Large model-based financial data analysis method, system and medium

Abstract

The application discloses a large model-based financial data analysis method and system and a medium, mainly relates to the technical field of financial data analysis, and is used to solve the problem of the isolation of each module in the prior art architecture, that is, the lack of a cooperative mechanism among the three links of data query, rule application and knowledge calling. The method comprises the following steps: acquiring initial financial problem data, disassembling the initial financial problem data into a financial problem string, a rule problem string and a query problem string; obtaining corresponding financial knowledge; obtaining corresponding financial early warning rules; integrating all corresponding query sub-SQL into a final query SQL by a large model; executing the final query SQL by an execution engine to obtain a query result; taking the financial knowledge, the financial early warning rules and the query result as input parameters of a financial data summary function, and then obtaining a prompt word composed of the financial knowledge, the financial early warning rules and the query result; inputting the prompt word into the large model to obtain a returned summary string.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of financial data analysis, and particularly relates to a financial data analysis method and system based on a large model and a medium. BACKGROUND

[0002] The current financial data analysis field mainly relies on the mode of combining traditional rule engines with manual intervention. The existing technical bottlenecks mainly manifest in three aspects: first, the generation of query statements needs to predefine all possible data requirements, which cannot dynamically adapt to new analysis dimensions; second, the early warning rules are separated from business knowledge, resulting in a lack of context association in risk judgment; and third, the result analysis link is still highly dependent on manual work, and the intelligent system can only provide raw data tables.

[0003] The traditional method has three core defects: first, the rigidity of query construction leads to low analysis efficiency. When facing complex requirements such as "comparing the accounts receivable turnover rate of each region with the industry benchmark", it is necessary to manually write multi-table association SQL and calculate the indicators, which is time-consuming and prone to errors. Second, the risk warning is separated from the business knowledge. The existing system fails to report risks in scenarios such as "inventory turnover days exceeding the credit period" which need to combine contract terms and accounting policies due to the lack of knowledge association mechanism. Third, the analysis conclusion generation is mechanized. The existing scheme either outputs unexplained data lists or applies fixed report templates, and cannot dynamically organize professional discussions according to specific problems. These problems essentially result from the isolation of each module in the technical architecture - the lack of coordination mechanism among data query, rule application, and knowledge invocation three links, which leads to the system being unable to perform contextually coherent multi-dimensional analysis like professional financial personnel. SUMMARY

[0004] The present application provides a financial data analysis method and system based on a large model to solve the problem of the isolation of each module in the existing technical architecture - the lack of coordination mechanism among data query, rule application, and knowledge invocation three links, which leads to the system being unable to perform contextually coherent multi-dimensional analysis like professional financial personnel.

[0005] In a first aspect, the present application provides a financial data analysis method based on a large model, the method comprising:

[0006] The initial financial problem data is obtained, the initial financial problem data is disassembled into a financial problem string, a rule problem string, and a query problem string, the financial problem string is taken as an input parameter of a financial business knowledge retrieval function, and then corresponding financial knowledge is obtained based on a preset financial knowledge vector library, the rule problem string is taken as an input parameter of a financial early warning threshold retrieval function, and then corresponding financial early warning rules are obtained based on a preset financial early warning vector library, the query problem string is taken as an input parameter of a financial data SQL generation function, and then corresponding query sub-SQL is generated based on database table information by using a large model, the initial financial problem data and the query sub-SQL are taken as input parameters of a financial report data query function, then the initial financial problem data is used, and the large model integrates all the query sub-SQL into final query SQL, the final query SQL is executed by an execution engine to obtain a query result, the financial knowledge, the financial early warning rules, and the query result are taken as input parameters of a financial data summary function, and then prompt words composed of the financial knowledge, the financial early warning rules, and the query result are obtained, the prompt words are input into the large model, and a returned summary string is obtained.

[0007] In an implementation manner of the present application, the financial problem string is taken as an input parameter of the financial business knowledge retrieval function, and then corresponding financial knowledge is obtained based on the preset financial knowledge vector library, specifically including:

[0008] The financial knowledge documents are stored in the preset financial knowledge vector library through slicing vectorization;

[0009] The financial business knowledge retrieval function retrieves and recalls slices from the preset financial knowledge vector library according to the semantic of the financial problem string, calls the large model, and processes the slices into the financial knowledge according to the financial problem string.

[0010] In an implementation manner of the present application, the financial problem string is taken as an input parameter of the financial business knowledge retrieval function, and then corresponding financial knowledge is obtained based on the preset financial knowledge vector library, specifically including:

[0011] The financial early warning rule documents are stored in the preset financial early warning vector library through slicing vectorization;

[0012] The financial early warning threshold retrieval function retrieves and recalls slices from the preset financial early warning vector library according to the semantic of the rule problem string, calls the large model, and processes the slices into the financial early warning rules according to the rule problem string.

[0013] In an implementation manner of the present application, the financial early warning threshold retrieval function retrieves and recalls slices from the preset financial early warning vector library according to the semantic of the rule problem string, specifically including:

[0014] The RAG enhanced retrieval technology is used to make a semantic similarity comparison in the preset financial early warning vector library, and the slice with the highest similarity is recalled.

[0015] In one implementation of this application, the query question string is used as the input parameter of the financial data SQL generation function, and then, based on the database table information, the corresponding query sub-SQL is generated using a large model, specifically including:

[0016] Input the query question string and database table information into the preset prompt template to obtain SQL prompt words. Input the SQL prompt words into the large model to obtain the corresponding query sub-SQL.

[0017] In one implementation of this application, the method further includes:

[0018] The initial financial problem data and query results are used as input parameters to the financial chart generation function to obtain chart generation prompts containing the initial financial problem data and query results. The chart generation prompts are then input into the large model to obtain a preset display chart containing the initial financial problem data and query results.

[0019] Secondly, this application provides a financial data analysis system based on a large model, the system comprising:

[0020] The segmentation module is used to obtain the initial financial problem data and break it down into financial problem strings, rule problem strings, and query problem strings.

[0021] The knowledge acquisition module is used to take the string of financial questions as the input parameter of the financial business knowledge retrieval function, and then obtain the corresponding financial knowledge based on the preset financial knowledge vector library.

[0022] The rule acquisition module is used to take the rule question string as the input parameter of the financial early warning threshold retrieval function, and then obtain the corresponding financial early warning rule based on the preset financial early warning vector library;

[0023] The sub-SQL acquisition module is used to take the query question string as the input parameter of the financial data SQL generation function, and then generate the corresponding query sub-SQL based on the database table information and the large model.

[0024] The result acquisition module is used to take the initial financial problem data and query sub-SQL as input parameters for the financial statement data query function; then, using the initial financial problem data, the large model integrates all the corresponding query sub-SQL into the final query SQL; the execution engine executes the final query SQL to obtain the query results.

[0025] The return module is used to take financial knowledge, financial early warning rules, and query results as input parameters to the financial data summary function, and then obtain prompt words composed of financial knowledge, financial early warning rules, and query results. Input the prompt words into the large model to obtain the returned summary string.

[0026] In one implementation of this application, the knowledge acquisition module includes a knowledge acquisition unit.

[0027] This is used to store financial knowledge documents into a preset financial knowledge vector library by slicing and vectorizing them.

[0028] The financial business knowledge retrieval function semantically retrieves slices from a pre-defined financial knowledge vector library based on the financial question string; it then calls the large model to process the slices into financial knowledge based on the financial question string.

[0029] In one implementation of this application, the system further includes a display module.

[0030] This function takes initial financial problem data and query results as input parameters to generate a financial chart, obtains chart generation prompts containing the initial financial problem data and query results, inputs the chart generation prompts into a large model, and obtains a preset display chart containing the initial financial problem data and query results.

[0031] Thirdly, this application provides a non-volatile computer storage medium storing computer instructions, which, when executed, implement a financial data analysis method based on a large model as described above.

[0032] As can be seen from the above technical solutions, this application has the following advantages:

[0033] It achieves intelligent collaboration across the entire financial analysis process: by using big data model technology to connect the three traditionally isolated modules of data query, rule application, and knowledge retrieval, it forms a contextual and coherent analytical capability similar to that of a professional financial person.

[0034] Improved analysis efficiency: Automatically completes the entire process of problem decomposition → knowledge retrieval → rule matching → SQL generation → result integration, saving time and costs compared to manual operation.

[0035] The accuracy of the analysis results has been enhanced: through the dual guarantee of vectorized knowledge base retrieval and large-scale intelligent reasoning, the financial analysis conclusions are both in line with professional standards and adaptable to business scenarios.

[0036] It lowers the barrier to entry: non-technical personnel can obtain a complete analysis report containing data, rules, and professional explanations simply by entering a natural language question, without needing to master SQL or financial expertise.

[0037] It achieves dynamic early warning capability: by matching the early warning rule base in real time, it can automatically identify financial risk points and generate early warning prompts during the analysis process.

[0038] It possesses continuous evolution characteristics: based on the feedback learning mechanism of a large model, the system will continuously optimize the processing accuracy of each link as the frequency of use increases. Attached Figure Description

[0039] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart of a financial data analysis method based on a large model provided in an embodiment of this application.

[0041] Figure 2 This is a schematic diagram of the internal structure of a financial data analysis system based on a large model, provided in an embodiment of this application. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] Those skilled in the art should understand that the embodiments described below are merely preferred embodiments of this disclosure and do not imply that this disclosure can only be implemented through these preferred embodiments. These preferred embodiments are merely used to explain the technical principles of this disclosure and are not intended to limit the scope of protection of this disclosure. Based on the preferred embodiments provided by this disclosure, all other embodiments obtained by those skilled in the art without creative effort should still fall within the scope of protection of this disclosure.

[0044] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0045] The technical solutions proposed in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0046] The embodiment provides a financial data analysis method based on a large model, such as Figure 1 As shown in the embodiments of this application, the method mainly includes the following steps:

[0047] Step 110: Obtain the initial financial problem data and break it down into financial problem string, rule problem string, and query problem string.

[0048] In some embodiments, the task breakdown result for the problem "Which industrial enterprises have triggered the debt-to-asset ratio warning line" can be:

[0049] Financial question string: "What are some relevant financial concepts regarding the debt-to-equity ratio?";

[0050] The rule question string is: "What is the warning line for the debt-to-asset ratio of industrial enterprises?"

[0051] The query string is something like: "How should industrial companies query their financial statements?"

[0052] The query string can contain multiple characters.

[0053] It should be noted that the decomposition process can be implemented using existing large-scale models capable of decomposing statements.

[0054] More specifically, the disassembly process can be described as follows:

[0055] Receive initial financial questions described in natural language (such as user questions or system logs). Remove irrelevant symbols, standardize terminology (e.g., change "Q3" to "Third Quarter"), and correct spelling errors.

[0056] Use NLP (Natural Language Processing) tools (such as Baidu ERNIE, spaCy) to identify the following in sentences: Core actions (e.g., "calculate," "verify"): associated with financial question strings. Constraints (e.g., "complies with XX rule"): associated with rule question strings. Data requirements (e.g., "sales revenue in Q3 2024"): associated with query question strings.

[0057] Example:

[0058] Question: "How can I verify whether the gross profit margin for Q1 2024 complies with the new accounting standards?"

[0059] Action: "Verify" → Financial Issues;

[0060] Constraint: "Compliance with the new accounting standards" → Rule-related issue;

[0061] Data: "Gross margin in Q1 2024" → Question.

[0062] Import the financial question, rule question, and query question into their respective preset prompt word templates to obtain financial question prompt words, rule question prompt words, and query question prompt words. Input each prompt word into the large model (a pre-trained Transformer architecture model) and use the large model to obtain the output financial question string, rule question string, and query question string.

[0063] Alternatively, the method for training a Transformer architecture model can be to obtain sample prompts and corresponding sample strings, input them into the Transformer architecture model, and obtain the trained Transformer architecture model.

[0064] Step 120: Use the string of financial questions as the input parameter of the financial business knowledge retrieval function, and then obtain the corresponding financial knowledge based on the preset financial knowledge vector library.

[0065] This step can be specifically described as follows:

[0066] The FinancialKnowledgeRetrieval function takes a string of financial questions as input and returns financial knowledge. This function is actually a Reinforced Financial Knowledge Retrieval (RAG) process. Specifically, before execution, financial-related knowledge documents are sliced ​​and vectorized into a financial knowledge vector database. During function execution, fragments are retrieved from the vector database based on the semantics of the input questions. These retrieved fragments are then processed and summarized into final financial knowledge by a larger model (a pre-trained Transformer architecture model; the training and invocation process is shown in step 110). This process includes identifying which financial indicators are point-in-time figures and which belong to different periods. The function also clarifies the business meanings of point-in-time figures and period figures.

[0067] As will be understood by those skilled in the art, this step mainly involves: inputting a string of financial questions → retrieving financial knowledge vectors → calling up a large model summary → outputting structured financial knowledge.

[0068] This step utilizes a vectorized financial knowledge base to achieve semantic search, avoiding the limitations of traditional keyword matching (such as the difference in business meaning between "point-in-time figures" and "period figures"). Example: When a user asks "What are point-in-time figures?", the system can accurately retrieve relevant financial document snippets (such as balance sheet indicators).

[0069] The retrieval results are further processed using large models (such as pre-trained Transformer architecture models) to generate summary knowledge that meets user needs (such as comparing the business meaning of point-in-time numbers and period numbers). This avoids redundant information from directly returning the original document, improving readability. Furthermore, the function FinancialKnowledgeRetrieval encapsulates the entire process of retrieval, recall, and summarization, facilitating reuse and expansion. The financial knowledge base can be updated independently (e.g., adding new accounting standards) without retraining the model.

[0070] Step 130: Use the rule question string as the input parameter of the financial early warning threshold retrieval function, and then obtain the corresponding financial early warning rule based on the preset financial early warning vector library.

[0071] This step can be specifically described as follows:

[0072] The FinancialWarningThresholdSearch function takes a rule question string as input and returns the financial warning rules. The technical implementation of this function is the same as step 110, which is an enhanced retrieval process (RAG). The difference is the scope of stored documents. This function retrieves a vector library (a preset financial warning vector library) by slicing and storing some financial warning rule texts in a vector database. It then retrieves the most relevant segments from the vector database and calls a large model to generate the final knowledge result. For example, the original document might be some explanations and principles regarding debt-to-equity ratio warnings for certain industries. The final result is the debt-to-equity ratio and current ratio warning lines for a certain type of enterprise, summarized by the large model.

[0073] Specifically, it includes:

[0074] Using RAG-enhanced retrieval technology, semantic similarity comparisons are performed from a pre-defined financial warning vector library to recall the slices with the highest similarity.

[0075] Based on the above description, those skilled in the art will understand that this step achieves semantic search through a vectorized financial early warning rule base, avoiding the rigid matching of traditional rule engines (such as the threshold differences for "debt-to-equity ratio early warning" in different industries). Example: When a user asks, "What is the current ratio early warning line for the retail industry?", the system can accurately recall relevant industry-specific rule fragments.

[0076] In addition, this step involves a second processing of the search results by the large model, transforming lengthy rule documents (such as "A warning is needed when the debt-to-asset ratio > 70%...") into concise conclusions (such as "Warning line for debt-to-asset ratio in the manufacturing industry: 70%)". This avoids the complexity of directly returning the original clauses and improves operability.

[0077] In addition, the solution involved in this step supports returning differentiated warning thresholds by industry (such as manufacturing and retail), which solves the problem of the disconnect between general rules and actual business.

[0078] In addition, the function FinancialWarningThresholdSearch shares the RAG framework with step 110, and only the vector library needs to be replaced to adapt to different scenarios (such as financial knowledge retrieval → early warning rule retrieval).

[0079] Step 140: Use the query question string as the input parameter of the financial data SQL generation function, and then use the large model to generate the corresponding query sub-SQL based on the database table information.

[0080] In some embodiments, this step may specifically be as follows:

[0081] The FinancialDataSQLGeneration function generates SQL queries from a single table. It takes a query question string as input and returns a sub-SQL query. Technically, this function uses the input parameter description and database table information to create a prompt template, which is then used to generate the required sub-SQL query by calling a large language model.

[0082] Understandably, this step utilizes Natural Language to SQL (Text-to-SQL) technology to achieve intelligent querying of financial data, with the following core benefits:

[0083] 1. Users do not need to master SQL syntax. They can generate accurate query statements using natural language (such as "query sales revenue in 2024"), improving the efficiency of non-technical personnel.

[0084] 2. This step can generate SQL by combining real-time table structure information, avoiding hard-coded query logic and adapting to the differences in table fields of different financial systems (such as "revenue" may correspond to the revenue or income field).

[0085] 3. This step can limit the generation of single-table query sub-SQL, reducing the risk of complex table joins, and at the same time constrain the model output through prompt word templates to prevent the generation of illegal statements (such as DELETE operations).

[0086] 4. The FinancialDataSQLGeneration function involved in this step can be reused in other data query scenarios (such as inventory, HR), only requiring an update to the table information prompt template.

[0087] Step 150: Use the initial financial problem data and query sub-SQL as input parameters to the financial statement data query function; then, using the initial financial problem data, the large model integrates all the corresponding query sub-SQL into the final query SQL; the execution engine executes the final query SQL to obtain the query results.

[0088] This step can be specifically described as follows:

[0089] The FinancialReportDataQuery function takes the original question description string and the query sub-SQL generated in step 140 as input parameters. This function generates complex queries using the reference SQL and returns the query results. The technical implementation principle involves combining the original question, relevant knowledge, and the query sub-SQL generated in previous steps to form a prompt template. This template is then used to call a larger model to generate the final data query SQL. The final query SQL is executed by the BI (Business Intelligence) engine to obtain the query results.

[0090] Understandably, this step utilizes multi-step SQL integration and execution technology to achieve intelligent and complex queries of financial data, with the following core benefits:

[0091] 1. Integrate scattered sub-SQL queries (such as single-table queries) into cross-table join queries through a large model, solving the problem of errors that are prone to occur when manually writing complex SQL (such as errors in multi-table JOIN logic).

[0092] 2. This step can combine the original problem description and sub-SQL to generate the final SQL, ensuring that the query results match the user's expectations. Figure 1 (For example, "Year-on-Year Revenue Growth Rate" needs to be linked to historical data tables).

[0093] 3. This step executes the optimized SQL through the BI engine, avoiding the performance bottleneck of manually concatenating SQL (such as slow queries caused by excessive nesting of subqueries).

[0094] 4. This step automates the entire process from natural language input to final data output, and is suitable for generating dynamic reports (such as quarterly financial analysis dashboards).

[0095] Step 160: Use financial knowledge, financial early warning rules, and query results as input parameters to the financial data summary function to obtain prompt words composed of financial knowledge, financial early warning rules, and query results. Input the prompt words into the large model to obtain the returned summary string.

[0096] This step can be specifically described as follows:

[0097] The FinancialDataSummary function takes the database query results from step 140, the results from step 110, and the results from step 120 as input parameters. These parameters are combined with built-in financial-related summary prompts to form the final prompt. This prompt, along with the input parameter structure, is then used to call the larger model. The larger model API will then return a summary text string. This function is used to summarize and analyze the obtained data, rules, and warning values.

[0098] The method also includes subsequent demonstration and analysis, the specific process of which can be as follows:

[0099] The initial financial problem data and query results are used as input parameters to the financial chart generation function to obtain chart generation prompts containing the initial financial problem data and query results. The chart generation prompts are then input into the large model to obtain a preset display chart containing the initial financial problem data and query results.

[0100] By combining the preset prompts corresponding to the displayed chart and calling the larger model, a chart object described in JSON format is returned. This JSON can be input into a BI tool to be rendered as a chart.

[0101] As described above, this embodiment achieves an end-to-end closed loop from natural language processing to structured knowledge, data querying, and visualization through a modular and intelligent financial data processing workflow. The core beneficial effects are as follows:

[0102] 1. Problem decomposition and semantic understanding (step 110):

[0103] Precise semantic segmentation: It breaks down complex financial problems into three sub-problems: financial knowledge, rules, and queries, avoiding the limitations of traditional single models in handling multimodal needs.

[0104] Dynamic adaptability: Supports the generation of multiple query question strings (such as "query debt-to-equity ratio" + "query current ratio") to adapt to complex analysis scenarios.

[0105] Reusable NLP framework: Based on the decomposition logic of pre-trained large models (such as Transformer), it can be transferred to the analysis of problems in other domains.

[0106] 2. Intelligent Retrieval of Financial Knowledge (Step 120):

[0107] Semantic Search: Deep semantic matching of financial knowledge base is achieved through RAG technology to solve the ambiguity problem of keyword retrieval (such as the distinction between "point-in-time" and "period").

[0108] Knowledge distillation: The large model summarizes the retrieved fragments into structured knowledge (such as "debt-to-equity ratio = total liabilities / total assets"), reducing the user's understanding cost.

[0109] Independent update mechanism: The knowledge base and model are decoupled, which supports rapid updates of the knowledge base when accounting standards change without retraining the model.

[0110] 3. Dynamic matching of financial early warning rules (step 130):

[0111] Industry-specific adaptation: Returns customized warning thresholds by industry (e.g., debt-to-asset ratio ≤ 60% for manufacturing, ≤ 50% for retail) to avoid misjudgments based on general rules.

[0112] Simplifying rules: condensing lengthy clauses (such as "a warning is required if the current ratio is less than 1 for three consecutive months") into actionable conclusions to improve business operability.

[0113] Shared RAG framework: Reuse the technical framework with step 120 to reduce development costs.

[0114] 4. Natural Language to SQL (Step 140):

[0115] Zero-code query: Business personnel can directly generate SQL using natural language (such as "query the gross profit margin for the third quarter") without IT support.

[0116] Dynamic table structure adaptation: Combines real-time database table information to automatically map field names (such as "Revenue" → revenue or income) to adapt to multiple financial systems.

[0117] Secure and controllable: Limit queries to a single table to avoid performance risks and data leaks caused by complex table joins.

[0118] 5. Complex SQL Integration and Execution (Step 150):

[0119] Automated cross-table join queries: Integrate single-table sub-SQL into cross-table queries (e.g., "year-on-year revenue growth rate" requires joining current and historical data tables), reducing manual concatenation errors.

[0120] meaning Figure 1 Consistency guarantee: Combine the original question with sub-SQL to generate the final SQL, ensuring that the result meets the user's requirements.

[0121] BI Engine Optimization: Execute optimized SQL through BI tools to improve query efficiency (such as avoiding slow queries).

[0122] 6. Data Summary and Visualization (Step 160):

[0123] Multimodal output: Integrates financial knowledge, rules, and query results into structured summaries (such as "Q3 gross margin decreased by 5%, below the industry warning line") to assist in decision-making.

[0124] Intelligent chart generation: Automatically converts query results into visual charts (such as line charts to show revenue trends), lowering the barrier to data interpretation.

[0125] Standardized JSON output: Supports direct rendering by BI tools, enabling rapid sharing and reuse of analysis results.

[0126] In addition, this application Figure 2 This application provides a financial data analysis system based on a large model. For example... Figure 2 As shown in the embodiments of this application, the system mainly includes:

[0127] The segmentation module 210 is used to obtain the initial financial problem data and decompose the initial financial problem data into financial problem strings, rule problem strings, and query problem strings.

[0128] The knowledge acquisition module 220 is used to take the string of financial questions as the input parameter of the financial business knowledge retrieval function, and then obtain the corresponding financial knowledge based on the preset financial knowledge vector library;

[0129] The knowledge acquisition module 220 includes knowledge acquisition units.

[0130] This is used to store financial knowledge documents into a preset financial knowledge vector library by slicing and vectorizing them.

[0131] The financial business knowledge retrieval function semantically retrieves slices from a pre-defined financial knowledge vector library based on the financial question string; it then calls the large model to process the slices into financial knowledge based on the financial question string.

[0132] The rule acquisition module 230 is used to take the rule question string as the input parameter of the financial early warning threshold retrieval function, and then obtain the corresponding financial early warning rule based on the preset financial early warning vector library;

[0133] The sub-SQL acquisition module 240 is used to take the query question string as the input parameter of the financial data SQL generation function, and then generate the corresponding query sub-SQL based on the database table information and the large model.

[0134] The result acquisition module 250 is used to take the initial financial problem data and query sub-SQL as input parameters of the financial statement data query function; then, using the initial financial problem data, the large model integrates all the corresponding query sub-SQL into the final query SQL; the execution engine executes the final query SQL to obtain the query results;

[0135] Return module 260 is used to take financial knowledge, financial early warning rules, and query results as input parameters of the financial data summary function, and then obtain prompt words composed of financial knowledge, financial early warning rules, and query results. Input the prompt words into the large model to obtain the returned summary string.

[0136] The system also includes a display module.

[0137] This function takes initial financial problem data and query results as input parameters to generate a financial chart, obtains chart generation prompts containing the initial financial problem data and query results, inputs the chart generation prompts into a large model, and obtains a preset display chart containing the initial financial problem data and query results.

[0138] In addition, embodiments of this application also provide a non-volatile computer storage medium storing executable instructions, which, when executed, implement the above-described financial data analysis method based on a large model.

[0139] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A financial data analysis method based on a large model, characterized in that, The method includes: Obtain initial financial problem data and break it down into financial problem strings, rule problem strings, and query problem strings; The financial question string is used as the input parameter of the financial business knowledge retrieval function, and then the corresponding financial knowledge is obtained based on the preset financial knowledge vector library; The rule question string is used as the input parameter of the financial early warning threshold retrieval function, and then the corresponding financial early warning rule is obtained based on the preset financial early warning vector library; The query string is used as the input parameter of the financial data SQL generation function, and then the corresponding query sub-SQL is generated based on the database table information and the large model. The initial financial problem data and query sub-SQL are used as input parameters for the financial statement data query function; then, using the initial financial problem data, the large model integrates all the corresponding query sub-SQL into the final query SQL; the execution engine executes the final query SQL to obtain the query results; Financial knowledge, financial early warning rules, and query results are used as input parameters to the financial data summary function, thereby obtaining prompt words composed of financial knowledge, financial early warning rules, and query results. The prompt words are then input into the large model to obtain the returned summary string.

2. The financial data analysis method based on a large model according to claim 1, characterized in that, The financial question string is used as the input parameter of the financial business knowledge retrieval function, and then the corresponding financial knowledge is obtained based on a pre-set financial knowledge vector library, specifically including: Financial knowledge documents are sliced ​​and vectorized and stored in a pre-defined financial knowledge vector library; The financial business knowledge retrieval function semantically retrieves slices from a pre-defined financial knowledge vector library based on the financial question string; it then calls the large model to process the slices into financial knowledge based on the financial question string.

3. The financial data analysis method based on a large model according to claim 1, characterized in that, The rule question string is used as the input parameter of the financial early warning threshold retrieval function, and then the corresponding financial early warning rules are obtained based on the preset financial early warning vector library, specifically including: The financial early warning rule document is sliced ​​and vectorized and stored in a preset financial early warning vector library; The financial early warning threshold retrieval function semantically retrieves slices from a preset financial early warning vector library based on the rule question string; it then calls the large model to process the slices into financial early warning rules based on the rule question string.

4. The financial data analysis method based on a large model according to claim 3, characterized in that, The financial early warning threshold retrieval function semantically retrieves and recalls slices from a pre-set financial early warning vector library based on rule question strings, specifically including: Using RAG-enhanced retrieval technology, semantic similarity comparisons are performed from a pre-defined financial warning vector library to recall the slices with the highest similarity.

5. The financial data analysis method based on a large model according to claim 1, characterized in that, The query string is used as the input parameter to the financial data SQL generation function. Based on the database table information, the corresponding query sub-SQL is generated using a large model, specifically including: Input the query question string and database table information into the preset prompt template to obtain SQL prompt words. Input the SQL prompt words into the large model to obtain the corresponding query sub-SQL.

6. The financial data analysis method based on a large model according to claim 1, characterized in that, The method further includes: The initial financial problem data and query results are used as input parameters to the financial chart generation function to obtain chart generation prompts containing the initial financial problem data and query results. The chart generation prompts are then input into the large model to obtain a preset display chart containing the initial financial problem data and query results.

7. A financial data analysis system based on a large model, characterized in that, The system includes: The segmentation module is used to obtain the initial financial problem data and break it down into financial problem strings, rule problem strings, and query problem strings. The knowledge acquisition module is used to take the string of financial questions as the input parameter of the financial business knowledge retrieval function, and then obtain the corresponding financial knowledge based on the preset financial knowledge vector library. The rule acquisition module is used to take the rule question string as the input parameter of the financial early warning threshold retrieval function, and then obtain the corresponding financial early warning rule based on the preset financial early warning vector library; The sub-SQL acquisition module is used to take the query question string as the input parameter of the financial data SQL generation function, and then generate the corresponding query sub-SQL based on the database table information and the large model. The result acquisition module is used to take the initial financial problem data and query sub-SQL as input parameters for the financial statement data query function; then, using the initial financial problem data, the large model integrates all the corresponding query sub-SQL into the final query SQL; the execution engine executes the final query SQL to obtain the query results. The return module is used to take financial knowledge, financial early warning rules, and query results as input parameters to the financial data summary function, and then obtain prompt words composed of financial knowledge, financial early warning rules, and query results. Input the prompt words into the large model to obtain the returned summary string.

8. The financial data analysis system based on a large model according to claim 7, characterized in that, The knowledge acquisition module includes knowledge acquisition units. This is used to store financial knowledge documents into a preset financial knowledge vector library by slicing and vectorizing them. The financial business knowledge retrieval function semantically retrieves and recalls slices from a pre-set financial knowledge vector library based on the string of financial questions. The large model is invoked, and the financial question string is sliced ​​and processed into financial knowledge.

9. The financial data analysis system based on a large model according to claim 7, characterized in that, The system also includes a display module. This function takes initial financial problem data and query results as input parameters to generate a financial chart, obtains chart generation prompts containing the initial financial problem data and query results, inputs the chart generation prompts into a large model, and obtains a preset display chart containing the initial financial problem data and query results.

10. A non-volatile computer storage medium, characterized in that, It stores computer instructions, which, when executed, implement a financial data analysis method based on a large model as described in any one of claims 1-6.

Images