Intelligent report generation system and method based on multi-module cooperation
Through the multi-module collaborative intelligent report generation system, the problems of report generation integrity, data authenticity, format compatibility and output stability are solved, and efficient and automated long report generation is achieved, which improves the integrity and credibility of report generation.
Patent Information
- Application Number
- CN202510835784.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-20
- Publication Date
- 2025-10-17
AI Technical Summary
The existing intelligent report generation system has problems such as insufficient report generation integrity, lack of data authenticity verification, poor format compatibility, low system intelligence and large volatility of large model output.
An intelligent report generation system with multi-module collaboration is adopted, including a multimodal OCR recognition engine, a trusted data purification system, a dynamic evolutionary intelligent agent and a long report generation engine. The multimodal OCR recognition engine is used for format verification, the trusted data purification system is used for multi-source verification and stability constraints, the dynamic evolutionary intelligent agent realizes online rule learning, and the long report generation engine generates complete and coherent long reports.
It achieves complete output of extra-long reports, improves data accuracy and format compatibility, reduces output volatility, improves the intelligence of the system and the repeatability of reports, reduces manual intervention, and improves generation efficiency.
Smart Images
Figure CN120805859A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of report generation, in particular to an intelligent report generation system and method based on multi-module cooperation. BACKGROUND
[0002] The prior art has the following core defects in the field of intelligent report generation:
[0003] 1. Inadequate report generation integrity: traditional AI large models are limited by output length (such as GPT-4 maximum token number is 32k), key information is easily truncated when generating long reports, and multi-module data splicing relies on manual work, with poor format consistency.
[0004] 2. Lack of data authenticity verification: there are false or contradictory information in Internet public data, existing systems lack multi-source cross verification mechanism, local database update lags behind, and cannot dynamically correct errors.
[0005] 3. Poor format compatibility: OCR system has high error rate in recognizing non-standard formats (such as scanned documents tilt, watermark blocking), and lacks data rationality checking, unable to detect abnormal values.
[0006] 4. Low system intelligence: traditional intelligent agents rely on preset rules, cannot learn and optimize analysis logic online, database update requires downtime maintenance, affecting business continuity.
[0007] 5. Large model output volatility: the output result difference rate of the same input multiple requests is high, lacking stability constraint mechanism, with poor report repeatability.
[0008] Therefore, an intelligent report generation system and method based on multi-module cooperation are proposed. SUMMARY
[0009] The present application provides an intelligent report generation system and method based on multi-module cooperation to solve the problems raised in the background art.
[0010] The specific technical solutions are as follows:
[0011] An intelligent report generation system based on multi-module cooperation, comprising: a multi-modal OCR recognition engine, a trusted data purification system, a dynamic evolution intelligent agent and a long report generation engine, wherein:
[0012] The multi-modal OCR recognition engine is used for recognizing and rationality checking of multi-format documents;
[0013] The trusted data purification system is used for multi-source verification and stability constraint of the recognized data;
[0014] The dynamic evolution agent is used to realize online learning and updating of rules.
[0015] The long report generation engine is used to generate a complete and coherent long report.
[0016] The multi-module collaborative intelligent report generation system described above, wherein the multi-modal OCR recognition engine comprises:
[0017] An adversarial generation network (GAN) module is used to synthesize training data covering extreme scenarios such as scanning tilt, watermark occlusion, etc.
[0018] A self-attention segmentation algorithm module is used to dynamically divide text regions and decorative elements.
[0019] An industry common sense graph module is used to compare and identify abnormal data in real time.
[0020] The multi-module collaborative intelligent report generation system described above, wherein the abnormal data labeling highlights the problem area with a red border and generates a correction suggestion.
[0021] The multi-module collaborative intelligent report generation system described above, wherein the trusted data purification system comprises:
[0022] A multi-source verification module cross-verified the Internet data and local database in real time through the agent.
[0023] A trustworthiness calculation module uses the formula for calculation, where α is the local data weight coefficient, with a default value of 0.7, used to adjust the proportion of local database data in trustworthiness calculation; D_local is the trustworthiness index of the corresponding data in the local database; D_web is the trustworthiness index of a single Internet data source; N is the number of Internet data sources participating in cross-validation.
[0024] A stability constraint module applies KL divergence constraint to the large model output and uses deterministic decoding for key fields.
[0025] The multi-module collaborative intelligent report generation system described above, wherein the KL divergence constraint satisfies KL(P output ||P history )<δ, where P_output is the probability distribution of the current AI large model output result; P_history is the probability distribution of the historical effective output result; δ is the preset KL divergence threshold, used to limit the degree of shift of the output distribution.
[0026] The multi-module collaborative intelligent report generation system described above, wherein the dynamic evolution agent comprises:
[0027] A double-buffering mechanism module, the standby database is updated synchronously when the main database provides services;
[0028] An HNSW-based incremental learning algorithm module, supporting millisecond-level retrieval;
[0029] A feedback loop module, automatically triggering fine-tuning for artificial correction data, using reinforcement learning (DQN algorithm) to optimize decision-making strategies, and having a version rollback function.
[0030] The intelligent report generation system based on multi-module cooperation described above, wherein the long report generation engine comprises:
[0031] A memory window mechanism module, inserting a summary paragraph after generating 5000 tokens;
[0032] A dynamic chunking algorithm module, automatically dividing chapters according to semantic relevance (BERT similarity > 0.85);
[0033] An automatic continuation module, triggering a continuation request and inheriting the previous state vector when detecting that the output is truncated;
[0034] A consistency checking module, comparing the logic contradictions between the front and back sections through a twin network.
[0035] The automatic continuation module of the intelligent report generation system based on multi-module cooperation described above triggers a continuation request by detecting the end symbol (such as the mark) of the output sequence or the token count threshold (≥ 4900 tokens), and the inherited previous state vector contains the context representation of the last 10 layers of the Transformer encoder.
[0036] The consistency checking module trains the twin network using TripletLoss, detects contradictions by calculating the Euclidean distance between the semantic vectors of the front and back sections, and triggers the correction process when the distance > 0.7.
[0037] The present application also provides an intelligent report generation method based on multi-module cooperation, which applies the intelligent report generation system based on multi-module cooperation described above, comprising the following steps:
[0038] Recognize and verify the rationality of multi-format documents through a multi-modal OCR recognition engine;
[0039] Input the recognized data into a trusted data purification system for multi-source verification and stability constraints;
[0040] Use a dynamic evolutionary agent to perform online learning and updating of rules;
[0041] Generate a complete and coherent long report through a long report generation engine.
[0042] The intelligent report generation method based on multi-module cooperation, wherein, when generating a long report, a "memory window" mechanism is adopted to insert a summary paragraph after generating 5000 tokens, and a dynamic block algorithm is used to automatically divide chapters according to semantic correlation, and when it is detected that the output is truncated, an automatic continuation function is triggered and consistency verification is performed.
[0043] The present application has the following beneficial effects:
[0044] 1. Solve the problem of report integrity: break through the token limit of large models, realize the complete output of ultra-long reports (>100,000 words), and avoid key information truncation and logical break.
[0045] 2. Improve data reliability: through multi-source verification and reliability calculation, the data accuracy is improved from 83% of the traditional scheme to close to the theoretical extreme value, effectively filtering false or contradictory information.
[0046] 3. Enhance format compatibility: support robust recognition of 200+ document formats, significantly reduce non-standard format recognition error rate, and reduce human error through real-time abnormal data warning mechanism.
[0047] 4. Dynamic evolution capability: the rule iteration period is shortened from the weekly level of the traditional scheme to the minute level, and the system can continuously optimize the analysis logic through online learning to adapt to changes in business scenarios.
[0048] 5. Stable large model output: the output volatility is greatly reduced, the consistency of key data fields (such as amount and date) is significantly improved, and the report repeatability meets the needs of professional scenarios.
[0049] 6. Full-process automation: reduce manual intervention, full-process automatic processing from document recognition to report generation, reduce labor costs, improve output efficiency, and the analysis results are not subject to subjective intervention, more objective. BRIEF DESCRIPTION OF DRAWINGS
[0050] Figure 1 The composition schematic diagram of the intelligent report generation system based on multi-module cooperation provided for the embodiments of the present application;
[0051] Figure 2 The flowchart of the intelligent report generation method based on multi-module cooperation provided for the embodiments of the present application. DETAILED DESCRIPTION
[0052] The technical solutions of the present application will be further described below in conjunction with the drawings and through specific embodiments.
[0053] Among them, the drawings are only used for illustrative description, and the representation is only a schematic diagram, not a physical diagram, and cannot be understood as a limitation of the patent; in order to better illustrate the embodiments of the present application, some components of the drawings will be omitted, enlarged or reduced, and do not represent the size of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings can be omitted.
[0054] The same or similar reference numerals in the drawings of the embodiments of the present application correspond to the same or similar components; in the description of the present application, it should be understood that if the terms "upper", "lower", "left", "right", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore the terms describing the positional relationship in the drawings are only used for illustrative description, and cannot be understood as a limitation of the patent, for those skilled in the art, the specific meaning of the above terms can be understood according to the specific situation.
[0055] In the description of the present application, unless otherwise explicitly specified and limited, if the term "connection" and the like appear to indicate the connection relationship between components, the term should be broadly understood, for example, it can be fixedly connected, or it can be detachably connected, or it can be integrated; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium; it can be the communication or interaction relationship between two components. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific situation.
[0056] Embodiment
[0057] The intelligent report generation system based on multi-module cooperation provided by the embodiment, as shown in Figure 1 , includes a multi-modal OCR recognition engine, a trusted data purification system, a dynamic evolution agent and a long report generation engine, wherein:
[0058] The multi-modal OCR recognition engine is used for recognizing and checking the rationality of multi-format documents;
[0059] The trusted data purification system is used for multi-source verification and stability constraint of the recognized data;
[0060] The dynamic evolution agent is used for realizing online learning and updating of rules;
[0061] The long report generation engine is used for generating a complete and coherent long report.
[0062] By adopting the above technical solution, through the collaboration of multimodal OCR recognition engine, trusted data purification system, dynamic evolution intelligent agent and long report generation engine, the automatic processing of multi-format documents, data verification, rule evolution and long report generation are realized, solving the problems of insufficient report integrity, low data credibility and poor format compatibility in traditional solutions.
[0063] Specifically, in this embodiment, the multimodal OCR recognition engine includes: a generative adversarial network (GAN) module, a self-attention segmentation algorithm module, and an industry knowledge graph module, wherein:
[0064] Generative Adversarial Network (GAN) module, used to synthesize training data covering extreme scenarios such as scan tilt and watermark occlusion;
[0065] Self-attention segmentation algorithm module, used to dynamically divide text areas and decorative elements;
[0066] The industry common sense graph module is used to compare recognition results in real time and mark abnormal data.
[0067] By adopting the above technical solution, the GAN module and self-attention segmentation algorithm are used to improve the OCR's recognition ability for documents in extreme formats. The rationality of the recognition results can be verified by combining the industry common sense graph to reduce the recognition errors of non-standard format documents.
[0068] Specifically, in this embodiment, abnormal data annotation highlights the problem area with a red border and generates correction suggestions. Using the above technical solution, abnormal data is visually annotated through red border highlighting and correction suggestions, improving the accuracy of data processing and the efficiency of manual review.
[0069] Specifically, in this embodiment, the trusted data purification system includes: a multi-source verification module, a credibility calculation module, and a stability constraint module, wherein:
[0070] The multi-source verification module crawls Internet data in real time through the intelligent agent and cross-validates it with the local database;
[0071] The credibility calculation module uses the formula Calculation is performed, where α is the local data weight coefficient, with a default value of 0.7, which is used to adjust the proportion of local database data in the credibility calculation; D_local is the credibility index of the corresponding data in the local database; D_web is the credibility index of a single Internet data source; N is the number of Internet data sources participating in the cross-validation;
[0072] The stability constraint module imposes KL divergence constraints on large model outputs and adopts deterministic decoding for key fields.
[0073] The above technical scheme is adopted, based on multi-source data cross verification and a credibility calculation model, data authenticity is improved; through a stability constraint algorithm, volatility of large model output is reduced, and output consistency of key data fields is ensured.
[0074] Specifically, in the embodiment, the KL divergence constraint satisfies KL(P output ||P history )<δ, wherein P_output is a probability distribution of a current AI large model output result; P_history is a probability distribution of a historical effective output result; and δ is a preset KL divergence threshold, used for limiting a deviation degree of the output distribution.
[0075] The above technical scheme is adopted, through the KL divergence constraint, deviation of the large model output distribution and the historical result is controlled, and stability and repeatability of the output content are further ensured.
[0076] Specifically, in the embodiment, the dynamic evolution agent includes a double buffering mechanism module, an HNSW-based incremental learning algorithm module, and a feedback closed loop module, wherein:
[0077] The double buffering mechanism module synchronously updates the standby library when the main database provides services;
[0078] The HNSW-based incremental learning algorithm module supports millisecond-level retrieval;
[0079] The feedback closed loop module automatically triggers fine tuning based on artificial correction data, adopts reinforcement learning (DQN algorithm) to optimize a decision strategy, and has a version rollback function.
[0080] The above technical scheme is adopted, the double buffering mechanism and the HNSW incremental learning algorithm are adopted, online update and millisecond-level retrieval of the rule database are achieved, the feedback closed loop is combined to optimize the decision strategy of the agent, and dynamic adaptability of the system to data and business scenarios is improved.
[0081] Specifically, in the embodiment, the long report generation engine includes a memory window mechanism module, a dynamic blocking algorithm module, an automatic continuation module, and a consistency verification module, wherein:
[0082] The memory window mechanism module inserts a summary paragraph after generating 5000 tokens each time;
[0083] The dynamic blocking algorithm module automatically divides chapters according to semantic correlation (BERT similarity>0.85);
[0084] The automatic continuation module triggers a continuation request and inherits a previous state vector when detecting that the output is truncated;
[0085] The consistency verification module compares logic contradictions between front and rear paragraphs through a twin network.
[0086] By using the technical solution, the memory window mechanism, the dynamic block algorithm and the automatic continuation function are adopted to break through the token limit of the large model, guarantee the coherence of the long report generation, and use the twin network to check logical contradictions to improve the logicality of the report content.
[0087] Specifically, in the embodiment, the automatic continuation module triggers the continuation request by detecting the termination symbol (such as the mark) of the output sequence or the token count threshold (≥4900 tokens), and the inherited previous context vector contains the context representation of the last 10 layers of the Transformer encoder.
[0088] The consistency checking module trains the twin network by using TripletLoss, detects contradictions by calculating the Euclidean distance between the semantic vectors of the front and rear paragraphs, and triggers the correction process when the distance is greater than 0.7.
[0089] By using the technical solution, the continuation trigger condition and the state vector inheritance mechanism are accurately set to ensure the seamless connection of the truncated content, and the twin network trained by TripletLoss enhances the logical contradiction detection capability to further improve the consistency of the report content.
[0090] The embodiment also provides an intelligent report generation method based on multi-module cooperation, which applies the intelligent report generation system based on multi-module cooperation, as shown in Figure 2 The method comprises the following steps:
[0091] S1: identifying and reasonably checking a multi-format document by using a multi-modal OCR recognition engine;
[0092] S2: inputting the recognized data into a trusted data purification system to perform multi-source verification and stability constraint;
[0093] S3: using a dynamic evolution agent to perform online learning and updating of rules;
[0094] S4: generating a complete and coherent long report by using a long report generation engine.
[0095] By using the technical solution, the system architecture based on multi-module cooperation realizes the full-process automation from document recognition, data purification, rule evolution to report generation, reduces manual intervention, and improves the report generation efficiency.
[0096] Specifically, in the embodiment, when generating a long report, a “memory window” mechanism is used to insert a summary paragraph after generating 5000 tokens, and a dynamic block algorithm is used to automatically divide chapters according to semantic correlation, and when it is detected that the output is truncated, an automatic continuation function is triggered and consistency checking is performed.
[0097] According to the technical scheme, for a long report generation scene, a combination strategy of a memory window, dynamic blocking, automatic continuation and consistency checking is used to ensure logical coherence and content integrity of a super-long document.
[0098] Detailed description of working principle
[0099] (I) System architecture and module coordination logic
[0100] The present application realizes intelligent report generation through the cooperative work of four core modules:
[0101] 1. Multi-modal OCR recognition engine → 2. Trusted data purification system → 3. Dynamic evolution agent → 4. Long report generation engine, forming a closed-loop process from document processing to report output.
[0102] (II) Module core working principle
[0103] 1. Multi-modal OCR recognition engine
[0104] 1.1 Format compatibility optimization: Use the generative adversarial network (GAN) to synthesize training data for extreme scenarios such as scanning tilt (±45°), watermark blocking (coverage ≤30%), and combine the self-attention segmentation algorithm to dynamically divide the text area and decorative elements, improving the recognition accuracy of non-standard format documents.
[0105] 1.2 Rationality verification: Based on the industry common sense atlas, the recognition results are compared in real time, and the abnormal data (such as amount decimal point dislocation, negative unit price) are highlighted by red frame and generate correction suggestions.
[0106] 2. Trusted data purification system
[0107] 2.1 Multi-source verification mechanism:
[0108] The agent crawls Internet data (such as business information, industry reports) and cross-verified with local databases in real time, and the data credibility is quantified by the formula .
[0109] 2.2 Stability constraint:
[0110] The KL divergence constraint KL(P output ||P history )<δ is imposed on the output of the large model, limiting the shift of the output distribution, and deterministic decoding (such as BeamSearch) is used for key fields to ensure output consistency.
[0111] 3. Dynamic evolution agent
[0112] 3.1 Online learning framework: Adopt double buffering mechanism (main database service while standby database synchronization update), combined with HNSW-based incremental learning algorithm to achieve millisecond-level retrieval and update of rules.
[0113] 3.2 Feedback closed-loop design: Artificially corrected data automatically triggers reinforcement learning (DQN algorithm) to fine-tune agent decision-making strategy, and retains the last 10 model versions to support version rollback in abnormal situations.
[0114] 4. Long report generation engine
[0115] 4.1 Segment coherence guarantee: Insert summary paragraphs as context memory windows every 5000 tokens, and divide chapters by semantic relevance through dynamic chunking algorithm (BERT similarity > 0.85).
[0116] 4.2 Automatic continuation and verification: Detect output truncation (e.g. terminator or token count >= 4900), inherit the last 10 layers of Transformer encoder state vectors to trigger continuation, and calculate semantic vector Euclidean distance (threshold > 0.7) through twin network (TripletLoss training) to detect logical contradictions.
[0117] Method flow:
[0118] (I) Full process operation steps
[0119] 1. Document input and recognition
[0120] Input multi-format documents (PDF / pictures / scanned copies) into multi-modal OCR recognition engine, automatically preprocess through GAN enhancement and tilt correction, and complete recognition and anomaly labeling with industry common sense graph.
[0121] 2. Data purification and verification
[0122] Recognized data enters the trusted data purification system, and the agent crawls Internet data and cross- verifies with local database, generates high-credibility data through credibility calculation and KL divergence constraint.
[0123] 3. Rule learning and updating
[0124] Dynamic evolution agent updates rule base online through double buffering mechanism and HNSW algorithm based on processed data and human feedback, optimizing analysis logic.
[0125] 4. Long report generation
[0126] Long report generation engine generates segmented content according to memory window and dynamic chunking strategy, automatically continues truncated parts, and verifies logical consistency through twin network, finally integrates into a complete report.
[0127] (B) Key scenario application examples
[0128] Long report generation scenario: When processing documents exceeding 100,000 words, the system inserts summary paragraphs every 5,000 tokens, divides chapters by semantic relevance, and if output truncation is detected (e.g., token number close to 4900), automatically inherits the previous context state vector to trigger continuation writing, ensuring report coherence.
[0129] Overall technical effects
[0130] 1. Solve report integrity issues: Break through the token limit of large models, achieve complete output of ultra-long reports (>100,000 words), and avoid key information truncation and logical breaks.
[0131] 2. Improve data reliability: Through multi-source verification and reliability calculation, the data accuracy rate is improved from 83% in traditional solutions to close to the theoretical extreme value, effectively filtering false or contradictory information.
[0132] 3. Enhance format compatibility: Support robust recognition of 200+ document formats, significantly reduce non-standard format recognition error rates, and reduce human error through real-time anomaly data warning mechanisms.
[0133] 4. Dynamic evolution capability: The rule iteration period is shortened from weekly in traditional solutions to minutes, and the system can continuously optimize analysis logic through online learning to adapt to changes in business scenarios.
[0134] 5. Stable large model output: Output volatility is significantly reduced, key data fields (such as amounts and dates) have significantly improved consistency, and report repeatability meets professional scenario requirements.
[0135] 6. Full-process automation: Reduce manual intervention, automate the entire process from document recognition to report generation, reduce labor costs, improve output efficiency, and analysis results are free from subjective interference, making them more objective.
[0136] In summary: This invention combines multi-modal OCR recognition engine, trusted data purification system, dynamic evolution agent, and long report generation engine to form a complete intelligent report generation system.
[0137] In the multi-modal OCR recognition engine, generative adversarial networks (GAN) are used to synthesize training data and self-attention segmentation algorithms to improve format compatibility and recognition accuracy. The trusted data purification system uses multi-source verification mechanisms and stability constraint algorithms to improve data reliability and output stability. The dynamic evolution agent uses double buffering mechanisms and incremental learning algorithms to enable online learning and rapid updating of rules. The long report generation engine uses memory window mechanisms and dynamic block algorithms to ensure the coherence and completeness of long report generation.
[0138] The above merely preferred embodiments of the present application and are not intended to limit the embodiments and protection scope of the present application. It should be noted by those skilled in the art that any equivalent substitutions and obvious changes made according to the present application description and drawings should be included in the protection scope of the present application.
Claims
1. An intelligent report generation system based on multi-module collaboration, characterized in that: include: Multimodal OCR recognition engine, trusted data purification system, dynamically evolving intelligent agent, and long report generation engine, including: The multimodal OCR recognition engine is used to recognize and verify the rationality of multi-format documents; The trusted data purification system is used to perform multi-source verification and stability constraints on the identified data; The dynamic evolutionary agent is used to realize online learning and updating of rules; The long report generation engine is used to generate a complete and coherent long report.
2. The intelligent report generation system based on multi-module collaboration according to claim 1 is characterized in that: The multimodal OCR recognition engine includes: Generative Adversarial Network module, used to synthesize training data for extreme scenarios such as scan tilt and watermark occlusion; Self-attention segmentation algorithm module, used to dynamically divide text areas and decorative elements; The industry common sense graph module is used to compare recognition results in real time and mark abnormal data.
3. The intelligent report generation system based on multi-module collaboration according to claim 2 is characterized in that: The abnormal data annotation highlights the problem area with a red border and generates correction suggestions.
4. The intelligent report generation system based on multi-module collaboration according to claim 1 is characterized in that: The trusted data purification system includes: Multi-source verification module, which uses intelligent agents to crawl Internet data in real time and cross-verify with local databases; Credibility calculation module, using the formula Calculation is performed, where α is the local data weight coefficient, with a default value of 0.7, which is used to adjust the proportion of local database data in the credibility calculation; D_local is the credibility index of the corresponding data in the local database; D_web is the credibility index of a single Internet data source; N is the number of Internet data sources participating in the cross-validation; The stability constraint module imposes KL divergence constraints on large model outputs and adopts deterministic decoding for key fields.
5. The intelligent report generation system based on multi-module collaboration according to claim 4 is characterized in that: The KL divergence constraint satisfies KL(P output ||P history )<δ, where P_output is the probability distribution of the output results of the current AI large model; P_history is the probability distribution of historical valid output results; δ is the preset KL divergence threshold, which is used to limit the degree of deviation of the output distribution.
6. The intelligent report generation system based on multi-module collaboration according to claim 1 is characterized in that: The dynamically evolving intelligent agent includes: Double buffering mechanism module, the standby database is updated synchronously while the primary database provides services; HNSW-based incremental learning algorithm module supports millisecond-level retrieval; The feedback closed-loop module automatically triggers fine-tuning by manually correcting data, uses reinforcement learning to optimize decision-making strategies, and has a version rollback function.
7. The intelligent report generation system based on multi-module collaboration according to claim 1 is characterized in that: The long report generation engine includes: The memory window mechanism module inserts a summary paragraph after every 5000 tokens generated; Dynamic segmentation algorithm module, automatically dividing chapters according to semantic relevance; The automatic rewriting module triggers a rewriting request and inherits the previous state vector when it detects that the output is truncated; The consistency check module compares the logical contradictions between the previous and next segments through the twin network.
8. The intelligent report generation system based on multi-module collaboration according to claim 7 is characterized in that: The automatic continuation module triggers a continuation request by detecting the terminator or token count threshold of the output sequence, and the inherited previous state vector contains the context representation of the last 10 layers of Transformer encoders; The consistency check module uses TripletLoss to train the twin network, and realizes contradiction detection by calculating the Euclidean distance between the semantic vectors of the previous and next paragraphs. When the distance is greater than 0.7, the correction process is triggered.
9. An intelligent report generation method based on multi-module collaboration, characterized in that: Applying the intelligent report generation system based on multi-module collaboration as described in any one of claims 1 to 8 comprises the following steps: Recognize and verify the rationality of multi-format documents through a multimodal OCR recognition engine; The identified data is input into the trusted data purification system for multi-source verification and stability constraints; Use dynamic evolutionary agents to learn and update rules online; Generate complete and coherent long reports through the long report generation engine.
10. The intelligent report generation method based on multi-module collaboration according to claim 9 is characterized in that: When generating long reports, a "memory window" mechanism is used to insert a summary paragraph after every 5,000 tokens are generated. A dynamic block algorithm is used to automatically divide chapters based on semantic relevance. When output truncation is detected, the automatic continuation function is triggered and a consistency check is performed.