need to master complex rules. They only need to upload sample tables to automatically generate parsing logic, which lowers the usage threshold, reduces the need for manual rule configuration, and reduces system maintenance costs.

[0065] In summary, this step achieves template-free intelligent parsing of budget forms by integrating image recognition and semantic matching technologies, providing efficient and flexible basic support for subsequent budget report generation.

[0066] S200: According to the analysis logic, basic data is collected by the data collection agent, and analysis tasks are assigned to professional agents by the task assignment agent to analyze the basic data.

[0067] The core purpose of this step is to achieve efficient integration and professional analysis of multi-source data. Through the collaboration of data collection agents and task allocation agents, the problems of difficult cross-source data integration, low analysis efficiency and insufficient professional capabilities are solved.

[0068] The data collection agent connects to the ERP system, database and unstructured files (such as PDF / Excel images) through the adapter mode, automatically generates customized data reading protocols, realizes multi-source data access, and provides unified input for subsequent analysis.

[0069] It should be noted that data cleansing and alignment are also required for accessing multi-source data, and this is not a limitation of the present invention. A data cleansing rules engine is used to remove null and duplicate values ​​and perform data format conversion (e.g., standardizing date formats and units). A budget indicator knowledge base is used to align the field calibers of different data sources (e.g., "budget amount" is standardized to "10,000 yuan").

[0070] The data collection agent improves data integration efficiency by automatically generating cleaning rules and ensures input consistency for subsequent parsing tasks.

[0071] Using task-assignment agents, the budget report generation task is broken down into subtasks (such as basic calculations and data comparison and analysis). A task allocation table is generated, clarifying the task content, priorities, and dependencies of each agent. A cluster of agents is activated to process tasks in parallel, forming a closed-loop feedback mechanism. Each specialized agent executes tasks based on domain knowledge, improving the accuracy of basic indicator calculations.

[0072] In summary, this step achieves efficient integration and professional analysis of multi-source data through the collaboration of data collection and task allocation agents, providing stable and efficient underlying support for budget report generation.

[0073] S300: Generate a budget report through a large model based on the basic data and analysis results.

[0074] The core purpose of this step is to generate high-quality, personalized budget reports through large models to solve the problems of delayed budget reports and inability to meet the needs of multiple roles in traditional technologies.

[0075] Through the large model, basic data (such as parsed budget indicators) and parsing results (such as analysis logic trees) are integrated to build a unified data input framework.

[0076] Based on the analysis results, logically rigorous text content (such as data analysis, conclusion deduction, and improvement suggestions) is generated to ensure the professionalism of the generated content. Furthermore, based on the underlying data, data in the budget report is populated to enhance its credibility. In this way, based on the underlying data and analysis results, a logically rigorous and content-rich budget report is generated, achieving an organic integration of data and conclusions.

[0077] It should be noted that for the analysis results, corresponding text content can be generated, and basic chart forms such as bar charts and line charts can also be generated to achieve an intuitive display of the budget report. The present invention does not impose any restrictions on this.

[0078] In this step, the large model automatically generates text and charts, reducing manual intervention and improving the efficiency of budget report generation.

[0079] In summary, this step achieves the logical rigor, content depth, and multi-role adaptation of the budget report through the multimodal generation of the large model, providing an intelligent and adaptive solution for budget management.

[0080] As a preferred embodiment of the present invention, a table-like structure is obtained based on image recognition, specifically:

[0081] According to the budget table sample, obtaining the row and column structure of the budget table sample through image recognition;

[0082] Obtaining a header field according to the field information filled in the row and column structure;

[0083] According to the header structure, the topological relationship between the header fields is analyzed and a header hierarchy tree is constructed to obtain the mapping relationship between the header fields.

[0084] The core purpose of this implementation is to achieve template-free dynamic parsing of budget tables. Through image recognition and topological relationship analysis technology, the table structure and field information are automatically extracted, and a header hierarchy tree is constructed to solve the problem of traditional fixed template dependence and inability to adapt to customized table structures.

[0085] Utilize deep learning OCR models (such as the Transformer architecture) to scan budget spreadsheets and accurately identify row and column structure and field locations. Image matching and coordinate mapping techniques address cross-page spreadsheet structural integrity issues (e.g., merged cells extending across pages). This allows for the extraction of table row and column counts and field distribution, generating structured spreadsheet metadata.

[0086] Extract field names and locations from table templates, and extract header fields. Note that this requires combining text location algorithms with visual feature analysis (such as font bolding and alignment) to distinguish header fields from annotations, eliminating annotations and obtaining the complete header fields. Accurately identifying complex table structures improves parsing efficiency and accuracy, reducing manual intervention.

[0087] Graph Neural Networks (GNNs) are used to analyze the topological relationships between header fields (e.g., parent-child hierarchies and cross-column dependencies) and construct a header hierarchy tree. Furthermore, logical validation rules are supplemented by incorporating the characteristics of industry budget tables (e.g., the "department budget - project budget - detailed item" hierarchy).

[0088] This step clarifies the mapping relationships between fields by generating a table header hierarchy tree. Using image recognition and topology analysis technology, table structures can be dynamically parsed without templates, adapting to complex business scenarios (such as multi-level headers and cross-page associations), providing a unified input framework for subsequent multi-source data integration.

[0089] Users don't need to master complex rules; simply upload a table sample to automatically generate parsing logic, lowering the barrier to entry. Table sample parsing responds to changes in the table structure (such as added fields and row and column adjustments) in real time, dynamically updating the parsing logic. This eliminates the need for manual rule configuration due to template changes and reduces system maintenance costs.

[0090] In summary, this embodiment achieves template-free intelligent analysis of budget tables by integrating image recognition, topology analysis and dynamic strategy optimization technology.

[0091] As an example of this implementation, the basic indicator fields are identified as follows:

[0092] In response to the budget report indicator input, identifying the basic indicator field based on the correlation between the budget report indicator and the header field through semantic recognition; and / or,

[0093] According to the budget report history database, the basic indicator fields are identified by searching the header fields and according to the frequency of occurrence of the header fields.

[0094] The core purpose of this embodiment is to realize the intelligent recognition of basic indicator fields in budget forms. Through semantic matching and historical data retrieval technology, basic indicator fields are identified for budget report generation to solve the problem of redundant report parameters.

[0095] Example 1: In response to budget report indicator input, natural language processing (NLP) is used to calculate the semantic relevance between budget report indicators and header fields using semantic similarity algorithms (such as cosine similarity and BERT embedding). Furthermore, the business meaning of the fields is extracted (e.g., "budget amount" corresponds to "financial category" indicators) by integrating with the budget indicator knowledge base, enhancing the accuracy of semantic matching.

[0096] Specifically, based on the input budget report indicator (such as "Total Department Budget"), the NLP model matches the header fields (such as "Budget Amount" and "Department Name") and calculates a relevance score. Based on the score, the highly matching fields are selected and used as the basic indicator fields (such as "Budget Amount").

[0097] Example 2: Historical Data Retrieval and Frequency Analysis. Using vector retrieval, header fields are encoded as vectors to construct a vector index for the budget report history database. Frequency statistics are performed to analyze the frequency of occurrence of header fields in the historical database and identify high-frequency fields (e.g., "Budget Amount" and "Execution Progress").

[0098] Specifically, search for header fields in historical budget reports and count their occurrences (e.g., "Budget Amount" appears 95% of the time). Combine this frequency with business importance (e.g., "Budget Amount" is a key indicator) to select high-frequency fields as foundational indicator fields. Ensure that key indicators (e.g., "Budget Amount") are prioritized to reduce the risk of omissions.

[0099] It should be noted that the technical means in the above two embodiments can be used in conjunction. Set a weighted fusion strategy to weight the semantic relevance score and the historical frequency statistics to generate a comprehensive score, for example, Score = a × semantic score + 𝛽 × frequency score. Through the weighted fusion strategy, semantic understanding and historical data are balanced to improve recognition robustness.

[0100] This embodiment uses semantic understanding and historical data mining technology to achieve intelligent identification of basic indicator fields in budget forms, screen out fields that are strongly related to budget report indicators, and provide efficient and accurate data support for subsequent budget report generation.

[0101] Specifically, semantic matching is performed based on the basic indicator fields to obtain parsing logic, which is specifically:

[0102] Based on the basic indicator fields, natural language processing technology is used to analyze the business meaning of the basic indicator fields;

[0103] Based on the business meaning, the budget indicator knowledge base is called to obtain multiple parsing logics;

[0104] The parsing logic is screened and determined according to the matching degree between the parsing logic and the header field.

[0105] The core purpose of this embodiment is to achieve dynamic semantic parsing and logic generation of basic indicator fields through the collaboration of natural language processing (NLP) and knowledge base, and solve the problem of traditional fixed template dependence and inability to adapt to customized table indicators.

[0106] Natural language processing parses business meaning. Semantic embedding models, using pre-trained models such as BERT and Sentence-BERT, encode basic indicator fields (such as "budget amount") into high-dimensional vectors, capturing their business meaning (such as "financial indicators"). Context-aware analysis, combining the field's header hierarchy and adjacent fields (such as "department name" and "project number"), enhances the accuracy of semantic parsing.

[0107] Specifically, input the basic indicator field (such as "new energy project budget") and generate semantic vectors through the NLP model , extract the business attributes of the field (such as "energy industry" and "budget category") and build multi-dimensional business labels .

[0108] Call the budget indicator knowledge base to obtain candidate logic. Knowledge base retrieval, based on the budget indicator knowledge base (such as industry standard indicator rules, historical parsing logic), matches the field semantic vector through vector similarity (cosine similarity). , generate a candidate parsing logic set . Execute rule extension, combined with field business tags , calling association rules (such as "energy industry budget must include carbon emission costs") to supplement candidate logic.

[0109] Specifically, for the field "new energy project budget," the knowledge base is searched for parsing rules related to "new energy" and "budget" (e.g., "budget = cost + taxes + profit"). This generates candidate logic L = {"budget = cost + taxes", "budget = infrastructure budget + operating budget", ...}.

[0110] Calculate the matching degree to filter logic, calculate the semantic similarity, and use cosine similarity , quantify the matching degree between the field and the candidate logic. Combined with the historical analysis records of the field (for example, "budget amount" often matches "department budget sum"), dynamically adjust the matching weight . Comprehensive semantic similarity, business tag matching, and historical usage frequency are combined to generate a comprehensive score.

[0111] .

[0112] Based on the comprehensive score of each candidate logic , select the highest scoring logic As the final parsing logic, it should be noted that if the matching degree is lower than the threshold, the fuzzy matching mechanism (such as manual rule backup) is triggered.

[0113] This example uses natural language processing and a knowledge base to dynamically analyze and generate logic for budget form indicator fields. Based on the matching between the fields and the parsing logic, the optimal parsing logic is selected to ensure the accuracy and consistency of the generated results.

[0114] As a preferred embodiment of the present invention, basic data is collected by a data collection agent, specifically:

[0115] According to the header hierarchy tree, basic data is collected sequentially from bottom to top according to the levels of the header fields in the header hierarchy tree.

[0116] The core purpose of this implementation is to achieve bottom-up hierarchical collection and standardized integration of basic data, guide the data collection sequence through the hierarchical relationship of the header hierarchy tree, and solve the problems of chaotic data collection and repeated collection.

[0117] Identify the parent-child dependency relationships between fields (for example, "total budget = sum of department budgets"), prioritize collecting bottom-level fields (department budgets), and then gradually aggregate to the top-level field (total budget).

[0118] Specifically, input the header hierarchy tree , traverse each node , triggering data collection tasks sequentially from the lowest level to the highest level. Bottom-up layered collection ensures that the underlying data is processed first, reducing errors caused by field dependencies (such as the need for "total budget" to be calculated based on "department budget") and ensuring data logical consistency.

[0119] Dynamically adjust collection strategies (e.g., prioritize frequently used fields) based on historical data collection efficiency (e.g., field collection time and error rate). Automatically switch to a backup data source (e.g., database snapshot) when a field collection times out (e.g., ERP interface response time > 5 seconds).

[0120] Monitor data collection status (e.g., timeout, missing fields), triggering retries or manual verification. If a field is missing (e.g., "Carbon Emission Budget" is not filled in), a manual verification process is triggered to avoid data chain breaks.

[0121] In summary, this embodiment achieves efficient and standardized collection of budget data through hierarchical driven collection.

[0122] As a preferred embodiment of the present invention, the task assignment agent assigns parsing tasks to each professional agent, specifically:

[0123] The analysis task includes at least one of a basic calculation task and a data comparison and analysis task;

[0124] Based on the basic calculation tasks, the financial agent performs basic data calculations to obtain basic indicators.

[0125] According to the data comparison and analysis task, the basic indicators are compared through the operation intelligent agent to obtain a comparison chart.

[0126] The core purpose of this implementation is to achieve professional analysis and efficient scheduling of analytical tasks. Through the task allocation agent, basic computing tasks and data comparison and analysis tasks are assigned to the financial agent and the operation agent respectively, solving the problems of insufficient analytical capabilities and slow analytical efficiency.

[0127] Based on the parsing logic tree generated from the header hierarchy tree and semantic matching results, the budget report generation task is broken down into subtasks (such as basic calculation tasks and data comparison and analysis tasks). Based on the task type (calculation / comparison), ensure that the task matches the agent's professional capabilities.

[0128] Specifically, tasks such as "Budget amount = sum of departmental budgets" are classified as basic calculation tasks and assigned to the Finance agent. Tasks such as "actual expenditure vs. budget trend comparison" are classified as data comparison and analysis tasks and assigned to the Operations agent. The Finance agent focuses on calculations, while the Operations agent focuses on comparison and analysis, improving task processing accuracy.

[0129] The financial agent calls the standardized data provided by the data collection agent (such as ERP system data and database records) and executes calculation formulas (such as "total budget = sum of department budgets").

[0130] Through the operational agent, historical data (such as "Budget for the same period last year" and "Industry Average Budget") is retrieved and comparative calculations are performed (for example, "Year-on-year Growth Rate = (Current Budget - Previous Budget) / Previous Budget"). Chart types (such as line charts and pie charts) are matched based on data characteristics (such as trends and percentages) to achieve dynamic visualization output.

[0131] This implementation method adapts to dynamic business scenarios through the dynamic disassembly and professional division of labor of task allocation agents, realizes standardized integration and in-depth analysis of multi-source data, and achieves professional analysis and efficient scheduling of budget report generation tasks.

[0132] As an example of this implementation, allocating parsing tasks to professional agents through a task allocation agent also includes:

[0133] By allocating tasks to agents, we monitor the task execution status, load balance, and resource consumption of each professional agent to obtain a state vector.

[0134] According to the state vector, abnormal events are identified. When the task execution state is greater than 5 seconds, it is judged as a processing timeout, triggering computing power expansion or task reallocation.

[0135] When the load balance degree is greater than 0.3, it is judged as load imbalance and task redistribution is triggered.

[0136] When resource consumption exceeds 90%, it is judged as insufficient resources and computing power expansion is triggered.

[0137] The core purpose of this embodiment is to realize dynamic monitoring and abnormal response of multi-agent systems. The task allocation agent monitors the task execution status, load balance and resource consumption of each professional agent in real time, constructs state vectors and identifies abnormal events, thereby triggering computing power expansion or task redistribution, and solving the problems of task timeout, load imbalance and resource shortage in the traditional fixed resource allocation mode.

[0138] The task allocation agent collects three core indicators of each professional agent in real time (task execution time, load balance, and resource consumption) and generates a state vector The state vector is updated in real time through polling or event-driven mechanisms (such as Prometheus monitoring alerts) to ensure the timeliness of monitoring data.

[0139] Based on industry experience and historical data, set abnormal event trigger thresholds (such as execution time > 5 seconds, load balancing > 0.3, resource consumption > 90%).

[0140] When an agent's task execution time exceeds 5 seconds (e.g., the "Cross-department budget summary" task takes 6.2 seconds), a timeout event is triggered. When the load balance exceeds 0.3 (e.g., the load difference between the operations agent and the finance agent is 0.35), a load imbalance event is triggered. When resource consumption exceeds 90% (e.g., GPU computing power usage exceeds 92%), a resource shortage event is triggered.

[0141] For timeout events, priority is given to expanding computing power (such as starting backup GPU nodes). If this fails, task redistribution is triggered. For load imbalance events, the task allocation strategy is optimized through reinforcement learning models (such as adjusting task priorities and reducing the number of tasks for high-load agents).

[0142] Computing capacity expansion uses containerization technologies (such as Kubernetes) to dynamically expand resources (e.g., adding GPU nodes) to alleviate resource shortages. Task redistribution uses dynamic load balancing algorithms (e.g., the minimum number of connections strategy) to migrate tasks to less-loaded agents (e.g., migrating the "budget trend analysis" task from a highly loaded operational agent to an idle agent).

[0143] This embodiment shortens task completion time by expanding computing power and redistributing tasks. The dynamic expansion strategy improves GPU computing power utilization, resolving resource shortages. The task redistribution strategy reduces load balancing, corrects load imbalances, and avoids single points of failure.

[0144] In summary, this embodiment achieves efficient and stable operation of the multi-agent system through multi-dimensional monitoring and dynamic response mechanisms.

[0145] As a preferred embodiment of the present invention, the training of the large model is specifically as follows:

[0146] Based on the generated budget report, a quality score is obtained by evaluating the logical coherence, analytical depth, and indicator calculation accuracy;

[0147] Obtain user feedback scores through user explicit evaluation scores and user implicit behavior scores;

[0148] Obtaining a processing efficiency score based on the budget report generation speed and resource consumption;

[0149] A reward function is set based on the weighted quality score, user feedback score, and processing efficiency score to guide the training direction of the large model.

[0150] The core purpose of this implementation is to optimize the training direction and output quality of large models through multi-dimensional evaluation and reward function design, and to address the limitations of traditional training methods that rely on a single metric (such as accuracy) and ignore user feedback and processing efficiency.

[0151] In the quality scoring process, NLP models (such as BERT) are used to examine the budget report's paragraph cohesion and semantic consistency (e.g., "is the budget amount logically related to the execution progress") to assess logical coherence. The depth of the comparative analysis in the report (e.g., "does the year-on-year growth analysis include the industry average") is verified and assessed using the budget indicator knowledge base. The generated results are compared with the standard answer (e.g., "is the total budget equal to the sum of departmental budgets") to calculate the accuracy of the indicator calculation.

[0152] In user feedback scoring, we collect user ratings (e.g., a 5-point scale) and labels (e.g., "content is incomplete") for budget reports to generate explicit user evaluation scores. We also quantify satisfaction through user behavior analysis (e.g., report viewing time, number of revisions, and download rate) (e.g., viewing time <30 seconds is considered negative feedback) to generate implicit user behavior scores.

[0153] The processing efficiency score measures the time it takes to generate a single report (average time < 5 seconds is considered excellent) to assess generation speed. GPU memory usage, CPU utilization, and energy consumption are monitored (memory usage < 8GB is considered efficient) to assess resource consumption.

[0154] Based on business priorities, quality scores (weights a ), user feedback rating (weight β ), processing efficiency score (weight γ ) are weighted to get the final reward .

[0155] This implementation addresses the comprehensive needs of business scenarios (e.g., the need for both accuracy and speed in financial reporting) through multi-dimensional assessment, overcoming the limitations of traditional single metrics. Furthermore, by processing efficiency scores and monitoring resource consumption, it adapts to high-load scenarios (e.g., elastic allocation of cloud service resources) and addresses the challenges of enterprise-level scenarios.

[0156] In summary, this implementation achieves comprehensive optimization of large-scale model training through multi-dimensional evaluation and reward function design, providing a high-quality, high-efficiency, and high-user-satisfaction solution for budget report generation.

[0157] The present invention also provides a storage medium,

[0158] The storage medium stores a computer program, which, when executed, implements the steps of the multi-agent-based adaptive budget report generation method.

[0159] Therefore, any effect of the multi-agent-based adaptive budget report generation method can be achieved, which will not be elaborated here.

[0160] The present invention again provides a processing device, comprising:

[0161] Memory for storing computer programs;

[0162] A processor is used to implement the steps of the multi-agent based adaptive budget report generation method when executing the computer program.

[0163] Therefore, any effect of the multi-agent-based adaptive budget report generation method can be achieved, which will not be elaborated here.

[0164] Anything not described in the present invention can be achieved by adopting or drawing on existing technologies.

[0165] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

[0166] The foregoing is merely an embodiment of the present invention and is not intended to limit the present invention. It will be apparent to those skilled in the art that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims of the present invention.

Claims

1. A multi-agent based adaptive budget report generation method, characterized in that: include: According to the collected budget table sample, the table sample parsing agent obtains the row and column structure of the budget table sample based on image recognition; Obtaining a header field according to the field information filled in the row and column structure; According to the header structure, the topological relationship between the header fields is analyzed and a header hierarchy tree is constructed to obtain the mapping relationship between the header fields and obtain the table sample structure; In response to input of a budget report indicator, identifying a basic indicator field based on a correlation between the budget report indicator and a header field through semantic recognition; and / or identifying a basic indicator field based on a frequency of occurrence of the header field by searching a budget report history database; Perform semantic matching based on the basic indicator fields and use natural language processing technology to analyze the business meaning of the basic indicator fields; Based on the business meaning, the budget indicator knowledge base is called to obtain multiple parsing logics; Filter and determine the parsing logic according to the matching degree between the parsing logic and the header field; According to the analysis logic, the data collection agent collects basic data, and the task assignment agent assigns analysis tasks to each professional agent to analyze the basic data; Based on the basic data and analysis results, a budget report is generated through the large model.

2. The method for generating a multi-agent-based adaptive budget report according to claim 1, characterized in that: Basic data is collected through the data collection agent, specifically: According to the header hierarchy tree, basic data is collected sequentially from bottom to top according to the levels of the header fields in the header hierarchy tree.

3. The method for generating a multi-agent-based adaptive budget report according to claim 1, characterized in that: The task assignment agent assigns parsing tasks to each professional agent, specifically: The analysis task includes at least one of a basic calculation task and a data comparison and analysis task; Based on the basic calculation tasks, the financial agent performs basic data calculations to obtain basic indicators. According to the data comparison and analysis task, the basic indicators are compared through the operation intelligent agent to obtain a comparison chart.

4. The method for generating a multi-agent-based adaptive budget report according to claim 3, characterized in that: Assigning parsing tasks to professional agents through the task assignment agent also includes: By allocating tasks to agents, we monitor the task execution status, load balance, and resource consumption of each professional agent to obtain a state vector. According to the state vector, abnormal events are identified. When the task execution state is greater than 5 seconds, it is judged as a processing timeout, triggering computing power expansion or task reallocation. When the load balance degree is greater than 0.3, it is judged as load imbalance and task redistribution is triggered. When resource consumption exceeds 90%, it is judged as insufficient resources and computing power expansion is triggered.

5. The method for generating a multi-agent-based adaptive budget report according to claim 1, wherein: The training of the large model is specifically as follows: Based on the generated budget report, a quality score is obtained by evaluating the logical coherence, analytical depth, and indicator calculation accuracy; Obtain user feedback scores through user explicit evaluation scores and user implicit behavior scores; Obtaining a processing efficiency score based on the budget report generation speed and resource consumption; A reward function is set based on the weighted quality score, user feedback score, and processing efficiency score to guide the training direction of the large model.

6. A storage medium, characterized in that The storage medium stores a computer program, which, when executed, implements the steps of the multi-agent-based adaptive budget report generation method according to any one of claims 1 to 5.

7. A processing device, characterized in that: include: memory for storing computer programs; A processor is configured to implement the steps of the multi-agent based adaptive budget report generation method as described in any one of claims 1 to 5 when executing the computer program.

Images