Professional research and judgment report generation method and system based on large language model

Abstract

The invention discloses a major language model-based professional research and judgment report generation method and system. The method comprises the following steps of inputting a research and judgment report generation task instruction into a major language model; inputting the task instruction and the data text into an embedding layer of the large language model to obtain text embedding; combining the obtained prompt with text embedding, inputting the combined prompt and text embedding into a large language model, and performing relation extraction on personnel and cases in a research and judgment task; outputting a relation extraction result of the model to meet a target specified in the task instruction; a current task instruction is disassembled, and analysis and query are performed through an F query function according to instructions oriented to various scenes in an instruction library; and according to the data set processed in the step S400, finally realizing the output of the research and judgment report in a standard format. The method has the advantages that automatic and efficient professional research and judgment report generation is realized, and the labor cost is saved; the report content is combined with domain knowledge, and high specialty and logicality are achieved; and the applicability of the multi-scene improvement report is flexibly supported.

Description

Technical Field

[0001] The present invention relates to the field of generating professional research and judgment reports based on a large language model, and in particular to a method and system for generating professional research and judgment reports based on a large language model. Background Art

[0002] With the rapid development of big data and artificial intelligence technologies, many industries have an increasing demand for professional reports based on massive data analysis that are valuable for decision-making references. However, traditional methods rely on manual writing, which is inefficient and susceptible to human subjectivity. In recent years, large language models (such as the GPT series) have demonstrated powerful generation capabilities in the field of natural language processing, but direct application to professional report generation still has problems such as insufficient domain knowledge and inaccurate generated content. Therefore, it has become an urgent need for the development of the industry to develop a method for generating professional research and judgment reports based on large language models, combining domain knowledge to improve the accuracy and practicality of generated content. Summary of the invention

[0003] The present invention writes a proprietary prompt word library, creates an application instruction set and an analysis and judgment template library for analysis and judgment business scenarios, builds a rule library for data classification, accesses multi-source heterogeneous data supporting analysis and judgment reports through various tools, integrates professional knowledge bases, and optimizes the generation process to solve the problems of low efficiency, insufficient accuracy, serious hallucinations, and unavailable results in report generation in the prior art.

[0004] The purpose of the present invention is to provide a method and system for generating a professional research and judgment report based on a large language model, so as to solve the above-mentioned problems existing in the prior art.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] A method for generating a professional research and judgment report based on a large language model comprises the following steps:

[0007] S100, inputting the task instructions generated by the research and judgment report into the large language model;

[0008] S200, design prompt word library module, instruction set module and analysis template library module;

[0009] S300, input the task instruction and data text into the embedding layer of the large language model to obtain text embedding; combine the obtained prompt with the text embedding, input it into the large language model, extract the relationship between the personnel and cases in the judgment task; output the relationship extraction result of the model to meet the goal specified in the task instruction;

[0010] S400, disassemble the current task instruction, and analyze and query the instructions for various scenarios in the instruction library through the F query function;

[0011] S500. Based on the data set processed in step S400, start calling the analysis and judgment template data, apply the template format, type, and style, and finally output the analysis and judgment report in a standard format.

[0012] Preferably, the task instruction includes: a task description and a task objective.

[0013] Preferably, the prompt word library module includes: Cprompt prompt, Xprompt prompt and Pprompt prompt;

[0014] Cprompt prompts: task prompts for personnel impulsivity, emotions, and behavioral out-of-control;

[0015] Xprompt prompts: task prompts for criminal cases, clues, and entity relationships;

[0016] Pprompt: Task prompt designed for general analysis and judgment logic.

[0017] Preferably, step S200 further includes:

[0018] When a task instruction is passed in, the query function interacts with the Cprompt prompt, Xprompt prompt, and Pprompt prompt to find the prompt related to the current task; the cosine similarity is used to measure the relevance of each prompt prompt with the current prompt prompt; the Cprompt prompt, Xprompt prompt, and Pprompt prompt are sorted according to the similarity to ensure that the most relevant prompt is used first; if no match can be found, the Pprompt prompt is used by default to judge the prompt for the task.

[0019] Preferably, the prompt prompt interacts with the Cprompt prompt, the Xprompt prompt, and the Pprompt prompt through a query function, and the query function formula is as follows:

[0020] prompt=∑F(X)

[0021] Through an F function, i.e., a query function, all prompts related to the current prompt (X) are found, and the prompt set related to the task is obtained; the cosine similarity between the embedded vector of each prompt and the embedded vector of the current prompt is calculated to obtain a distance score;

[0022] Based on the scores, the prompt words are applied.

[0023] Preferably, in step S400, the query function formula is Instruction = ∑F(M, E);

[0024] Through an F query function, according to M, i.e., the tasks analyzed and disassembled by the large model and E, i.e., the tasks in the instruction library, a search is performed to match the associated task list, i.e., Instruction. According to Instruction, each task is processed again according to step 300. Finally, Instruction is executed and a collection of data processed by multiple task instructions is obtained.

[0025] Preferably, when applying the template format, type, and style in S500, the formula used is as follows:

[0026] data=∑LLM(D,Mb)

[0027] Through an LLM, i.e., large model processing function, according to D, i.e., the data set to be processed, and Mb, i.e., the model list, the data to be processed and the template are reasoned and analyzed through LLM, the most appropriate analysis and judgment template is matched, and then D is filled into the selected template, and finally the analysis and judgment report data is output.

[0028] Preferably, the distance formula between vectors is calculated by cosine similarity:

[0029]

[0030] A system for generating professional research and judgment reports based on a large language model based on the same concept includes:

[0031] Data acquisition module, data fusion module, index construction module, prompt word library module, instruction set module, analysis and judgment template library module, large model execution process control module and analysis and judgment report generation module;

[0032] Data acquisition module, used for collecting and processing data;

[0033] Data fusion module, used to align and fuse the processed text data and structured data;

[0034] Index building module, used to build a multi-level index system;

[0035] The prompt word library module is used to build a professional prompt word library;

[0036] Instruction set module, used to build instructions;

[0037] The template library module is used to build a template library containing classification, style, and data format content;

[0038] The large model execution process control module is used to connect the prompt word library, professional instruction set, analysis template library and related components of the large language model to form a complete processing chain;

[0039] The analysis and judgment report generation module is used to generate the analysis and judgment report.

[0040] Preferably, the data acquisition module includes:

[0041] Data analysis module, used to determine the target areas and task types for information extraction;

[0042] Structured data processing module, used for preprocessing of structured data;

[0043] Unstructured data processing module, used for preprocessing of unstructured data;

[0044] The data fusion module includes: data alignment, vector processing and LoRA fine-tuning;

[0045] The research and judgment report generation module includes: research and judgment report generation, research and judgment reasoning and report summary generation.

[0046] The specific implementation of this system is as follows:

[0047] The first step is to determine the target areas and task types for information extraction

[0048] Based on the data analysis module, we first need to identify the target areas and task types for information extraction. Task types include but are not limited to policy analysis, budget planning, social event analysis, etc. To this end, we need to analyze the data types and structures to be processed and identify the expected output formats, such as classification labels, numerical scores, or event associations.

[0049] At the same time, multiple factors such as data scale, data quality and data privacy are comprehensively considered to ensure that the collected data complies with privacy protection regulations and can be processed efficiently and accurately. Data collection covers text, tables, charts and other forms, and the collection scope includes various resources such as literature, articles, government reports, news, etc. in related fields. By sorting and classifying these data, the original data input for the vector knowledge base is finally formed. This step aims to provide a comprehensive and structured original data foundation for subsequent data processing and model training.

[0050] Step 2: Structured data processing

[0051] Based on the structured data processing module, data preprocessing is performed. The processing of structured data includes multiple steps such as data loading, cleaning and standardization:

[0052] 1) Data loading and preliminary inspection: First, load the structured data from the data source and perform a preliminary inspection to obtain the basic statistical information of the data (such as mean, standard deviation, data type, etc.).

[0053] 2) Data cleaning: remove irrelevant or erroneous records, fill in missing values, and handle outliers (detect and eliminate abnormal data through the Z-Score method).

[0054] 3) Data standardization: Standardize numerical data, unify date formats, currency units, etc., ensure data consistency, and eliminate deviations between different units.

[0055] 4) Deduplication: Delete duplicate records to avoid data redundancy.

[0056] 5) Dealing with the class imbalance problem: Ensure that the sample sizes of different categories are balanced through oversampling or undersampling methods.

[0057] 6) Feature Engineering: Select and construct features that have a significant impact on model performance and optimize data quality.

[0058] Through this series of data preprocessing steps, the data quality can be effectively improved, the accuracy of subsequent model training can be ensured, and reliable data support can be provided for tasks such as policy analysis and budget planning.

[0059] Step 3: Unstructured Data Preprocessing

[0060] Based on the unstructured data processing module, data preprocessing is performed. Unstructured data (such as text, images, audio, etc.) is more difficult to process and usually has no fixed data model. Therefore, for the preprocessing of unstructured data, intelligent technology is needed to deal with the problems of difficulty in data standardization and difficulty in understanding.

[0061] 1) Data loading and preliminary check: Load unstructured data from different formats (Word, Excel, PDF, TXT, etc.) and perform basic file checks to ensure data integrity.

[0062] 2) Data cleaning: For example, when extracting text from an HTML file, it is necessary to remove irrelevant information such as HTML tags and special characters.

[0063] 3) Text preprocessing, specific steps:

[0064] ① Word segmentation: Segment the text into words and extract the smallest meaningful unit.

[0065] ② Stemming and word form restoration: Standardize the vocabulary into stem form (for example, restore "running" to "run").

[0066] ③Entity recognition: Extract key entities (such as date, place, person, etc.) through named entity recognition (NER) technology.

[0067] ④Data enhancement: Generate more training samples and improve the generalization ability of the model through enhancement technology. 4) Text vectorization: Convert the processed text into vector form for use in the training and reasoning of machine learning models.

[0068] Processing unstructured data through intelligent technology can ensure the consistency and accuracy of text data and provide high-quality support for policy analysis and intelligent retrieval.

[0069] Step 4: Fusion of text and structured data

[0070] Based on the data fusion module, the processed text data and structured data need to be aligned and fused so that they can be processed together in the same model. To this end, the early fusion method is adopted:

[0071] 1) Concatenate text embeddings (such as vectors generated by BERT or other pre-trained language models) and structured data embeddings (such as vector representations of numerical and categorical data) into a long vector.

[0072] 2) The long vector is input into the GLM-4 model for processing.

[0073] 3) The self-attention mechanism (Se l f-Attention) is adopted. Through the multi-head attention network, the model can give different attention levels according to different data sources, thereby efficiently integrating information from text and structured data.

[0074] This fusion method can effectively integrate text information and structured data, improve the processing capabilities of multimodal data, and enable the model to make more accurate decision inferences.

[0075] Step 5: Build a multi-level index system

[0076] In order to facilitate the retrieval and recall of information, this embodiment constructs a multi-level and multi-dimensional index system. The system combines keywords, structured paragraph information and semantic information, and after embedding processing, all data is stored in a vector database (such as Weaviate).

[0077]

[0078] Relevant data is thus recalled by similarity. This enables the most relevant data to be returned quickly and accurately when a query is made.

[0079] This indexing system effectively improves the efficiency and accuracy of data retrieval, and can provide a reliable data foundation for subsequent decision support and intelligent reasoning.

[0080] Step 6: Build a professional reminder vocabulary

[0081] For different analysis scenarios (such as fighting, fraud, emotional impulse, etc.), a professional prompt word library is established for classification management. By professionally tuning the prompt words, the reasoning accuracy of large models (such as GLM4-9B) in specific scenarios can be improved.

[0082] Through the management and optimization of professional prompt vocabulary, the reasoning accuracy of large language models in complex scenarios can be effectively improved, providing precise support for business decisions.

[0083] Step 7: Build a professional instruction set

[0084] Based on the thinking method of large language models, professional instruction sets are developed for analysis scenarios and data.

[0085] By honing appropriate instructions, the model is guided to summarize, infer and generate indicator data, entity data, event data, etc.

[0086] This instruction set can effectively improve the model's ability to process complex data and guide the model to perform accurate reasoning and generation tasks.

[0087] Step 8: Construction of the research template library

[0088] For different types of research and judgment reports, a template library is built that includes classification, style, data format, etc. By annotating and uniformly managing report templates, the large language model is supported to follow a consistent structure and style when generating reports.

[0089] The template library provides a unified structure and requirements, making the generated research and analysis reports more standardized, improving the report quality and readability.

[0090] Step 9: Execute process control based on the workflow model

[0091] Through the workflow management system, the prompt word library, professional instruction set, analysis and judgment template library, etc. are connected with the relevant components of the large language model to form a complete processing chain. The large language model is introduced through multiple links to realize functions such as data retrieval, feature processing, and rule processing, and finally output a summary and analysis report in natural language format. This process control system can ensure the efficient execution of the large model and ensure the high quality and professionalism of the results.

[0092] Through the implementation of the above steps, a complete knowledge extraction and analysis system based on a large language model was finally formed, providing powerful intelligent support for various decision support tasks.

[0093] The beneficial effects of the present invention are as follows: the present invention discloses a method for generating a professional research and judgment report based on a large language model, comprising the following steps: inputting a task instruction for generating a research and judgment report into a large language model; designing a prompt word library module, an instruction set module, and a research and judgment template library module; inputting the task instruction and data text into the embedding layer of the large language model to obtain text embedding; combining the obtained prompt with the text embedding, inputting it into the large language model, and extracting the relationship between the personnel and cases in the research and judgment task; outputting the relationship extraction result of the model to meet the target specified in the task instruction; disassembling the current task instruction, and analyzing and querying it through the F query function according to the instructions for various scenarios in the instruction library; according to the data set processed by step S400, starting to call the research and judgment template data, applying the template format, type, and style, and finally realizing the output of the standard format of the research and judgment report. The advantages of the present invention are: realizing automatic and efficient generation of professional research and judgment reports, saving labor costs; the report content combines domain knowledge, has high professionalism and logic; flexibly supports multi-scenario and multi-language output, and improves the applicability of the report. BRIEF DESCRIPTION OF THE DRAWINGS

[0094] Figure 1 It is a flow chart of a method for generating a professional research and judgment report based on a large language model of the present invention;

[0095] Figure 2 It is a functional block diagram of the research and judgment report generation system of the present invention;

[0096] Figure 3 is a flow chart of structured data preprocessing of the present invention;

[0097] Figure 4 is a flow chart of unstructured data preprocessing of the present invention;

[0098] Figure 5 It is the workflow-based large model execution control diagram of the present invention. DETAILED DESCRIPTION

[0099] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation methods described herein are only used to explain the present invention and are not used to limit the present invention.

[0100] Reference Figure 1 , Figure 2 , Figure 3 , Figure 4 and Figure 5A method for generating a professional research and judgment report based on a large language model is shown, comprising the following steps:

[0101] S100, inputting the task instructions generated by the research and judgment report into the large language model;

[0102] S200, design prompt word library module, instruction set module and analysis template library module;

[0103] S300, input the task instruction and data text into the embedding layer of the large language model to obtain text embedding; combine the obtained prompt with the text embedding, input it into the large language model, extract the relationship between the personnel and cases in the judgment task; output the relationship extraction result of the model to meet the goal specified in the task instruction;

[0104] S400, disassemble the current task instruction, and analyze and query the instructions for various scenarios in the instruction library through the F query function;

[0105] S500. Based on the data set processed in step S400, start calling the analysis and judgment template data, apply the template format, type, and style, and finally output the analysis and judgment report in a standard format.

[0106] Preferably, the task instruction includes: a task description and a task objective.

[0107] Preferably, the prompt word library module includes: Cprompt prompt, Xprompt prompt and Pprompt prompt;

[0108] Cprompt prompts: task prompts for personnel impulsivity, emotions, and behavioral out-of-control;

[0109] Xprompt prompts: task prompts for criminal cases, clues, and entity relationships;

[0110] Pprompt: Task prompt designed for general analysis and judgment logic.

[0111] Preferably, step S200 further includes:

[0112] When a task instruction is passed in, the query function interacts with the Cprompt prompt, Xprompt prompt, and Pprompt prompt to find the prompt related to the current task; the cosine similarity is used to measure the relevance of each prompt prompt with the current prompt prompt; the Cprompt prompt, Xprompt prompt, and Pprompt prompt are sorted according to the similarity to ensure that the most relevant prompt is used first; if no match can be found, the Pprompt prompt is used by default to judge the prompt for the task.

[0113] Preferably, the prompt prompt interacts with the Cprompt prompt, the Xprompt prompt, and the Pprompt prompt through a query function, and the query function formula is as follows:

[0114] prompt=∑F(X)

[0115] Through an F function, i.e., a query function, all prompts related to the current prompt (X) are found, and the prompt set related to the task is obtained; the cosine similarity between the embedded vector of each prompt and the embedded vector of the current prompt is calculated to obtain a distance score;

[0116] Based on the scores, the prompt words are applied.

[0117] Preferably, in step S400, the query function formula is Instruction = ∑F(M, E);

[0118] Through an F query function, according to M, i.e., the tasks analyzed and disassembled by the large model and E, i.e., the tasks in the instruction library, a search is performed to match the associated task list, i.e., Instruction. According to Instruction, each task is processed again according to step 300. Finally, Instruction is executed and a collection of data processed by multiple task instructions is obtained.

[0119] Preferably, when applying the template format, type, and style in S500, the formula used is as follows:

[0120] data=∑LLM(D,Mb)

[0121] Through an LLM, i.e., large model processing function, according to D, i.e., the data set to be processed, and Mb, i.e., the model list, the data to be processed and the template are reasoned and analyzed through LLM, the most appropriate analysis and judgment template is matched, and then D is filled into the selected template, and finally the analysis and judgment report data is output.

[0122] Preferably, the distance formula between vectors is calculated by cosine similarity:

[0123]

[0124] A system for generating professional research and judgment reports based on a large language model based on the same concept includes:

[0125] Data acquisition module, data fusion module, index construction module, prompt word library module, instruction set module, analysis and judgment template library module, large model execution process control module and analysis and judgment report generation module;

[0126] Data acquisition module, used for collecting and processing data;

[0127] Data fusion module, used to align and fuse the processed text data and structured data;

[0128] Index building module, used to build a multi-level index system;

[0129] The prompt word library module is used to build a professional prompt word library;

[0130] Instruction set module, used to build instructions;

[0131] The template library module is used to build a template library containing classification, style, and data format content;

[0132] The large model execution process control module is used to connect the prompt word library, professional instruction set, analysis template library and related components of the large language model to form a complete processing chain;

[0133] The analysis and judgment report generation module is used to generate the analysis and judgment report.

[0134] Preferably, the data acquisition module includes:

[0135] Data analysis module, used to determine the target areas and task types for information extraction;

[0136] Structured data processing module, used for preprocessing of structured data;

[0137] Unstructured data processing module, used for preprocessing of unstructured data;

[0138] The data fusion module includes: data alignment, vector processing and LoRA fine-tuning;

[0139] The research and judgment report generation module includes: research and judgment report generation, research and judgment reasoning and report summary generation.

[0140] By adopting the above technical solution disclosed in the present invention, the following beneficial effects are obtained:

[0141] The present invention discloses a method for generating a professional research and judgment report based on a large language model, comprising the following steps: inputting a task instruction generated by the research and judgment report into the large language model; designing a prompt word library module, an instruction set module and a research and judgment template library module; inputting the task instruction and data text into the embedding layer of the large language model to obtain text embedding; combining the obtained prompt with the text embedding, inputting it into the large language model, and extracting the relationship between the personnel and the case in the research and judgment task; outputting the relationship extraction result of the model to meet the target specified in the task instruction; disassembling the current task instruction, and analyzing and querying it through the F query function according to the instructions for various scenarios in the instruction library; according to the data set processed by step S400, starting to call the research and judgment template data, applying the template format, type and style, and finally realizing the output of the standard format of the research and judgment report. The advantages of the present invention are: realizing automatic and efficient generation of professional research and judgment reports, saving labor costs; the report content combines domain knowledge, has high professionalism and logic; flexibly supports multi-scenario and multi-language output, and improves the applicability of the report.

[0142] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be considered as the scope of protection of the present invention.

Claims

1. A method for generating a professional research and judgment report based on a large language model, characterized in that: The following steps are involved: S100, inputting the task instructions generated by the research and judgment report into the large language model; S200, design prompt word library module, instruction set module and analysis template library module; S300, inputting the task instruction and data text into the embedding layer of the large language model to obtain text embedding; combining the obtained prompt with the text embedding, inputting the result into the large language model, and extracting the relationship between the personnel and cases in the judgment task; Output the relationship extraction results of the model to meet the objectives specified in the task instructions; S400, disassemble the current task instruction, and analyze and query the instructions for various scenarios in the instruction library through the F query function; S500. Based on the data set processed in step S400, start calling the analysis and judgment template data, apply the template format, type and style, and finally output the analysis and judgment report in a standard format.

2. The method for generating a professional research and judgment report based on a large language model according to claim 1 is characterized in that: The task instruction includes: task description and task objective.

3. The method for generating a professional research and judgment report based on a large language model according to claim 1 is characterized in that: The prompt word library module includes: Cprompt prompt, Xprompt prompt and Pprompt prompt; The Cprompt prompt: a task prompt for personnel impulsivity, emotions, and uncontrolled behavior; The Xprompt prompts: task prompts for criminal cases, clues, and entity relationships; The Pprompt prompt is a task prompt designed for general analysis and judgment logic.

4. The method for generating a professional research and judgment report based on a large language model according to claim 3 is characterized in that: The step S200 further includes: When a task instruction is passed in, the query function interacts with the Cprompt prompt, the Xprompt prompt and the Pprompt prompt to find a prompt related to the current task; the cosine similarity is used to measure the relevance of each prompt prompt with the current prompt prompt; the Cprompt prompt, the Xprompt prompt and the Pprompt prompt are sorted according to the similarity to ensure that the most relevant prompt is used first; if no match can be found, the Pprompt prompt is used by default to judge the prompt of the task.

5. The method for generating a professional research and judgment report based on a large language model according to claim 4 is characterized in that: The prompt prompt interacts with the Cprompt prompt, the Xprompt prompt, and the Pprompt prompt through a query function, and the query function formula is as follows: prompt=∑F(X) Through an F function, i.e., a query function, all prompts related to the current prompt (X) are found, and the prompt set related to the task is obtained; the cosine similarity between the embedded vector of each prompt and the embedded vector of the current prompt is calculated to obtain a distance score; Based on the scores, the prompt words are applied.

6. The method for generating a professional research and judgment report based on a large language model according to claim 5 is characterized in that: In the step S400, the query function formula is Instruction = ∑F(M, E); Through an F query function, according to M, i.e., the tasks analyzed and disassembled by the large model and E, i.e., the tasks in the instruction library, a search is performed to match the associated task list, i.e., Instruction. According to the Instruction, each task is processed again according to the step 300. Finally, the Instruction is executed and a collection of data after processing multiple task instructions is obtained.

7. The method for generating a professional research and judgment report based on a large language model according to claim 6 is characterized in that: When applying the template format, type, and style in S500, the formula used is as follows: data=∑LLM(D,Mb) Through an LLM, i.e., large model processing function, according to D, i.e., the data set to be processed, and Mb, i.e., the model list, the data to be processed and the template are reasoned and analyzed through LLM, the most appropriate analysis and judgment template is matched, and then D is filled into the selected template, and finally the analysis and judgment report data is output.

8. The method for generating a professional research and judgment report based on a large language model according to claim 7 is characterized in that: The distance formula between vectors is calculated by the cosine similarity:

9. A system for generating professional research and judgment reports based on a large language model, characterized in that: include: Data acquisition module, data fusion module, index construction module, prompt word library module, instruction set module, analysis and judgment template library module, large model execution process control module and analysis and judgment report generation module; The data acquisition module is used to collect and process data; The data fusion module is used to align and fuse the processed text data and structured data; The index building module is used to build a multi-level index system; The prompt word library module is used to build a professional prompt word library; The instruction set module is used to construct instructions; The analysis and judgment template library module is used to construct a template library containing classification, style, and data format content; The large model execution process control module is used to connect the prompt word library, professional instruction set, analysis template library and related components of the large language model to form a complete processing chain; The research and judgment report generating module is used to generate a research and judgment report.

10. The system for generating professional research and judgment reports based on a large language model according to claim 9, characterized in that: The data acquisition module comprises: Data analysis module, used to determine the target areas and task types for information extraction; Structured data processing module, used for preprocessing of structured data; Unstructured data processing module, used for preprocessing of unstructured data; The data fusion module includes: data alignment, vector processing and LoRA fine-tuning; The research and judgment report generation module includes: research and judgment report generation, research and judgment reasoning and report summary generation.

Images