Business potential target mining method and device based on large model, equipment and medium

Abstract

The application discloses a business potential target mining method and device based on a large model, equipment and medium, relates to the technical field of artificial intelligence and data mining, and can quickly and objectively output two types of refined mining results of explicit demand judgment and implicit demand judgment. The method comprises the following steps: obtaining standard business data obtained by preprocessing multi-source heterogeneous original business data through multiple stages; mapping the text data in the standard business data and the business knowledge base into semantic vectors through a text embedding model; in response to a business potential target mining request, generating a query vector based on the business characteristics of the target to be mined, recalling relevant knowledge fragments in the vector database by using a hybrid retrieval strategy; inputting the standard business data, the query vector and the recalled knowledge fragments into a preset large language model, guiding the large language model to perform logical reasoning based on the scoring rules in the dynamic rule base through a preset prompt word template, and outputting the mining result of the business potential target.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and data mining technology, and in particular to a method, apparatus, device and medium for mining potential business targets based on a large model. Background Technology

[0002] With increasing global market volatility and a more complex international trade environment, various uncertainties have significantly impacted the operations of real economy enterprises. This has led to continuous upgrades in business models and spurred a surge in diversified business demands, creating vast opportunities for business expansion for various market players. Accurate identification of potential business targets is a core prerequisite for market players to seize business opportunities and achieve large-scale development. Currently, there is an urgent need for efficient and intelligent technological means to address the shortcomings of traditional target identification methods and improve the efficiency and accuracy of identifying potential business targets.

[0003] In existing technologies, the discovery of potential business targets is mainly based on manual operation. It adopts a manual mining model that combines information retrieval and experience judgment. This involves manually reviewing publicly disclosed documents of enterprises, tracking industry market dynamics, integrating various relevant information of enterprises, and conducting targeted communication. Based on the experience of practitioners, potential targets are classified, prioritized, and their needs are verified. Finally, a traceable list of potential business targets is formed, which provides basic support for subsequent business docking.

[0004] However, existing technologies have significant shortcomings, failing to meet the demands for high efficiency and accuracy in identifying potential business targets. Firstly, the data is fragmented and massive, scattered across different information platforms, resulting in extremely low efficiency for manual integration and difficulties in cross-scenario information synchronization with insufficient real-time performance. Secondly, target selection relies on the subjective experience of practitioners, lacking scientific quantitative analysis methods, leading to low selection accuracy and effective identification rates, with a large amount of manpower wasted on ineffective operations. Furthermore, limitations in human semantic understanding make it difficult to identify implicit needs and match differentiated requirements. In addition, the entire process suffers from poor collaboration, with manual recording prone to errors, untimely dynamic tracking, and a lack of standardized traceability mechanisms. In summary, existing manual target discovery methods lack data-driven quantitative analysis methods and deep semantic understanding capabilities, making it difficult to overcome the inherent shortcomings of manual operations in terms of efficiency, accuracy, and collaboration, and unable to effectively solve the problems of accurate identification and real-time response to potential targets in complex scenarios. Summary of the Invention

[0005] In view of this, this application provides a method, apparatus, device and medium for mining potential business targets based on a large model. The main purpose is to solve the problem that the existing manual mining mode lacks data-driven quantitative analysis methods and deep semantic understanding capabilities, making it difficult to overcome the inherent defects of manual operation in terms of efficiency, accuracy and collaboration, and unable to effectively solve the problem of accurate identification and real-time response of potential targets in complex scenarios.

[0006] The first aspect provides a method for mining potential business targets based on a large model, which includes: Standard business data is obtained by acquiring heterogeneous raw business data from multiple sources and preprocessing it through multiple stages. Based on a pre-built business knowledge base, the standard business data and the text data in the business knowledge base are mapped into semantic vectors through a text embedding model and stored in a vector database. In response to a request to mine potential business targets, a query vector is generated based on the business characteristics of the target to be mined. A hybrid retrieval strategy is used to recall relevant knowledge fragments in the vector database. The hybrid retrieval strategy includes keyword matching and semantic similarity retrieval. The standard business data, the query vector, and the recalled knowledge fragments are input into a preset large language model. The large language model is guided by a preset prompt word template to perform logical reasoning based on the scoring rules in the dynamic rule base, and output the mining results of potential business targets. The mining results include explicit demand determination results or implicit demand determination results.

[0007] Secondly, a business potential target mining device based on a large model is provided, the device comprising: The acquisition unit is used to acquire standard business data obtained by multi-stage preprocessing of multi-source heterogeneous raw business data; A storage unit is used to map the standard business data and the text data in the business knowledge base into semantic vectors through a text embedding model based on a pre-built business knowledge base, and store them in a vector database. The generation unit is used to respond to a business potential target mining request, generate a query vector based on the business characteristics of the target to be mined, and use a hybrid retrieval strategy to recall relevant knowledge fragments in the vector database. The hybrid retrieval strategy includes keyword matching and semantic similarity retrieval. The reasoning unit is used to input the standard business data, the query vector, and the recalled knowledge fragments into a preset large language model, and guide the large language model to perform logical reasoning based on the scoring rules in the dynamic rule base through preset prompt word templates, and output the mining results of potential business targets; the mining results include explicit demand determination results or implicit demand determination results.

[0008] Thirdly, a business potential target mining device based on a large model is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor. When the processor executes the program, it implements the above-mentioned business potential target mining method based on a large model.

[0009] Fourthly, a storage medium is provided on which a computer program is stored, which, when executed by a processor, implements the above-described method for mining potential business targets based on a large model.

[0010] By employing the above technical solutions, this application provides a method, apparatus, device, and medium for mining potential business targets based on a large model. Compared with existing methods that use manual mining to achieve potential business target mining, this application obtains standard business data obtained from multi-source heterogeneous raw business data through multi-stage preprocessing; based on a pre-built business knowledge base, the standard business data and text data in the business knowledge base are mapped into semantic vectors through a text embedding model and stored in a vector database; in response to a potential business target mining request, a query vector is generated based on the business characteristics of the target to be mined, and relevant knowledge fragments are recalled in the vector database using a hybrid retrieval strategy, which includes keyword matching and semantic similarity retrieval; the standard business data, query vectors, and recalled knowledge fragments are input into a preset large language model, and the large language model is guided by preset prompt word templates to perform logical reasoning based on scoring rules in a dynamic rule base, outputting the mining results of potential business targets; the mining results include explicit demand determination results or implicit demand determination results. The entire process transforms multi-source heterogeneous raw business data into high-quality standard business data through multi-stage preprocessing, effectively eliminating format heterogeneity and business noise. Then, relying on a text embedding model, the standard business data and business knowledge base text are uniformly transformed into semantic vectors and stored in the database, realizing the semantic fusion and efficient aggregation of heterogeneous business information. At the same time, a hybrid retrieval strategy combining keyword matching and semantic similarity retrieval is used to carry out knowledge retrieval. This retains the retrieval reliability of accurate keyword matching and can effectively reach deep-related knowledge through semantic mining, significantly improving the comprehensiveness and accuracy of the retrieval of target related knowledge fragments. Subsequently, the pre-set prompt word template guides the large language model to complete intelligent logical reasoning in combination with the scoring rules of the dynamic rule base. This avoids the defects of large subjective bias, low efficiency and inconsistent standards of manual analysis, and can quickly and objectively output two types of refined mining results: explicit demand judgment and implicit demand judgment. This comprehensively enhances the recognition accuracy and response processing efficiency of potential targets in complex business scenarios.

[0011] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0012] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating a business potential target mining method based on a large model in one embodiment of this application; Figure 2 yes Figure 1 A flowchart illustrating a specific implementation method of step 101; Figure 3 yes Figure 1 A flowchart illustrating a specific implementation method for step 102; Figure 4 yes Figure 1 A flowchart illustrating a specific implementation method for step 103; Figure 5 yes Figure 1 A flowchart illustrating a specific implementation method for step 104; Figure 6 This is a schematic diagram of the structure of a business potential target mining system based on a large model in one embodiment of this application; Figure 7 This is a schematic diagram of the structure of a business potential target mining device based on a large model in one embodiment of this application; Figure 8 This is a schematic diagram of the device structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0013] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0014] In existing technologies, the discovery of potential business targets mainly relies on manual operation, employing a manual discovery model that combines information retrieval with experience-based judgment. However, existing manual discovery models lack data-driven quantitative analysis methods and deep semantic understanding capabilities, making it difficult to overcome the inherent shortcomings of manual operation in terms of efficiency, accuracy, and collaboration. Consequently, they cannot effectively solve the problems of accurate identification and real-time response to potential targets in complex scenarios.

[0015] To address this issue, this embodiment provides a method for mining potential business targets based on a large model, such as... Figure 1 As shown, it includes the following steps: 101. Obtain standard business data by preprocessing multi-stage raw business data from multiple heterogeneous sources.

[0016] In this embodiment, the multi-source heterogeneous raw business data is collected from a distributed environmental monitoring network. This distributed environmental monitoring network includes a time-series numerical stream continuously output by a fixed sensor array, equipment operation logs generated by field inspection terminals, and scanned document images formed from manually observed records. The sensor array generates structured data packets containing coordinates, timestamps, and multi-dimensional environmental parameters every second; the equipment operation logs record equipment status and operation events in key-value pairs; and the scanned document images contain handwritten annotations and tabular content, presented in an unstructured text format.

[0017] The specific multi-stage preprocessing process for heterogeneous raw business data from multiple sources includes the following steps: First, format normalization is performed on the heterogeneous raw business data from multiple sources. Optical character recognition (OCR) technology is applied to scanned document images to convert them into editable character sequences, and layout analysis algorithms are used to reconstruct the paragraph's logical structure. Field mapping rules are used on device operation logs to extract key information such as device codes and event types into standardized data objects. Protocol parsing and unit verification are performed on time-series numerical streams. This stage completely eliminates format differences at the data source, forming a unified data representation framework.

[0018] Subsequently, deep cleaning and noise suppression were carried out on the heterogeneous original business data from multiple sources: residual markers and invalid special symbols in character sequences were accurately removed using regular expressions; non-valid observation content such as test identifiers and equipment self-test information was filtered out based on a preset rule base; the same source observation records that were repeatedly collected across collection nodes were identified and merged based on the text fingerprint algorithm; and meaningless character fragments and redundant descriptions unrelated to the business scenario were removed by combining text length threshold constraints and semantic relevance models. Invalid interference information was removed from multiple dimensions to fully ensure the purity, validity and business relevance of the final data content.

[0019] Next, the heterogeneous raw business data from multiple sources undergoes full-domain semantic standardization. The cleaned text is uniformly converted to a preset character encoding, standardizing paragraph breaks and line breaks; non-standard expressions are mapped to standard tags using a domain-specific terminology dictionary; and full-width punctuation is simultaneously converted to half-width format. This process consolidates scattered expressions into semantically consistent standard data units, significantly improving the semantic understanding accuracy of subsequent analysis.

[0020] Through the aforementioned multi-stage preprocessing steps, the original, scattered, and mixed-format raw business data is transformed into a high-quality standard dataset. For example, in the risk analysis of a company's hedging business, the system successfully integrated the heterogeneous inputs from twenty trading accounts. After eliminating format conflicts and noise interference, the accuracy of hedging risk pattern recognition was significantly improved.

[0021] 102. Based on a pre-built business knowledge base, the standard business data and the text data in the business knowledge base are mapped into semantic vectors through a text embedding model and stored in a vector database.

[0022] During the business knowledge base construction phase, it is necessary to integrate multi-source professional literature resources. These resources include, but are not limited to, the full text of national hedging business standards, historical hedging case analysis reports, hedging operation specification documents, and records of experience summaries from domain experts. All textual materials, after preliminary organization, form a well-structured knowledge set, covering professional topics such as hedging indicator thresholds, hedging strategy analysis logic, and risk transmission paths, providing an authoritative knowledge foundation for semantic understanding.

[0023] Standard domain data, preprocessed through multiple stages, is processed synchronously. This data originates from standardized business reports generated by the hedging business system, account status log text output from trading terminals, and manually recorded descriptions of hedging operations after character recognition and punctuation normalization. Each text has undergone terminology standardization; for example, "hedging" and "hedging transaction" are uniformly mapped to "hedging" to ensure consistency and semantic purity.

[0024] Specifically, in the process of mapping standard business data and text data from the business knowledge base into semantic vectors through a text embedding model, long documents are segmented into semantically complete independent units based on character length thresholds and syntactic dependency analysis results. For example, a standard hedging document containing multiple chapters is segmented into logical paragraphs, with each unit focusing on a single hedging indicator definition or judgment rule; the risk exposure description paragraphs in hedging business reports are also fully preserved as independent semantic units. Subsequently, a pre-trained deep text embedding model is used to vectorize and encode each unit. This model, based on a self-attention architecture, deeply captures lexical contextual relationships, accurately mapping text units into high-dimensional numerical vectors, allowing semantically similar expressions to be tightly clustered in the vector space.

[0025] Correspondingly, the generated semantic vectors are structurally bound to their corresponding text units, data source identifiers, and timestamp information to form complete data records. All records are persistently stored in a dedicated vector database, and a hierarchical, navigable small-world graph index is constructed based on vector dimensions and data scale. This index efficiently organizes the vector space, significantly improving subsequent retrieval efficiency. The storage process ensures that each vector can be accurately traced back to the original text content and metadata, providing technical support for knowledge tracing.

[0026] In practical applications, when a business description is received stating that "a trading company is experiencing persistent exchange rate fluctuations and an abnormally large increase in its hedging exposure," the vector database pre-stores relevant knowledge units. These units include causal analyses of similar historical events, correlation rules for hedging indicators, and records of hedging handling experience. Vectorized storage allows the business description to quickly locate semantically related knowledge.

[0027] 103. In response to a request to mine potential business targets, a query vector is generated based on the business characteristics of the target to be mined, and relevant knowledge fragments are retrieved from the vector database using a hybrid retrieval strategy.

[0028] Upon receiving a request to mine potential business targets, this embodiment first reads the business feature data corresponding to the target to be mined. This feature data includes basic attribute information, related data identifiers, and inherent scene descriptions within the corresponding business scenario, which can completely define the scope of the target to be analyzed and mined. Then, the read business feature data is uniformly input into a pre-trained text embedding model. Relying on the model's own semantic mapping capabilities, the business features described in natural language are uniformly transformed into structured query vectors with fixed dimensions and semantic associations, thereby realizing the digital and vectorized expression of the business features of the target to be mined.

[0029] After generating the query vector, a pre-defined hybrid retrieval mechanism will be activated, relying on the constructed vector database to retrieve related knowledge fragments. During the retrieval execution phase, two independent retrieval processes will run simultaneously. One is keyword matching retrieval, which extracts core feature words and fixed identifier words from the original business feature text that generated the query vector. The extracted words are then precisely compared with the original text content corresponding to each knowledge fragment in the vector database to filter out knowledge data whose text content contains core keywords. The other simultaneously executed process is semantic similarity retrieval, which compares the generated query vector one by one with the semantic vectors corresponding to various types of knowledge fragments pre-stored in the vector database. By calculating the spatial similarity between vectors, knowledge fragments whose semantic relevance meets a pre-defined threshold are selected. After both retrieval processes complete data filtering, the two sets of retrieval results will be integrated and categorized. Duplicate knowledge fragments will be removed, and effective content with both text keyword relevance and deep semantic relevance will be retained, resulting in a complete set of knowledge fragments that fits the current target features to be mined.

[0030] In practical applications, for cases where a monitoring point reports price fluctuations in the hedging target and a continuous expansion of risk exposure, a hybrid search strategy simultaneously activates two paths: keyword matching precisely locates standard document fragments containing terms such as hedging and risk exposure, while semantic search uncovers historical case records describing drastic fluctuations in the target asset price and the exacerbation of risk due to corporate inventory management or cash flow pressures. The two types of search results are merged and ranked according to preset weight coefficients, and the text semantic unit with the highest comprehensive score is selected as the recalled knowledge fragment. The merged knowledge fragment provides a basis for subsequent hedging risk assessment, ensuring the reliability of explicit keyword matching while extending to the depth of implicit semantic associations.

[0031] 104. Input the standard business data, the query vector, and the recalled knowledge fragments into a preset large language model. Guide the large language model to perform logical reasoning based on the scoring rules in the dynamic rule base through preset prompt word templates, and output the mining results of potential business targets.

[0032] The standard business data consists of a collection of structured and unstructured data that has undergone multiple stages of cleaning, denoising, and standardization, including various observation records, attribute information, and related data in the business scenario; the query vector is a semantic vector generated based on the business features of the target to be mined, which accurately corresponds to the core direction of the target to be mined; the recalled knowledge fragments are knowledge content that is highly relevant to the target to be mined and obtained through a hybrid retrieval strategy, covering various effective information such as the inherent rules of the business scenario, historical observation cases, and attribute relationships.

[0033] To ensure that the large language model can perform inference according to preset logic and avoid inference bias, a preset prompt word template is used to guide the model's inference process. This prompt word template pre-embeds the calling instructions, inference logic requirements, and result output specifications of the dynamic rule base. It guides the large language model to retrieve the corresponding scoring rules from the dynamic rule base and, in conjunction with the three types of input data, perform standardized logical inference. The guidance logic of the prompt word template aligns with business scenario requirements, requiring the model to first parse the core information of the standard business data, then combine it with the target features to be mined corresponding to the query vector to match relevant content in the recalled knowledge fragments, and finally, complete multi-dimensional inference analysis based on the scoring rules in the dynamic rule base.

[0034] Taking a hedging scenario as an example, the target to be mined is the potential risk characteristics corresponding to a company's hedging business data. The standard business data input to the large language model consists of pre-processed data on the company's transaction prices, position sizes, and risk exposures over a period of time. This data is formatted uniformly and free of redundant noise. The query vector is a semantic vector generated based on the objective of "mining corporate hedging risk characteristics," precisely targeting risk correlation analysis within the hedging data. The recalled knowledge fragments include hedging risk assessment criteria, historical hedging risk cases of similar companies, and correlation thresholds for various types of hedging business data. Guided by the prompt word template, the large language model retrieves the scoring rules for hedging risk assessment from the dynamic rule base. It first analyzes the numerical trends of various hedging indicators in the standard business data, then combines this with the risk assessment criteria in the recalled knowledge fragments, compares them with the risk mining requirements corresponding to the query vector, and gradually conducts logical reasoning.

[0035] During the inference process, the large language model scores each observation indicator in the standard business data according to the scoring rules in the dynamic rule base. It then analyzes the correlations between the indicators based on historical cases in the recalled knowledge fragments to determine if any abnormal features exist, distinguishing between explicit and implicit anomalies. For example, if the pollutant concentration in the standard business data exceeds the threshold set in the scoring rules for a certain period, and this is combined with the normal concentration range under similar climatic conditions in the recalled knowledge fragments, it is determined that the pollutant concentration for that period is explicitly abnormal. Conversely, if the values ​​of all indicators in the standard business data are within the normal threshold range, but the changing trends of different indicators show correlation patterns similar to historical anomaly cases, and this is combined with the correlation analysis criteria in the scoring rules, it is determined that there is a risk of implicit anomalies.

[0036] After inference is complete, the large language model will output the results of mining potential business objectives according to the preset output specifications of the prompt word template. The mining results include two categories: explicit demand determination and implicit demand determination. Explicit demand determination corresponds to clearly identifiable abnormal features, including specific abnormal indicators, abnormal time periods, and judgment criteria. Implicit demand determination corresponds to potential abnormal risks that require deep correlation analysis to identify, including the related factors of the abnormal risks, possible subsequent changes, and preliminary judgment logic. The entire inference process requires no manual intervention, relying entirely on the model's intelligent reasoning capabilities, combined with standardized scoring rules from a dynamic rule base, to ensure the accuracy, standardization, and objectivity of the output results.

[0037] The business potential target mining method based on a large model provided in this application, compared with the existing method of using manual mining to achieve business potential target mining, obtains standard business data by preprocessing multi-source heterogeneous original business data through multiple stages; based on a pre-built business knowledge base, the standard business data and text data in the business knowledge base are mapped into semantic vectors through a text embedding model and stored in a vector database; in response to the business potential target mining request, a query vector is generated based on the business characteristics of the target to be mined, and relevant knowledge fragments are recalled in the vector database using a hybrid retrieval strategy, which includes keyword matching and semantic similarity retrieval; the standard business data, query vectors, and recalled knowledge fragments are input into a preset large language model, and the large language model is guided by a preset prompt word template to perform logical reasoning based on the scoring rules in the dynamic rule base, and output the mining results of the business potential target; the mining results include explicit demand determination results or implicit demand determination results. The entire process transforms multi-source heterogeneous raw business data into high-quality standard business data through multi-stage preprocessing, effectively eliminating format heterogeneity and business noise. Then, relying on a text embedding model, the standard business data and business knowledge base text are uniformly transformed into semantic vectors and stored in the database, realizing the semantic fusion and efficient aggregation of heterogeneous business information. At the same time, a hybrid retrieval strategy combining keyword matching and semantic similarity retrieval is used to carry out knowledge retrieval. This retains the retrieval reliability of accurate keyword matching and can effectively reach deep-related knowledge through semantic mining, significantly improving the comprehensiveness and accuracy of the retrieval of target related knowledge fragments. Subsequently, the pre-set prompt word template guides the large language model to complete intelligent logical reasoning in combination with the scoring rules of the dynamic rule base. This avoids the defects of large subjective bias, low efficiency and inconsistent standards of manual analysis, and can quickly and objectively output two types of refined mining results: explicit demand judgment and implicit demand judgment. This comprehensively enhances the recognition accuracy and response processing efficiency of potential targets in complex business scenarios.

[0038] In practical applications, to address issues such as inconsistent formats, noise redundancy, and non-standard representation of multi-source heterogeneous raw business data, a phased preprocessing approach is necessary. This involves first collecting data and adapting its format, then cleaning the data to remove invalid interference, and finally standardizing the data to obtain compliant standard business data. Specifically, for example... Figure 2 As shown, step 101 above includes the following steps: 201. Use a pre-built heterogeneous data adaptation gateway to obtain multi-source heterogeneous raw business data from distributed nodes, perform format preprocessing on the multi-source heterogeneous raw business data, and obtain preprocessed business data.

[0039] 202. Perform hybrid noise filtering on the preprocessed business data to remove structural markers, non-business symbols and invalid formatted text under preset business scenarios, and use fingerprint comparison technology based on hash features combined with semantic relevance discrimination model to perform deduplication and merging and low-quality data removal to obtain purified business data.

[0040] 203. Perform character encoding unification and typesetting symbol standardization processing on the purified business data, and rely on the industry professional terminology knowledge graph to use string matching and normalization algorithms to convert non-standard business expressions and punctuation marks into standard formats to obtain standard business data.

[0041] In this embodiment, the heterogeneous data adaptation gateway, as an access component for data preprocessing, can achieve fully automated access to multiple types and formats of data sources through various methods such as API calls, intelligent web crawling, and direct database connections, covering data needs across all scenarios. Distributed nodes correspond to deployment nodes for various data sources, including enterprise public information release nodes, market information release nodes, and industry data release nodes. Enterprise public information release nodes include temporary announcements from listed companies, such as risk warnings, inventory impairment notices, and announcements regarding hedging business approvals / progress / adjustments, as well as periodic reports, such as the complete text of annual / semi-annual / quarterly reports and their split financial notes and operational analysis sections. They also include investor interaction platforms, such as the original questions and detailed responses from companies on platforms like Panorama.com and Shanghai Stock Exchange e-interaction. Market information release nodes cover futures product databases from the three major domestic futures exchanges and international exchanges such as LME and COMEX. Industry data release nodes include futures raw material supply chain maps and enterprise operational statistics released by industry associations. The raw business data output by each node exhibits multi-source heterogeneous characteristics, including both structured and unstructured data. For structured data, field mapping is used for access, while unstructured data is converted into a processable format through OCR recognition and text extraction technology. After the unified acquisition of raw business data is completed, the format of various types of raw business data is parsed and converted, and the data of different formats is uniformly converted into an intermediate format to obtain the pre-processed business data.

[0042] Specifically, during the process of performing mixed noise filtering on the pre-processed business data, a combination of regular expression matching and semantic recognition is used to batch remove structural tags and non-business symbols from the data, including HTML tags, XML tags, and special symbols. Simultaneously, invalid formatted text in preset business scenarios is filtered out, such as various formatted disclaimers in corporate announcements and other content irrelevant to the core topic. Subsequently, a hash-based fingerprint comparison technology is used to perform a full-domain comparison of all pre-processed data, identifying and merging duplicate data such as different publishing channels for the same content or multiple submissions of the same question, avoiding redundant calculations. At the same time, a semantic relevance discrimination model is used to perform semantic analysis on the data content. By setting a text length threshold, meaningless character sequences below the threshold and abnormal documents exceeding the threshold are removed, while irrelevant and invalid content is filtered out, completing the removal of low-quality data and ensuring the validity and purity of the data, resulting in purified business data.

[0043] Specifically, during the process of unifying character encoding and standardizing typesetting symbols in the purified business data, at the encoding level, all text is uniformly converted to a preset character encoding format, for example, all text is uniformly converted to UTF-8 encoding; at the format level, paragraph separators and line break logic are standardized, for example, text line breaks and paragraph separators are unified, and documents of different formats such as PDF and Word are converted into standard TXT text; at the content level, relying on the knowledge graph of professional terminology in the environmental field, string matching and normalization algorithms are used to accurately map non-standard business expressions to standard business tags, for example, unifying "HEDGING" into a standard Chinese expression; at the same time, the conversion of full-width punctuation to half-width format is completed to eliminate symbol ambiguity and obtain standard business data, for example, unifying full-width question marks and half-width question marks.

[0044] In practical application scenarios, raw business data includes text-based unstructured data, semi-structured log data, and structured numerical data. Specifically, in the process of format preprocessing of multi-source heterogeneous raw business data, for text-based unstructured data, optical character recognition and document parsing technologies are used to convert portable document formats and text processing document formats into editable character streams; for semi-structured log data and structured numerical data, character mapping and key-value pair extraction technologies are used to convert them into unified data objects to eliminate the format heterogeneity of the data source.

[0045] The aforementioned text-based unstructured data includes the full text of temporary reports issued by monitoring stations, parameter notes and environmental analysis sections extracted from periodic monitoring reports, and the original text of questions submitted by the public observation platform and detailed station responses; the semi-structured log data originates from equipment operation logs generated by mobile inspection terminals, recording equipment status and operation event sequences; the structured numerical data consists of a time-series numerical stream continuously output by a fixed sensor array, including coordinate information, timestamps, and multi-dimensional environmental parameter indicators.

[0046] For unstructured text-based data, optical character recognition (OCR) and document parsing technologies are used to convert original files in standard electronic document formats and text-editable document formats into editable character streams. After extracting text content from the monitoring report using OCR, document parsing algorithms are used to reconstruct the paragraph logical structure and table information, ensuring the integrity and parsability of the text content.

[0047] For semi-structured log data and structured numerical data, character mapping and key-value pair extraction techniques are used for transformation. Based on preset field mapping rules, the inspection terminal logs extract key information such as device codes and event types into unified data objects; the numerical stream output by the sensor array is converted into a standardized numerical sequence through protocol parsing and unit verification. This process systematically eliminates the format heterogeneity of the data source, forming a unified data representation framework with a clear structure and standardized fields.

[0048] Furthermore, to achieve traceability in data governance and facilitate subsequent auditing and issue tracing, after obtaining standard business data from multi-source heterogeneous raw business data and preprocessing it through multiple stages, the standard business data can be labeled with its data source, preprocessing time, and processing node metadata. A distributed message queue can then be used to push the preprocessed structured numerical data to a time-series database, and the preprocessed text-based unstructured data to a pre-defined large language model. Pushing the preprocessed structured numerical data to the time-series database via the distributed message queue supports the normal operation of downstream quantitative scoring models, while pushing the preprocessed text-based unstructured data to the pre-defined large language model provides suitable data support for subsequent risk signal extraction and implicit demand mining.

[0049] In practical applications, to achieve effective integration and efficient retrieval of standard business data and business knowledge base text, it is necessary to map the standard business data and text data in the business knowledge base into semantic vectors through a text embedding model and store them in a vector database, based on a pre-built business knowledge base. Specifically, for example... Figure 3 As shown, step 102 above includes the following steps: 301. Perform semantic segmentation processing on the standard business data and text data in the business knowledge base. Based on the preset character length threshold and the syntactic dependency analysis results, divide the continuous text into text semantic units that maintain semantic integrity.

[0050] 302. Call the pre-trained text embedding model to vectorize and encode each text semantic unit, generating a business semantic vector corresponding to the text semantic unit.

[0051] 303. The business semantic vector is structurally bound to its corresponding text semantic unit, data source identifier and timestamp metadata, and persistently stored in the vector database.

[0052] 304. Based on the dimensional parameters and data scale of the business semantic vector, construct a hierarchical navigable small world graph index or a product quantized inverted index to efficiently organize the vector space and accelerate retrieval optimization, so as to support the real-time response requirements of semantic similarity calculation in the hybrid retrieval strategy.

[0053] In this embodiment, the business knowledge base stores professional knowledge related to the business domain, including text data such as operating procedures, risk assessment standards, and historical operating cases. Standard business data includes preprocessed texts of listed company announcements, investor interaction Q&A records, and fragments of industry research reports. To avoid semantic loss during long text vectorization, continuous text is segmented into semantic units that maintain semantic integrity based on a preset character length threshold and syntactic dependency analysis results. For example, the text "Hedging operations need to be set according to the company's inventory size; when the inventory size is large, the hedging ratio can be appropriately increased to reduce price fluctuation risk" in the business knowledge base is divided into two semantic units based on syntactic dependency relations: "Hedging operations need to be set according to the company's inventory size" and "When the inventory size is large, the hedging ratio can be appropriately increased to reduce price fluctuation risk". Another example is the text "A listed company's Q1 2025 inventory impairment announcement shows that the inventory size increased by 30% compared to the same period last year, and it plans to initiate hedging operations to hedge price risk" in standard business data, which is divided into two semantic units to ensure that each unit is semantically complete and unambiguous.

[0054] The aforementioned text embedding model is based on a self-attention architecture and extracts semantic association features between words through a context-aware mechanism. For example, when encoding the text semantic unit "Setting operations requires setting hedging ratios in conjunction with the company's inventory scale," the text embedding model will perceive the semantic associations between "setting operations," "inventory scale," and "hedging ratio," generating a business semantic vector that can accurately represent this semantic. When encoding the text semantic unit "A listed company's inventory scale in Q1 2025 increased by 30% compared to the same period of the previous year" in standard business data, it will associate semantics such as "listed company," "time," "inventory scale," and "growth ratio," generating a corresponding business semantic vector to ensure that the business semantic vector can accurately map the text semantics and fit the semantic requirements of relevant business scenarios.

[0055] Furthermore, business semantic vectors are structurally bound to their corresponding text semantic units, data source identifiers, and timestamp metadata, and persistently stored in a vector database. This structural binding establishes a correspondence between semantic vectors and the original text and metadata, enabling the retrieval of text semantic unit content based on the recalled semantic vectors during subsequent searches. The text semantic unit content serves as the context for reasoning within a pre-defined large language model. For example, a business semantic vector representing operational specifications is bound to its corresponding text semantic unit, source identifier (business knowledge base), and timestamp; a business semantic vector representing the inventory scale of a listed company is bound to its corresponding text semantic unit, source identifier (listed company announcement), and timestamp. After binding, these are batch-persistently stored in the vector database, achieving ordered archiving of semantic vectors and aligning with the hierarchical and categorized storage requirements of the data storage module.

[0056] Specifically, based on the dimensional parameters and data scale of the business semantic vector, and combined with different business scenarios, two types of indexes can be constructed: hierarchical navigable small world graph indexes or product-quantized inverted indexes. Both hierarchical navigable small world graph indexes and product-quantized inverted indexes are constructed by adapting the dimensional parameters of the business semantic vectors to the actual data scale. The two indexing methods can be flexibly selected or combined according to real-time changes in the vector data.

[0057] For a hierarchical navigable small-world graph index, considering relevant business scenarios, the business semantic vectors generated by the text embedding model have a fixed dimension. However, as business progresses, the data scale continuously expands. For example, a large number of business semantic vectors corresponding to listed company announcements and investor Q&A records are added daily. Simultaneously, updates to the business knowledge base also add semantic vectors related to operational specifications and historical cases. If this massive amount of vector data is stored unordered, semantic similarity calculations during subsequent mixed searches will be too time-consuming, failing to meet real-time response requirements. Specifically, when constructing a hierarchical navigable small-world graph index, all vectors are first layered according to their semantic relevance based on the dimension parameters of the business semantic vectors. Vectors within each layer maintain high semantic similarity, and connections are established between different layers through associated nodes, forming a multi-level navigation structure. For example, business semantic vectors related to listed company inventory are grouped into one layer, and semantic vectors related to operational ratios are grouped into another layer. Adjacency relationships are established within each layer through vector similarity sorting, and rapid jumps between layers are achieved through core associated vectors. This hierarchical organization allows the retrieval process to quickly locate the level of the target vector without traversing all vectors, significantly improving retrieval efficiency.

[0058] For constructing a product-quantized inverted index, addressing the characteristics of high-dimensional business semantic vectors and massive data scale, the product-quantized inverted index reduces the computational load during retrieval, thus accelerating the search process. In specific implementation, firstly, based on the dimensional parameters of the business semantic vector, the high-dimensional vector is decomposed into multiple low-dimensional sub-vectors. Each low-dimensional sub-vector is quantized to generate a corresponding quantization code. Then, the quantization codes of all vectors are aggregated to construct an inverted index table, grouping vectors with the same quantization code for storage. For example, a high-dimensional business semantic vector representing inventory scale and operations in a listed company's announcement is decomposed into two low-dimensional sub-vectors: inventory-related and operation-related. Each sub-vector is quantized to generate a quantization code, and vectors with the same quantization code are grouped together and stored under the corresponding entry in the inverted index table. When subsequent mixed searches require semantic similarity calculations, the inverted index table can be used to quickly locate the vector group matching the quantization code of the query vector. Precise similarity calculations can then be performed within this group, eliminating the need to perform calculations on all vectors, effectively reducing computational complexity and ensuring real-time response in semantic similarity calculations.

[0059] In practical applications, to accurately acquire knowledge related to potential targets in hedging transactions and ensure that the retrieved knowledge fragments align with business terminology, a hybrid retrieval strategy is needed to retrieve relevant knowledge fragments from a vector database. Specifically, for example... Figure 4 As shown, step 103 above includes the following steps: 401. In response to a request to mine potential business targets, construct query text based on the structured business features and unstructured business description text of the target to be mined, call a pre-trained text embedding model to vectorize the query text, and generate a query vector aligned with the business semantic space.

[0060] 402. A hybrid retrieval strategy is adopted, simultaneously performing keyword matching retrieval based on inverted index and semantic retrieval based on vector similarity, generating keyword candidate sets containing text semantic unit identifiers and matching scores, and semantic candidate sets containing text semantic unit identifiers and similarity scores, respectively.

[0061] 403. Based on the preset keyword matching weight coefficient and semantic similarity weight coefficient, the corresponding text semantic units in the keyword candidate set and the semantic candidate set are weighted and fused to calculate the comprehensive score of each text semantic unit.

[0062] 404. Select a preset number of text semantic units as the recalled knowledge fragments in descending order of comprehensive score. The recalled knowledge fragments include the content of the text semantic units and their associated metadata.

[0063] In hedging scenarios, structured business features include quantitative information such as the type of hedging instrument, position size, and risk exposure, while unstructured business description text includes textual information such as hedging operation records, market analysis descriptions, and historical risk event explanations. These two types of information are integrated to form a complete query text. Subsequently, a pre-trained text embedding model is used to vectorize this query text, generating a query vector aligned with the semantic space of the hedging business.

[0064] Specifically, during the synchronous execution of the hybrid retrieval strategy in the vector database, keyword matching retrieval is performed on the metadata of text semantic units based on the inverted index to extract core terms from the query text. These terms are then matched against the metadata keyword database to generate a keyword candidate set containing text semantic unit identifiers and matching scores. Simultaneously, the cosine similarity between the query vector and all business semantic vectors in the database is calculated to generate a semantic candidate set containing semantically similar text semantic unit identifiers and similarity scores. These two retrieval paths run in parallel without interference, capturing explicit keyword associations and deep semantic associations respectively.

[0065] After the retrieval is completed, the corresponding text semantic units in the keyword candidate set and semantic candidate set are weighted and fused according to the preset keyword matching weight coefficient and semantic similarity weight coefficient. Among them, the keyword matching weight coefficient and semantic similarity weight coefficient are preset according to the actual needs of hedging business. The comprehensive score of each text semantic unit is obtained by weighted calculation. The comprehensive score can fully reflect the degree of correlation between the text semantic unit and the target to be mined.

[0066] Finally, a predetermined number of text semantic units are selected as the recalled knowledge fragments in descending order of their comprehensive scores. These recalled knowledge fragments include the content of the text semantic units and their associated metadata. In hedging scenarios, the text semantic unit content covers relevant knowledge such as hedging risk assessment standards, historical hedging case analysis, and market trend association rules. The associated metadata includes information such as the source of the text semantic unit, its creation time, and associated business scenarios, providing comprehensive knowledge support for the subsequent mining and analysis of potential targets in hedging transactions.

[0067] Furthermore, to improve the accuracy of the hybrid retrieval strategy, a dynamic weight adjustment mechanism is introduced to optimize the knowledge retrieval effect. This mechanism adjusts the keyword matching weight coefficient and semantic similarity weight coefficient in real time based on the industry volatility index of the target and the keyword density index of the business description text. The industry volatility index is obtained by querying the business knowledge base, and the keyword density index is calculated by statistically analyzing the frequency of occurrence of preset business keywords in the business description text, ensuring that the adjustment parameters are supported by objective data. Accordingly, the keyword matching weight coefficient and the semantic similarity weight coefficient are dynamically adjusted based on the industry volatility index of the target and the keyword density index of the business description text. Specifically, the industry volatility index is obtained by querying the business knowledge base, and the keyword density index is calculated by statistically analyzing the frequency of occurrence of preset business keywords in the business description text. When the industry volatility index is greater than a preset volatility threshold or the keyword density index is less than a preset density threshold, the semantic similarity weight coefficient is increased; when the keyword density index is greater than the preset density threshold, the keyword matching weight coefficient is increased. Weighted fusion is performed based on the dynamically adjusted weight coefficients to optimize the semantic relevance between the retrieved knowledge fragments and the potential business targets.

[0068] The aforementioned rule design enables the retrieval strategy to adapt to changes in business characteristics. It strengthens semantic association mining capabilities when industry fluctuations are significant or text keywords are sparse, and improves the accuracy of keyword matching when text keywords are dense, effectively avoiding retrieval bias caused by fixed weights. Furthermore, based on dynamically adjusted weight coefficients, a weighted fusion is performed on the corresponding text semantic units in the keyword candidate set and semantic candidate set, recalculating the comprehensive score of each text semantic unit. The adjusted fusion result significantly improves the semantic relevance between the recalled knowledge fragments and potential business targets, ensuring both the accuracy of keyword retrieval and enhancing the semantic retrieval's ability to capture implicit associations.

[0069] In practical applications, to ensure the accuracy of potential target discovery in hedging transactions, standard business data, query vectors, and recalled knowledge fragments can be organically combined by dynamically matching and adapting templates, filling in structured prompts, and constraining model reasoning. This ensures that the discovery results align with actual business needs and improves the accuracy of potential target discovery. Specifically, for example... Figure 5 As shown, step 104 above includes the following steps: 501. Based on the risk dimension characteristics and enterprise attribute parameters of the target to be mined, dynamically match and adapt templates from a preset library of multiple prompt word templates.

[0070] 502. Fill the key indicators in the standard business data, the business feature descriptions corresponding to the query vector, the recalled knowledge fragments, and the scoring rules in the dynamic rule base that match the current business scenario according to the structure of the adaptation template to generate structured prompt words.

[0071] 503. Input the structured prompt words into the preset large language model, constrain the preset large language model to explicitly reference the recalled knowledge fragments as the basis for reasoning during the reasoning process, and perform weighted calculation or conditional triggering judgment according to the scoring rules to output the mining results of potential business targets.

[0072] In this embodiment, the mining results of potential business targets can be calculated based on a calculation module using a preset large language model and a RAG retrieval module. This process employs a dual-drive model of retrieving professional knowledge and using prompt words to guide precise calculations.

[0073] Specifically, in the RAG retrieval process, when the large language model analyzes an aluminum alloy processing company, it first transforms the company's hedging needs into vectors. Then, it retrieves the most relevant knowledge fragments from the vector library, including aluminum futures delivery rules, raw material ratio scoring standards, and hedging cases of similar companies, and uses these as the factual basis for the calculation.

[0074] Specifically, in the prompt word engineering phase, standardized prompt word templates adapted to different business processes can be designed, clearly defining the computational logic and input / output format of the large language model. This standardized prompt word template comprises three parts: first, the business objective, such as calculating the urgency of hedging needs and recommending tools based on enterprise data and retrieved knowledge; second, the computational rules, such as calculating scores with weights of 30% for raw material proportion and 20% for price volatility; and third, the output requirements, such as including risk exposure type, priority score, and recommended tool combination. Finally, the retrieved knowledge fragments and preprocessed enterprise data are embedded together with the prompt words, guiding the large language model to output the results of mining potential business objectives according to fixed logic.

[0075] The entire implementation process is deeply integrated with the business workflow, achieving seamless collaboration across all stages and ensuring that the technology implementation aligns with business needs. During the screening phase, the prompts explicitly require first retrieving explicit requirements through keyword searches, then combining this with industry rules from the knowledge base to filter target companies. Simultaneously, the RAG search module retrieves business matching rules to quickly narrow down the range of target companies, improving screening efficiency.

[0076] During the demand analysis phase, prompts guide the large model to connect corporate financial data, market conditions, and retrieved industry transmission logic. For example, by combining the knowledge that capacity expansion leads to increased demand for raw materials, the implicit hedging gap of enterprises can be deduced. At the same time, prompts limit the use of cases from the knowledge base as strategy references, further enhancing the credibility and practicality of the recommendations.

[0077] During the results output phase, the prompt keywords are formatted as follows: basic enterprise information plus risk exposure details plus scoring criteria plus strategy recommendations. The knowledge sources retrieved by the RAG search module will be annotated simultaneously.

[0078] In terms of prompt word optimization, multiple templates can be designed according to risk type (cost side, inventory side) and enterprise size (large, medium and small). For example, prompt words for cross-border enterprises will specifically guide the use of knowledge base content on double hedging of exchange rates and commodities, which significantly improves the adaptability of scenarios and meets the hedging needs assessment of different types of enterprises.

[0079] At the same time, an iterative mechanism is set up to optimize the prompt word rules and knowledge base content based on feedback from business integration, forming a closed loop of retrieval, calculation, feedback, and optimization, continuously improving the accuracy of business opportunity assessment, and ensuring the accuracy and applicability of hedging business results calculation.

[0080] The following examples further illustrate how to apply the prompts. Template 1 is a prompt template for assessing hedging opportunities for companies with explicit needs. Its objective is to assess a company's explicit hedging needs based on publicly available information and knowledge base content, and output the priority of the opportunity, details of the risk exposure, and preliminary contact suggestions.

[0081] The input data for this template includes basic enterprise information, publicly available text information, key financial data, market data, and knowledge base support. Basic enterprise information includes the company name, industry, and main business. Publicly available text information includes hedging-related statements mentioned in announcements or interactive Q&A sessions. Key financial data includes the percentage of raw materials in operating costs, the percentage of inventory balance in total assets, and the percentage of cross-border business revenue. Market data includes the price volatility of the corresponding commodity over the past three months. Knowledge base support includes retrieved hedging cases from the same industry, commodity matching rules, and scoring criteria.

[0082] The calculation rules for Template 1 are as follows: if the text contains statements such as "intention to conduct hedging" or "researching hedging solutions", it is directly judged as a high-priority business opportunity; if only price or inventory risk is mentioned without mentioning hedging, the priority is determined by combining financial data; the urgency score is calculated based on the following weights: raw materials (30%), inventory (25%), price volatility (20%), cross-border revenue (15%), and explicit statement intensity (10%), with a maximum score of 10; referring to knowledge base examples, the optimal combination of hedging tools is matched, and the tool types are limited to futures, options, forward contracts, and option-embedded trades.

[0083] Template 1's output format requirements include five parts: enterprise business opportunity rating and score, core risk exposure, hedging needs analysis, recommended tool combination, and key suggestions for connection.

[0084] Template 2 is a template for identifying hedging opportunities for companies with implicit hedging needs. Its objective is to identify implicit hedging needs for companies that have not directly disclosed their hedging intentions but have potential risks through multi-dimensional data correlation analysis, and output demand gaps and strategic recommendations.

[0085] The input data for this template includes basic enterprise information, financial and operational data, market and industry data, and knowledge base support. The basic enterprise information includes the enterprise name, industry, and production capacity. The financial and operational data includes the proportion of raw materials in operating costs, the year-on-year growth rate of inventory, and the volatility of net profit in the past two quarters. The market and industry data includes the price trend of corresponding raw materials, the volatility in the past three months, and the industry hedging penetration rate. The knowledge base support includes the retrieved industry risk transmission logic, the rules for the correlation between capacity expansion and hedging demand, and cases for implicit demand assessment.

[0086] The calculation rules for Template 2 are as follows: there are three conditions for triggering implicit demand. If any one of these conditions is met, implicit demand can be determined to exist. These conditions are: 1) raw materials account for no less than 30% and the price volatility is no less than 20%; 2) the year-on-year growth rate of inventory is no less than 15% and the corresponding commodity price has fallen; and 3) new production capacity is no less than 10% and the external dependence on raw materials is no less than 50%. The potential risk loss amount is calculated by multiplying the raw material purchase volume by the price volatility. Based on the industry hedging strategies in the knowledge base, a hedging plan that is suitable for implicit demand is recommended. The risk hedging ratio and expected cost of the plan must be explained.

[0087] Template 2's output format requirements include five parts: implicit demand determination results, demand triggering conditions, potential risk loss calculation, customized hedging solutions, and demand awakening suggestions.

[0088] In practical applications, to ensure the accurate reach and visualization of the results of uncovering potential business objectives, a data visualization module is used as the terminal carrier connecting the results of large language model calculations with business execution, enabling a multi-dimensional presentation of the results of uncovering potential business objectives. Taking hedging business scenarios as an example, the data visualization module, based on the needs of hedging business, includes at least a comprehensive business opportunity overview dashboard and a business opportunity tracking and review dashboard.

[0089] Among them, the overall business opportunity overview dashboard is geared towards the management perspective, and its content includes industry business opportunity heat maps, pie charts showing the distribution of business opportunity types, real-time business opportunity dynamic monitoring, and connection information packages.

[0090] The industry opportunity heat map is categorized by sectors such as non-ferrous metals, energy and chemicals, and agricultural products. It uses color depth to represent the number of high-priority business opportunities and the average score of demand urgency in each industry. It also supports viewing business opportunity trends over time, making it easier for management to identify market demand hotspots.

[0091] The pie chart of business opportunity types distinguishes between explicit demand opportunities, implicit demand opportunities, and secondary development opportunities. Secondary development opportunities are those that have already been hedged but have expansion and optimization needs. At the same time, the average conversion rate of each type of business opportunity is marked with historical conversion data to provide data basis for management to allocate resources.

[0092] Real-time business opportunity dynamic monitoring displays the number of newly added high-priority business opportunities, the number of business opportunities that have been followed up, and the number of business opportunities that have been successfully implemented. High-priority business opportunities are those with a score of 8 or higher. At the same time, it automatically alerts high-priority business opportunities that have not been followed up for more than 7 days, effectively preventing the loss of business opportunities.

[0093] The integrated information package can automatically link to the contact information of key decision-makers in an enterprise, communication scripts, and compliance considerations. Key decision-makers include the chief financial officer and the chief risk officer, and compliance considerations include relevant content such as information disclosure requirements for hedging business.

[0094] Among them, the opportunity tracking and review dashboard is geared towards the operations level, focusing on the closed-loop management of the entire business process, supporting the analysis of opportunity conversion effects and process optimization. Its content includes opportunity conversion funnel charts, team and individual performance dashboards, and hedging effect review modules.

[0095] The business opportunity conversion funnel chart clearly shows the number and conversion rate of each stage: new business opportunities, initial contact, demand verification, solution design, and cooperation implementation. It can accurately pinpoint conversion bottlenecks. For example, if the conversion rate is low in the demand verification stage, the communication script needs to be optimized.

[0096] Among them, the team and individual performance dashboards display the number of follow-up business opportunities, the success rate of connection, and the amount of cooperation completed, according to the business team and individual dimensions. They also support filtering and comparison by track, providing reliable data support for corporate performance evaluation.

[0097] The hedging performance review module connects with the hedging data of partner companies, displaying indicators such as changes in profit volatility and the proportion of risk hedging costs before and after hedging, forming a comprehensive case library, which in turn feeds back into the big model to optimize the judgment logic and improve the accuracy of subsequent business opportunity discovery.

[0098] Furthermore, to achieve precise targeting of business opportunities and improve efficiency in connecting with them, the aforementioned data visualization module can also build an automated business opportunity push system based on the matching rules of people-jobs-tracks, ensuring that the right business opportunity reaches the right business personnel at the right time.

[0099] The automated business opportunity push system achieves precise matching through its push rules, specifically including three methods: matching by track, pushing by priority, and pushing by follow-up status.

[0100] Matching by track involves pre-assigning industry tracks to business personnel, such as non-ferrous metals industry specialists or cross-border hedging specialists. When the big model outputs business opportunities, the module automatically identifies the track to which the business opportunity belongs and pushes it to the corresponding business specialist.

[0101] Priority-based push notifications specifically mean that high-priority opportunities, i.e. those with a score of 8 or higher, are triggered for immediate push notifications across multiple channels, including WeChat Work and email. Medium and low-priority opportunities are pushed out daily in a way that balances push efficiency with the work pace of business personnel.

[0102] Specifically, the push notification based on the follow-up status means that when a company releases new announcements such as adjustments to hedging limits or changes in performance, the module automatically updates the business opportunity status and pushes an update reminder to the person following up on the business opportunity, along with an interpretation of the new information, to avoid business personnel conducting work based on old data.

[0103] Furthermore, to form a closed business loop, the aforementioned data visualization module also sets up a push feedback closed-loop management mechanism. After receiving a business opportunity push, business personnel need to mark the follow-up status in the system as viewed, reached, or not meeting the requirements. The module automatically records the operation time and results.

[0104] For business opportunities marked as reached, the module automatically adds them to the tracking and review dashboard, reminding business personnel to carry out follow-up actions according to preset nodes to ensure that business opportunity follow-up is uninterrupted.

[0105] For business opportunities marked as mismatched with demand, business personnel can enter the reasons for the mismatch. For example, if the company has signed a long-term price lock-in agreement, the system will synchronize the reason to the data storage module, feed back into the big model to optimize the judgment rules, reduce the probability of misjudging similar business opportunities, and continuously improve the accuracy of business opportunity discovery.

[0106] Furthermore, as a specific implementation of the above method, embodiments of this application provide a business potential target mining system based on a large model, such as... Figure 6 As shown in the diagram, the deployment structure and data flow path of each functional module in this embodiment of the invention are clearly illustrated. The entire architecture uses Spark as its core computing engine, enabling data scheduling and parallel processing across storage modules. A load balancing strategy is used to partition the dataset, improving processing efficiency. The bidirectional connection between the Milvus database and the large model computing module reflects the complete RAG mechanism process: the retrieval phase retrieves relevant knowledge fragments from Milvus, and the computing phase injects the knowledge into the context of the large model, forming enhanced inference results. This system design achieves full automation and intelligence from raw data access to intelligent analysis and result display, possessing good scalability and stability.

[0107] Figure 6 The left-hand area represents the data input and preprocessing stage. The raw data is stored in HDFS and cleaned, denoised, and formatted by the data preprocessing module. At the same time, the Milvus database serves as the retrieval source for data retrieval. After processing, the data is read by the Spark engine and distributed to the computing process.

[0108] Figure 6 The central area is divided into two main parts. On the left are storage module 1 and storage module 2, corresponding to the HDFS cluster and Milvus vector database cluster, respectively. Storage module 1 is used to store the original text dataset and intermediate processing results, supporting distributed storage of large-scale unstructured data. Storage module 2 is dedicated to the persistence of knowledge base data. It contains multiple data units, representing the semantic units of the text after semantic vectorization and their associated meta-information, forming the core knowledge carrier of the RAG mechanism.

[0109] Figure 6 The right-hand area is the large model computation module, containing storage module 3, which is the model result storage area, used to temporarily store the business opportunity analysis results output by the large model inference. This area establishes data interaction channels with the left-hand storage module through multiple connecting lines, indicating that in the RAG process, the large model extracts raw text and semantic vectors from HDFS and Milvus, and writes the final computation results into storage module 3, realizing the closed-loop execution of the computation task.

[0110] Figure 6The outermost layer on the right is the data visualization module, including the ACTION component. It connects to the Milvus database via indexing, enabling rapid retrieval of knowledge base content. Simultaneously, it connects to the TDSQL database through a data storage interface to save and manage structured analysis results. The front-end display system generates a visualization interface based on data from TDSQL, presenting and pushing the results of uncovering potential business objectives, ensuring that the analysis results can be used efficiently.

[0111] Furthermore, as a specific implementation of the above method, embodiments of this application provide a business potential target mining device based on a large model, such as... Figure 7 As shown, the device includes: an acquisition unit 61, a storage unit 62, a generation unit 63, and an inference unit 64.

[0112] The acquisition unit 61 is used to acquire standard business data obtained by multi-stage preprocessing of multi-source heterogeneous raw business data; Storage unit 62 is used to map the standard business data and the text data in the business knowledge base into semantic vectors through a text embedding model based on a pre-built business knowledge base, and store them in a vector database; The generation unit 63 is used to respond to a business potential target mining request, generate a query vector based on the business characteristics of the target to be mined, and use a hybrid retrieval strategy to recall relevant knowledge fragments in the vector database. The hybrid retrieval strategy includes keyword matching and semantic similarity retrieval. The reasoning unit 64 is used to input the standard business data, the query vector, and the recalled knowledge fragments into a preset large language model, and guide the large language model to perform logical reasoning based on the scoring rules in the dynamic rule base through a preset prompt word template, and output the mining results of potential business targets; the mining results include explicit demand determination results or implicit demand determination results.

[0113] The business potential target mining device based on a large model provided in this invention, compared with the existing method of using manual mining to achieve business potential target mining, obtains standard business data by preprocessing multi-source heterogeneous original business data through multiple stages; based on a pre-built business knowledge base, the standard business data and text data in the business knowledge base are mapped into semantic vectors through a text embedding model and stored in a vector database; in response to a business potential target mining request, a query vector is generated based on the business features of the target to be mined, and relevant knowledge fragments are recalled in the vector database using a hybrid retrieval strategy, which includes keyword matching and semantic similarity retrieval; the standard business data, query vectors, and recalled knowledge fragments are input into a preset large language model, and the large language model is guided by a preset prompt word template to perform logical reasoning based on the scoring rules in the dynamic rule base, and output the mining results of the business potential target; the mining results include explicit demand determination results or implicit demand determination results. The entire process transforms multi-source heterogeneous raw business data into high-quality standard business data through multi-stage preprocessing, effectively eliminating format heterogeneity and business noise. Then, relying on a text embedding model, the standard business data and business knowledge base text are uniformly transformed into semantic vectors and stored in the database, realizing the semantic fusion and efficient aggregation of heterogeneous business information. At the same time, a hybrid retrieval strategy combining keyword matching and semantic similarity retrieval is used to carry out knowledge retrieval. This retains the retrieval reliability of accurate keyword matching and can effectively reach deep-related knowledge through semantic mining, significantly improving the comprehensiveness and accuracy of the retrieval of target related knowledge fragments. Subsequently, the pre-set prompt word template guides the large language model to complete intelligent logical reasoning in combination with the scoring rules of the dynamic rule base. This avoids the defects of large subjective bias, low efficiency and inconsistent standards of manual analysis, and can quickly and objectively output two types of refined mining results: explicit demand judgment and implicit demand judgment. This comprehensively enhances the recognition accuracy and response processing efficiency of potential targets in complex business scenarios.

[0114] In specific application scenarios, the acquisition unit is specifically used for: The pre-built heterogeneous data adaptation gateway is used to obtain multi-source heterogeneous raw business data from distributed nodes, and the format preprocessing of the multi-source heterogeneous raw business data is performed to obtain preprocessed business data. Hybrid noise filtering is performed on the preprocessed business data to remove structural markers, non-business symbols, and invalid formatted text in preset business scenarios. Then, fingerprint comparison technology based on hash features combined with a semantic relevance discrimination model is used to deduplicatize and merge the data and remove low-quality data to obtain purified business data. The purified business data undergoes character encoding standardization and typesetting symbol standardization. Based on the industry professional terminology knowledge graph, string matching and normalization algorithms are used to map non-standard business expressions and punctuation marks into standard formats to obtain standard business data.

[0115] In specific application scenarios, the original business data includes text-based unstructured data, semi-structured log data, and structured numerical data. The acquisition unit is further specifically used for: For the aforementioned unstructured text data, optical character recognition and document parsing technologies are used to convert portable document formats and text processing document formats into editable character streams; For the semi-structured log data and structured numerical data, character mapping and key-value pair extraction techniques are used to transform them into unified data objects to eliminate the format heterogeneity of the data source. The device further includes: The push unit is used to label the standard business data obtained by multi-stage preprocessing of the original heterogeneous business data from multiple sources, and to label the standard business data with data source, preprocessing time and processing node metadata. It also pushes the preprocessed structured numerical data to the time series database through a distributed message queue, and pushes the preprocessed text-type unstructured data to the preset large language model.

[0116] In specific application scenarios, the storage unit is specifically used for: Semantic segmentation processing is performed on the text data in the standard business data and business knowledge base. Based on the preset character length threshold and the syntactic dependency analysis results, the continuous text is segmented into text semantic units that maintain semantic integrity. A pre-trained text embedding model is invoked to vectorize and encode each text semantic unit, generating a business semantic vector corresponding to the text semantic unit; the text embedding model is based on a self-attention architecture and extracts semantic association features between words through a context-aware mechanism; The business semantic vector is structurally bound to its corresponding text semantic unit, data source identifier, and timestamp metadata, and persistently stored in the vector database. The structural binding is used to locate and return the text semantic unit content based on the recalled semantic vector in the subsequent retrieval stage. The text semantic unit content serves as the context for reasoning in the preset large language model. Based on the dimensional parameters and data scale of the business semantic vector, a hierarchical navigable small world graph index or a product-quantized inverted index is constructed to efficiently organize the vector space and accelerate retrieval optimization, so as to support the real-time response requirements of semantic similarity calculation in the hybrid retrieval strategy.

[0117] In specific application scenarios, the generation unit is specifically used for: In response to a request to mine potential business targets, a query text is constructed based on the structured business features and unstructured business description text of the target to be mined. A pre-trained text embedding model is called to vectorize the query text and generate a query vector aligned with the business semantic space. A hybrid retrieval strategy is adopted, simultaneously performing keyword matching retrieval based on inverted index and semantic retrieval based on vector similarity, generating keyword candidate sets containing text semantic unit identifiers and matching scores, and semantic candidate sets containing text semantic unit identifiers and similarity scores, respectively. Based on the preset keyword matching weight coefficient and semantic similarity weight coefficient, the corresponding text semantic units in the keyword candidate set and the semantic candidate set are weighted and fused to calculate the comprehensive score of each text semantic unit. A predetermined number of text semantic units are selected as the recalled knowledge fragments in descending order of their comprehensive scores. The recalled knowledge fragments include the content of the text semantic units and their associated metadata.

[0118] In specific application scenarios, the generation unit is further used for: Based on the industry volatility index of the target to be mined and the keyword density index of the business description text, the keyword matching weight coefficient and the semantic similarity weight coefficient are dynamically adjusted; wherein, the industry volatility index is obtained by querying the business knowledge base, and the keyword density index is calculated by statistically analyzing the frequency of occurrence of preset business keywords in the business description text; when the industry volatility index is greater than a preset volatility threshold or the keyword density index is less than a preset density threshold, the semantic similarity weight coefficient is increased; when the keyword density index is greater than the preset density threshold, the keyword matching weight coefficient is increased. The weighted fusion is performed based on dynamically adjusted weight coefficients to optimize the semantic relevance between recalled knowledge fragments and potential business targets.

[0119] In specific application scenarios, the inference unit is specifically used for: Based on the risk dimension characteristics and enterprise attribute parameters of the target to be mined, a matching template is dynamically matched from a pre-set library of multiple prompt word templates; The key indicators in the standard business data, the business feature descriptions corresponding to the query vector, the recalled knowledge fragments, and the scoring rules in the dynamic rule base that match the current business scenario are filled in according to the structure of the adaptation template to generate structured prompt words; The structured prompts are input into the preset large language model, which is constrained to explicitly reference the recalled knowledge fragments as the basis for reasoning during the reasoning process. The model then performs weighted calculations or conditional triggering judgments according to the scoring rules, and outputs the mining results of potential business targets.

[0120] Based on the above-described large-model-based business potential target mining method, this application embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the above-described large-model-based business potential target mining method.

[0121] Based on this understanding, the technical solution of this application can be embodied in the form of a software product. The software product can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, or portable hard drive), and includes several instructions to cause a computer device (such as a personal computer, server, or network device) to execute the methods described in the various implementation scenarios of this application.

[0122] Based on the above-described large-model-based business potential target mining method and corresponding virtual device embodiments, in order to achieve the above objectives, this application embodiment also provides a physical device for large-model-based business potential target mining, which may be a computer, smartphone, tablet computer, smartwatch, server, or network device, etc. The physical device includes a storage medium and a processor; the storage medium is used to store computer programs; the processor is used to execute the computer programs to implement the above-described large-model-based business potential target mining method.

[0123] Optionally, the physical device may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB interfaces, card reader interfaces, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Wi-Fi interfaces), etc.

[0124] In an exemplary embodiment, see Figure 8 The aforementioned physical devices include a communication bus, a processor, a memory, and a communication interface. They may also include input / output interfaces and a display device. The various functional units can communicate with each other via the bus. The memory stores computer programs, and the processor executes the programs stored in the memory to perform the business potential target mining method based on a large model described in the above embodiments.

[0125] Those skilled in the art will understand that the entity device structure for mining potential business targets based on a large model provided in this embodiment does not constitute a limitation on the entity device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0126] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the physical devices involved in the large-scale model-based business potential target mining, supporting the operation of information processing programs and other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software in the information processing physical devices.

[0127] Through the above description of the implementation methods, those skilled in the art can clearly understand that this application can be implemented using software plus necessary general-purpose hardware platforms, or it can be implemented using hardware. By applying the technical solution of this application, compared with the existing methods, this application transforms multi-source heterogeneous original business data into high-quality standard business data through multi-stage preprocessing, effectively eliminating format heterogeneity and business noise. Then, relying on the text embedding model, the standard business data and business knowledge base text are uniformly transformed into semantic vectors and stored in the database, realizing the semantic fusion and efficient collection of heterogeneous business information. At the same time, a hybrid retrieval strategy combining keyword matching and semantic similarity retrieval is used to carry out knowledge retrieval, which not only retains the retrieval reliability of accurate keyword matching, but also achieves effective access to deep related knowledge through semantic mining, significantly improving the comprehensiveness and accuracy of the retrieval of target related knowledge fragments. Subsequently, by using preset prompt word templates to guide the large language model to complete intelligent logical reasoning in combination with the scoring rules of the dynamic rule base, the defects of large subjective bias, low efficiency and inconsistent standards of manual analysis can be avoided. It can quickly and objectively output two types of refined mining results: explicit demand judgment and implicit demand judgment, comprehensively enhancing the recognition accuracy and response processing efficiency of potential targets in complex business scenarios.

[0128] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application. Those skilled in the art will understand that the modules in the apparatus of the embodiment can be distributed within the apparatus of the embodiment as described, or can be modified to be located in one or more apparatuses different from this embodiment. The modules of the above-described embodiment can be combined into one module, or further divided into multiple sub-modules.

[0129] The serial numbers in this application are for descriptive purposes only and do not represent the superiority or inferiority of any particular implementation scenario. The above disclosures are merely a few specific implementation scenarios of this application; however, this application is not limited thereto, and any variations conceived by those skilled in the art should fall within the protection scope of this application.

Claims

1. A method for mining potential business targets based on a large model, characterized in that, include: Standard business data is obtained by acquiring heterogeneous raw business data from multiple sources and preprocessing it through multiple stages. Based on a pre-built business knowledge base, the standard business data and the text data in the business knowledge base are mapped into semantic vectors through a text embedding model and stored in a vector database. In response to a request to mine potential business targets, a query vector is generated based on the business characteristics of the target to be mined. A hybrid retrieval strategy is used to recall relevant knowledge fragments from the vector database. This hybrid retrieval strategy includes keyword matching and semantic similarity retrieval. The keyword matching weight coefficient and semantic similarity weight coefficient are dynamically adjusted based on the industry volatility index of the target to be mined and the keyword density index of the business description text. The industry volatility index is obtained by querying the business knowledge base, and the keyword density index is calculated by statistically analyzing the frequency of occurrence of preset business keywords in the business description text. When the industry volatility index is greater than a preset volatility threshold or the keyword density index is less than a preset density threshold, the semantic similarity weight coefficient is increased. When the keyword density index is greater than the preset density threshold, the keyword matching weight coefficient is increased. Weighted fusion is performed based on the dynamically adjusted weight coefficients to optimize the semantic relevance between the recalled knowledge fragments and the potential business targets. Based on the risk dimension characteristics and enterprise attribute parameters of the target to be mined, a matching template is dynamically matched from a pre-set library of multiple prompt word templates. Key indicators from the standard business data, business feature descriptions corresponding to the query vector, recalled knowledge fragments, and scoring rules matching the current business scenario from the dynamic rule base are filled in according to the structure of the matching template to generate structured prompt words. These structured prompt words are input into a pre-set large language model, which is constrained to explicitly reference the recalled knowledge fragments as reasoning basis during the reasoning process. Weighted calculations or conditional triggering judgments are performed according to the scoring rules to output the mining results of potential business targets. The mining results include explicit demand judgment results or implicit demand judgment results.

2. The method according to claim 1, characterized in that, The standard business data obtained by acquiring multi-source heterogeneous raw business data through multi-stage preprocessing includes: The pre-built heterogeneous data adaptation gateway is used to obtain multi-source heterogeneous raw business data from distributed nodes, and the format preprocessing of the multi-source heterogeneous raw business data is performed to obtain preprocessed business data. Hybrid noise filtering is performed on the preprocessed business data to remove structural markers, non-business symbols, and invalid formatted text in preset business scenarios. Then, fingerprint comparison technology based on hash features combined with a semantic relevance discrimination model is used to deduplicatize and merge the data and remove low-quality data to obtain purified business data. The purified business data undergoes character encoding standardization and typesetting symbol standardization. Based on the industry professional terminology knowledge graph, string matching and normalization algorithms are used to map non-standard business expressions and punctuation marks into standard formats to obtain standard business data.

3. The method according to claim 2, characterized in that, The original business data includes text-based unstructured data, semi-structured log data, and structured numerical data. The preprocessing of the multi-source heterogeneous original business data to obtain preprocessed business data includes: For the aforementioned unstructured text data, optical character recognition and document parsing technologies are used to convert portable document formats and text processing document formats into editable character streams; For the semi-structured log data and structured numerical data, character mapping and key-value pair extraction techniques are used to transform them into unified data objects to eliminate the format heterogeneity of the data source. After obtaining standard business data from the multi-source heterogeneous raw business data through multi-stage preprocessing, the method further includes: The standard business data is labeled with data source, preprocessing time and processing node metadata. The preprocessed structured numerical data is pushed to the time series database through a distributed message queue, and the preprocessed text-type unstructured data is pushed to the preset large language model.

4. The method according to claim 1, characterized in that, The pre-built business knowledge base maps the standard business data and text data in the business knowledge base into semantic vectors through a text embedding model and stores them in a vector database, including: Semantic segmentation processing is performed on the text data in the standard business data and business knowledge base. Based on the preset character length threshold and the syntactic dependency analysis results, the continuous text is segmented into text semantic units that maintain semantic integrity. A pre-trained text embedding model is invoked to vectorize and encode each text semantic unit, generating a business semantic vector corresponding to the text semantic unit; the text embedding model is based on a self-attention architecture and extracts semantic association features between words through a context-aware mechanism; The business semantic vector is structurally bound to its corresponding text semantic unit, data source identifier, and timestamp metadata, and persistently stored in the vector database. The structural binding is used to locate and return the text semantic unit content based on the recalled semantic vector in the subsequent retrieval stage. The text semantic unit content serves as the context for reasoning in the preset large language model. Based on the dimensional parameters and data scale of the business semantic vector, a hierarchical navigable small world graph index or a product-quantized inverted index is constructed to efficiently organize the vector space and accelerate retrieval optimization, so as to support the real-time response requirements of semantic similarity calculation in the hybrid retrieval strategy.

5. The method according to any one of claims 1-4, characterized in that, In response to a request to mine potential business targets, a query vector is generated based on the business characteristics of the target to be mined. A hybrid retrieval strategy is then used to recall relevant knowledge fragments from the vector database, including: In response to a request to mine potential business targets, a query text is constructed based on the structured business features and unstructured business description text of the target to be mined. A pre-trained text embedding model is called to vectorize the query text and generate a query vector aligned with the business semantic space. A hybrid retrieval strategy is adopted, simultaneously performing keyword matching retrieval based on inverted index and semantic retrieval based on vector similarity, generating keyword candidate sets containing text semantic unit identifiers and matching scores, and semantic candidate sets containing text semantic unit identifiers and similarity scores, respectively. Based on the preset keyword matching weight coefficient and semantic similarity weight coefficient, the corresponding text semantic units in the keyword candidate set and the semantic candidate set are weighted and fused to calculate the comprehensive score of each text semantic unit. A predetermined number of text semantic units are selected as the recalled knowledge fragments in descending order of their comprehensive scores. The recalled knowledge fragments include the content of the text semantic units and their associated metadata.

6. A business potential target mining device based on a large model, characterized in that, include: The acquisition unit is used to acquire standard business data obtained by multi-stage preprocessing of multi-source heterogeneous raw business data; A storage unit is used to map the standard business data and the text data in the business knowledge base into semantic vectors through a text embedding model based on a pre-built business knowledge base, and store them in a vector database. A generation unit, in response to a request to mine potential business targets, generates a query vector based on the business characteristics of the target to be mined, and uses a hybrid retrieval strategy to recall relevant knowledge fragments from the vector database. The hybrid retrieval strategy includes keyword matching and semantic similarity retrieval. Based on the industry volatility index of the target to be mined and the keyword density index of the business description text, the unit dynamically adjusts the keyword matching weight coefficient and the semantic similarity weight coefficient. The industry volatility index is obtained by querying the business knowledge base, and the keyword density index is calculated by statistically analyzing the frequency of occurrence of preset business keywords in the business description text. When the industry volatility index is greater than a preset volatility threshold or the keyword density index is less than a preset density threshold, the semantic similarity weight coefficient is increased; when the keyword density index is greater than the preset density threshold, the keyword matching weight coefficient is increased. Weighted fusion is performed based on the dynamically adjusted weight coefficients to optimize the semantic relevance between the recalled knowledge fragments and the potential business targets. The inference unit is used to dynamically match and adapt templates from a pre-set library of multiple prompt word templates based on the risk dimension characteristics and enterprise attribute parameters of the target to be mined; to fill in the key indicators in the standard business data, the business feature descriptions corresponding to the query vector, the recalled knowledge fragments, and the scoring rules in the dynamic rule base that match the current business scenario according to the structure of the adapting templates to generate structured prompt words; to input the structured prompt words into a pre-set large language model, constraining the pre-set large language model to explicitly reference the recalled knowledge fragments as the basis for inference during the inference process, and to perform weighted calculations or conditional triggering judgments according to the scoring rules, and output the mining results of potential business targets; the mining results include explicit demand judgment results or implicit demand judgment results.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

8. A computer storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images