A multi-source heterogeneous data full-link closed-loop processing method and system

Abstract

The application discloses a kind of multi-source heterogeneous data full-link closed loop processing method and system, comprising: analyzing multi-source heterogeneous data, obtains semantic feature and access boundary feature;Semantic feature and access boundary feature are respectively converted into vector data and metadata for storage;Collection historical check iteration number, determine search damping coefficient, and the elastic coupling calculation of vector data and metadata is carried out, obtains candidate tender collection;Based on candidate tender collection, generate preliminary business tender, and carry out tender compliance check to preliminary business tender, extract feature misplacement information;According to feature misplacement information, reverse update search damping coefficient, and based on the search damping coefficient after updating, obtain output compliance business tender.The application compared with prior art, it can be through feature misplacement information reverse update search damping coefficient, avoid irregular history sample being recalled repeatedly, reduce the feature misplacement rate of core business index, improve the compliance accuracy of business tender generation.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and data processing technology, and more specifically, to a method and system for full-link closed-loop processing of multi-source heterogeneous data. Background Technology

[0002] In the intelligent bidding and business management scenarios of modern enterprise-level engineering projects, the system needs to process massive, multi-source, and heterogeneous business data in real time. This data includes structured enterprise resource ledgers, financial and social security tax data, as well as unstructured bidding documents, contract documents, audio and video interaction records, and other formats. To support business collaboration and intelligent decision-making under multi-terminal access, modern enterprises have built a digital business platform based on large language models and multimodal perception technology. Through the parsing, association, and reorganization of the above-mentioned multi-source heterogeneous data, the platform supports processes such as project bidding, resource matching, and document preparation.

[0003] Existing technologies for automating the generation of business tender documents typically convert historical tender documents into vector data. Upon receiving a generation request, they retrieve matching historical texts by calculating a single semantic similarity and then directly feed the data into a large language model for concatenation and output. However, simple semantic retrieval can only match literal meanings and cannot identify the magnitude limits of numerical values. This makes it easy for the system to recall historical data that violates the rules, and the large language model will repeatedly extract these erroneous samples, resulting in misalignments of data such as amounts and times.

[0004] Therefore, how to provide a closed-loop processing method and system for multi-source heterogeneous data across the entire chain, which can update the retrieval damping coefficient in reverse through feature misalignment information, avoid repeated recall of illegal historical samples, reduce the feature misalignment rate of core business indicators, and improve the compliance accuracy of business tender documents, has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] To address the aforementioned technical issues, this invention provides a closed-loop processing method for multi-source heterogeneous data across the entire chain. This method can update the retrieval damping coefficient in reverse through feature misalignment information, preventing the repeated recall of illegal historical samples, reducing the feature misalignment rate of core business indicators, and improving the compliance accuracy of business tender document generation.

[0006] The first technical solution provided by this invention is as follows: This invention provides a closed-loop processing method for multi-source heterogeneous data across the entire chain, comprising the following steps: S1 receiving multi-source heterogeneous data, parsing the multi-source heterogeneous data, and obtaining semantic features and access boundary features; S2 converting the semantic features into vector data for storage, and converting the access boundary features into metadata for storage; S3 collecting historical verification iteration counts and determining a retrieval damping coefficient, wherein the retrieval damping coefficient corresponds one-to-one with the historical verification iteration counts; S4 responding to a tender document generation request, performing elastic coupling calculation on the vector data and the metadata based on the retrieval damping coefficient to obtain a candidate tender document set, wherein the retrieval damping coefficient is used to quantify the degree of attenuation of retrieval similarity by the access boundary features; S5 generating preliminary business tender documents based on the candidate tender document set, and performing tender document compliance verification on the preliminary business tender documents; if the tender document compliance verification fails, extracting feature misalignment information between the preliminary business tender documents and the candidate tender document set; S6 updating the retrieval damping coefficient in reverse according to the feature misalignment information, and re-obtaining the candidate tender document set based on the updated retrieval damping coefficient, and outputting a compliant business tender document.

[0007] Furthermore, in a preferred embodiment of the present invention, parsing the multi-source heterogeneous data to obtain semantic features and admission boundary features includes: The multi-source heterogeneous data is format-normalized to obtain the text stream to be processed; Retrieve the delimiter of the text stream to be processed, and divide the text stream to be processed based on the delimiter to obtain the basic text sequence; Based on a preset window length, the basic text sequence is overlapped and truncated to generate overlapping text blocks. There is a preset number of overlapping characters between adjacent overlapping text blocks. Perform high-dimensional space mapping on the overlapping text blocks to obtain the semantic features; The overlapping text blocks are subjected to entity recognition and context parsing to determine candidate entity phrases. The candidate entity phrases are then mapped and matched with the business dictionary to obtain the core elements for admission. The core access elements are converted using timestamps and numerical standards to generate access boundary features.

[0008] Furthermore, in a preferred embodiment of the present invention, the preliminary business tender document undergoes a tender compliance verification, including: Parse the preliminary business tender document to obtain the feature fields to be verified; Based on the tender document generation request, collect project category tags, traverse the preset rule relationship table according to the project category tags, and locate the target compliance rule row; Extract the compliance boundary indicators of the target compliance rule row and construct the compliance value range; The feature field to be verified is compared with the compliant numerical range by boundary comparison to obtain a multi-dimensional boundary overflow vector, and it is determined whether there is an overflow component greater than zero in the multi-dimensional boundary overflow vector. If it exists, it is determined that the preliminary business tender document has failed the tender document compliance verification, the overflow component with a value greater than zero is obtained, and feature misalignment information is generated.

[0009] Furthermore, in a preferred embodiment of the present invention, obtaining the multidimensional boundary overflow vector includes: Semantic text parsing is performed on the feature field to be verified to extract the non-numerical text of the feature field to be verified, and the non-numerical text is removed. The removed feature field to be verified is then converted into a standard numerical scalar. Extract the compliance boundary indicators of the compliance numerical range, and align the standard numerical scalar with the compliance boundary indicators in the same dimension; If the current compliance boundary indicator is a lower limit control, then the standard numerical scalar value is subtracted from the compliance boundary indicator to generate a positive overflow value; If the current compliance boundary indicator is an upper limit control, then the standard numerical scalar is subtracted from the compliance boundary indicator to generate a positive overflow value; The positive overflow values ​​of each dimension are reorganized into arrays to construct a multidimensional boundary overflow vector.

[0010] Furthermore, in a preferred embodiment of the present invention, the vector data and the metadata are elastically coupled to obtain a candidate bid set, including: Calculate the vector distance between the tender document generation request and the vector data to obtain semantic similarity; Extract the target boundary features from the tender document generation request, and calculate the absolute numerical difference between the target boundary features and the corresponding features in the metadata to obtain the boundary offset; The similarity decay value is generated by multiplying the retrieval damping coefficient with the boundary offset and performing an exponential operation on the product. The overall matching degree is obtained by subtracting the similarity decay value from the semantic similarity. The comprehensive matching degree of all the vector data is sorted in descending order to obtain a comprehensive matching sequence, and a candidate bid set is constructed based on the comprehensive matching sequence.

[0011] Furthermore, in a preferred embodiment of the present invention, outputting a compliant business tender document includes: Analyze the feature misalignment information and extract the multidimensional boundary overflow vector; Set the reverse iteration step size, and perform a scalar multiplication operation between the multidimensional boundary overflow vector and the reverse iteration step size to generate a damping increment vector. The current search damping coefficient is summed with the damping increment vector to obtain the search damping optimization coefficient. In response to the aforementioned retrieval damping optimization coefficient, the elastic coupling calculation is re-executed to obtain a revised candidate tender set; The modified candidate bid set is input into a large language model to generate a second bid text, and a compliant business bid is output.

[0012] Further, in a preferred embodiment of the present invention, the multidimensional boundary overflow vector is multiplied by the reverse iteration step size to generate a damping increment vector, including: Extract the overflow components of each dimension in the multidimensional boundary overflow vector, and based on the current dimension, retrieve historical compliance sample data that has passed compliance verification from the metadata. Statistical features are extracted from the historical compliance sample data to construct a historical compliance sample distribution matrix, which includes the mean and variance of the historical compliance sample data. Calculate the statistical deviation of the overflow component relative to the historical compliant sample distribution matrix to obtain the dynamic deviation scalar; The dynamic deviation scalar input is normalized and mapped using a preset nonlinear activation function to generate dynamic severity level weights. The reverse iteration step size is then multiplied by the dynamic severity level weights to generate a dynamic iteration step size vector. The multidimensional boundary overflow vector and the dynamic iteration step size vector are subjected to a Hadamard product operation to generate a damping increment vector.

[0013] Furthermore, in a preferred embodiment of the present invention, a convergence blocking verification step is included before re-performing the elastic coupling calculation, specifically as follows: Count the number of verification iterations for the current tender document generation request, and determine whether the number of verification iterations is greater than a preset convergence blocking threshold; If it is not greater than, then in response to the retrieval damping optimization coefficient, the re-execution of the elastic coupling calculation is triggered; If the value is greater than the limit, the verification iteration is interrupted, and the candidate bid set corresponding to the maximum comprehensive matching degree in each round of verification iteration is extracted. An alarm business bid is then output. The alarm business bid indicates that the candidate bid set still cannot automatically converge to the bid compliance verification standard after extreme iterations, and there is feature misalignment information that requires manual review and intervention.

[0014] The present invention provides a second technical solution as follows: This invention also provides a multi-source heterogeneous data end-to-end closed-loop processing system, the system comprising: The heterogeneous data parsing module is used to receive multi-source heterogeneous data, parse the multi-source heterogeneous data, and obtain semantic features and access boundary features. The feature transformation and storage module is used to transform the semantic features into vector data for storage, and to transform the admission boundary features into metadata for storage. The initial damping determination module collects the number of historical verification iterations and determines the retrieval damping coefficient, wherein the retrieval damping coefficient corresponds one-to-one with the number of historical verification iterations; The elastic coupling calculation module, in response to the tender document generation request, performs elastic coupling calculation on the vector data and the metadata based on the retrieval damping coefficient to obtain a set of candidate tender documents. The retrieval damping coefficient is used to quantify the degree of attenuation of the retrieval similarity by the admission boundary features. The compliance misalignment extraction module generates preliminary business bids based on the candidate bid set and performs bid compliance verification on the preliminary business bids. If the bid compliance verification fails, it extracts the feature misalignment information between the preliminary business bids and the candidate bid set. The damping reverse update module updates the retrieval damping coefficient in reverse according to the feature misalignment information, and re-obtains the candidate bid set based on the updated retrieval damping coefficient, and outputs compliant business bids.

[0015] Furthermore, in a preferred embodiment of the present invention, the elastic coupling calculation module includes: A semantic similarity calculation unit is used to calculate the vector distance between the tender document generation request and the vector data to obtain semantic similarity. The boundary offset calculation unit is used to extract the target boundary features from the tender document generation request, and calculate the absolute numerical difference between the target boundary features and the corresponding features in the metadata to obtain the boundary offset. A similarity attenuation generation unit is used to perform a product exponential operation on the retrieval damping coefficient and the boundary offset to generate a similarity attenuation value. A comprehensive matching calculation unit is used to obtain a comprehensive matching degree by subtracting the similarity decay value from the semantic similarity; The candidate set construction unit is used to sort the comprehensive matching degree of all the vector data in descending order, obtain the comprehensive matching sequence, and construct a candidate bid set based on the comprehensive matching sequence.

[0016] The multi-source heterogeneous data end-to-end closed-loop processing method and system of the present invention have the following advantages over the prior art: (1) Based on the feature misalignment information when the tender compliance verification fails, the retrieval damping coefficient of the front-end retrieval process is updated in reverse, so that the system can dynamically remove historical samples that cause the indicator violation, avoid invalid iteration caused by compliance verification error, and reduce the risk of cyclical violation of key numerical indicators such as amount and construction period. (2) By using the admission boundary feature as the decay constraint term for similarity calculation, the limitations of traditional retrieval that only focuses on literal semantics are changed. This allows the retrieval logic to impose a mandatory constraint on the numerical limits of historical data, reducing the misalignment rate of key features in the splicing of complex business terms and ensuring the compliance of parameters in the final output business tender. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating the steps of the multi-source heterogeneous data end-to-end closed-loop processing method provided in this embodiment of the invention; Figure 2 A flowchart illustrating the steps for defining semantic features and admission boundary features in embodiments of the present invention; Figure 3 A logical framework diagram for generating feature misalignment information provided in embodiments of the present invention; Figure 4 A logical framework diagram of a multidimensional boundary overflow vector provided in an embodiment of the present invention; Figure 5 The simulation comparison results of high-value data compliance in industrial automation tender documents provided in this embodiment of the invention are shown in the figure. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0020] It should be noted that when a component is referred to as being "fixed to" or "set on" another component, it can be directly on or indirectly set on the other component; when a component is referred to as being "connected to" another component, it can be directly connected to or indirectly connected to the other component.

[0021] It should be understood that the terms "length", "width", "upper", "lower", "front", "rear", "first", "second", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention.

[0022] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" or "several" means two or more, unless otherwise explicitly specified.

[0023] It should be noted that the structures, proportions, sizes, etc., shown in the accompanying drawings of this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed in the specification, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0024] like Figures 1 to 5As shown, this invention provides a closed-loop processing method for multi-source heterogeneous data across the entire chain, specifically including: S1 receiving multi-source heterogeneous data, parsing the multi-source heterogeneous data, and obtaining semantic features and access boundary features; S2 converting the semantic features into vector data for storage, and converting the access boundary features into metadata for storage; S3 collecting historical verification iteration counts, determining the retrieval damping coefficient, with the retrieval damping coefficient corresponding one-to-one with the historical verification iteration counts; S4 responding to the tender document generation request, performing elastic coupling calculation on the vector data and metadata based on the retrieval damping coefficient, obtaining a candidate tender document set, the retrieval damping coefficient being used to quantify the degree of attenuation of retrieval similarity by the access boundary features; S5 generating preliminary business tender documents based on the candidate tender document set, and performing tender document compliance verification on the preliminary business tender documents. If the tender document compliance verification fails, feature misalignment information between the preliminary business tender document and the candidate tender document set is extracted; S6 updating the retrieval damping coefficient in reverse according to the feature misalignment information, and re-obtaining the candidate tender document set based on the updated retrieval damping coefficient, outputting a compliant business tender document.

[0025] In a specific embodiment of the present invention, the system receives multi-source heterogeneous data generated in the bidding and tendering business, such as historical tender documents, project requirements, compliance standard libraries, etc. By parsing these data, the system extracts semantic features representing the core logic of the text, as well as access boundary features representing business red lines (such as budget amount, construction period, qualification level, etc.).

[0026] The semantic features mentioned above are vectorized and encoded into high-dimensional vector data and stored in a vector database. At the same time, the ingress boundary features are structured and converted into metadata that can be numerically compared. An index association between the semantic space and the numerical boundary space is established to facilitate subsequent coupled retrieval.

[0027] The system records and collects the historical verification iteration count of the current tender generation request. This historical verification iteration count is used as an input variable for the retrieval primary key or adaptive factor to determine the retrieval damping coefficient effective in the current round. In this embodiment, the system pre-configures a damping mapping table in memory. The retrieval damping coefficient and the historical verification iteration count have a one-to-one correspondence in the damping mapping table, either as a function mapping or an entry matching relationship. The value of the retrieval damping coefficient dynamically evolves as the historical verification iteration count increases, thereby quantifying the penalty for reducing the similarity of the current retrieval from historical violations and achieving adaptive adjustment of data retrieval accuracy.

[0028] In response to the tender document generation request, the system performs data retrieval through flexible coupling computation. The system applies a retrieval damping coefficient to the retrieval process, coupling vector data with metadata. During this process, the system quantifies the entry boundary features based on the retrieval damping coefficient, using it as a moderating factor to reduce the initially obtained semantic retrieval similarity.

[0029] It is important to note that in this embodiment, the process of constructing the retrieval damping coefficient mapping table is as follows: Based on historical multi-round tender generation records and compliance verification logs, statistical analysis is performed on the feature matching effect and compliance pass rate under different verification iteration numbers; by mining the correspondence between the iteration number and the degree of feature misalignment in historical data, a function mapping logic between the iteration number and the retrieval damping coefficient is pre-constructed and stored in the mapping table. This changes the traditional retrieval model that only focuses on the literal similarity of text, introduces the numerical constraints of the admission boundary features into the retrieval process, eliminates historical data that does not meet business indicator requirements, and directly obtains a set of candidate tenders with high matching degree.

[0030] The system generates preliminary business bids based on the candidate bid set and performs compliance verification on the bids. If the verification fails, the system performs a comparison operation to identify and extract the feature misalignment information between the preliminary business bid and the candidate bid set that caused the verification failure. The feature misalignment information intuitively represents the degree of deviation between the currently generated content and the business compliance red line. If the verification passes, the system directly confirms that the business bid is a compliant text, ends the current search and iteration process, and outputs the bid as the final result.

[0031] The system updates the retrieval damping coefficient in reverse based on the feature misalignment information, and then re-obtains the candidate bid set based on the updated retrieval damping coefficient. This closed-loop processing mechanism directly transforms the results of compliance verification failures into constraint weights in the retrieval process, enabling the system to automatically avoid historical samples that caused verification failures in subsequent searches, reduce indicator misalignments in the generated bids, and ultimately output business bids that meet compliance requirements.

[0032] It should be added that, in the embodiments of the present invention, elastic coupling computation refers to a composite matching algorithm that combines vector retrieval of unstructured semantics with numerical boundaries of structured metadata. It introduces a dynamically updatable retrieval damping coefficient as an elastic adjustment factor to penalize the basic semantic similarity, thereby calculating a comprehensive matching degree that satisfies both semantic relevance and meets the constraints of various high-value data.

[0033] Therefore, the technical solution of this invention, compared with the prior art, can update the retrieval damping coefficient in reverse through feature misalignment information, avoid repeated recall of illegal historical samples, reduce the feature misalignment rate of core business indicators, and improve the compliance accuracy of business tender document generation.

[0034] The following detailed explanation of the steps and flow of the full-link closed-loop processing method for multi-source heterogeneous data, with reference to specific embodiments.

[0035] Specifically, in specific embodiments of the present invention, such as Figure 2 As shown, the method for parsing multi-source heterogeneous data and obtaining semantic features and admission boundary features includes: normalizing the format of the multi-source heterogeneous data to obtain the text stream to be processed; retrieving the delimiter of the text stream to be processed and dividing the text stream based on the delimiter to obtain the basic text sequence; overlapping and truncating the basic text sequence based on a preset window length to generate overlapping text blocks, with a preset number of overlapping characters between adjacent overlapping text blocks; performing high-dimensional space mapping on the overlapping text blocks to obtain semantic features; performing entity recognition processing and context parsing on the overlapping text blocks to determine candidate entity phrases, and performing mapping and matching processing between the candidate entity phrases and the business dictionary to obtain the core elements for admission; and performing timestamp conversion and numerical standard conversion on the core elements for admission to generate admission boundary features.

[0036] In a specific embodiment of the present invention, the received unstructured or semi-structured multi-source heterogeneous data (such as PDF documents, scanned images, or rich text format tender documents) undergoes format normalization. A text parsing component or optical character recognition (OCR) engine is invoked to uniformly strip the original data of layout tags, hidden control characters, and non-text layers, forcibly flattening it into a one-dimensional plain text format, i.e., the text stream to be processed.

[0037] For this text stream, the system uses regular expression matching algorithms or string traversal methods to retrieve preset delimiters (such as newline characters) from the data stream. The system uses the retrieved delimiters as breakpoints to physically segment the global text, removing redundant whitespace characters and obtaining the basic text sequence arranged by sentences or paragraphs.

[0038] Considering that long sentences are very common in scenarios with strong compliance requirements, the system initiates a sliding window sampling logic on the basic text sequence to prevent key indicators from being hard-segmented and truncated. The system sends out sampling windows with a preset window length (e.g., a scalar of 512 characters) along the sequence, and retains a specified preset number of overlapping characters each time it slides. In this embodiment, an overlap band of 20 characters can be used. The output of this operation consists of a data cluster composed of multiple overlapping text blocks. By utilizing the overlap area between adjacent blocks, it ensures that core business indicators and their contextual modifiers that cross the segmentation boundary are not physically destroyed by the segmentation action.

[0039] Subsequently, the data flow enters the parallel feature extraction computation branch. On one hand, the system inputs overlapping text blocks into a pre-trained language encoding network. In this embodiment, a deep learning model with a Transformer architecture can be used to perform local and global weight allocation on the text through a self-attention mechanism, reducing its dimensionality and projecting it into a dense floating-point array under a high-dimensional space mapping, which serves as the semantic features for subsequent calculation of semantic matching degree.

[0040] On the other hand, the system employs named entity recognition sequence labeling algorithms (such as BiLSTM-CRF) to perform contextual concatenation parsing on overlapping text blocks, identifying candidate entity phrases in the text that represent time, funding, or qualification types. The system synchronously reads the business dictionary from the configuration file and compares these candidate entity phrases with the core indicator trie in the business dictionary. Only when a candidate entity phrase successfully matches a dictionary representation is it anchored as a compliance-related core access element, thereby filtering out core business indicator features belonging to rigid constraints.

[0041] Since the extracted core access elements are mostly expressed in natural language, the system performs strong type conversion logic at the underlying level. The system performs timestamp conversion on time-related elements and numerical standard conversion on amount / qualification-related elements. The pure numerical set after this type normalization process is the final access boundary feature, which is encapsulated and passed on to perform subsequent flexible coupling calculations.

[0042] It is important to note that the preset window length is obtained by reading the upper limit of the input dimension of the underlying tensor of the downstream language encoding network and extracting the statistical average of the character lengths of complete natural paragraphs in historical business texts. The system uses the upper limit of the input dimension as a constraint boundary, and selects values ​​that are close to this boundary and greater than the statistical average as the preset window length. This ensures that a single overlapping text block can accommodate the complete business context to the greatest extent possible without triggering truncation and overflow in the encoding network.

[0043] The specific method for obtaining the preset number of overlapping characters is as follows: the span of the core business indicators extracted from the business dictionary and history is calculated in advance, the maximum physical character span value representing the description of coherent entities such as funds and construction period is extracted, and the maximum physical span value is set as the preset number of overlapping characters.

[0044] Specifically, in a specific embodiment of the present invention, the preliminary business tender document undergoes compliance verification, including: parsing the preliminary business tender document to obtain the feature fields to be verified; collecting project category tags based on the tender document generation request, traversing a preset rule relationship table according to the project category tags, and locating the target compliance rule row; extracting the compliance boundary indicators of the target compliance rule row and constructing a compliance value range; performing boundary comparison processing between the feature fields to be verified and the compliance value range to obtain a multi-dimensional boundary overflow vector, and determining whether there is an overflow component greater than zero in the multi-dimensional boundary overflow vector; if so, determining that the preliminary business tender document has failed the tender document compliance verification, obtaining the overflow component greater than zero, and generating feature misalignment information.

[0045] In specific embodiments of the present invention, such as Figure 3 As shown, the text feature extraction algorithm or regular expression matching script is first called to scan and parse the newly generated preliminary business tender text, and forcibly extract the feature fields to be verified that involve substantive constraints in the text (such as specific numerical parameters such as "total project price", "committed delivery days", and "warranty period" generated in the text).

[0046] Subsequently, rule mapping and interval construction operations are performed. The system backtracks and reads the original tender document generation request, extracting project category tags such as "IT hardware procurement" or "infrastructure engineering". Using this project category tag as the primary key, the system traverses the preset rule relationship table in the underlying relational database, matching and locating the target compliance rule row that uniquely corresponds to this business category. Next, the system reads the various compliance boundary indicators loaded in the target compliance rule row, and constructs specific compliance value intervals for each business dimension accordingly. For example, based on the indicators, the project duration requirement is constructed as a closed or semi-closed interval of [0, 180] days.

[0047] The actual values ​​of each extracted feature field to be verified are substituted into the corresponding compliant value range for boundary comparison processing to obtain a multi-dimensional boundary overflow vector. In the data structure of this vector, each dimension (component) directly represents the degree of boundary overflow of a business indicator.

[0048] Finally, the system iterates through all elements within the multidimensional boundary overflow vector to determine if any overflow components are greater than zero. Component values ​​less than or equal to zero indicate that the actual feature has not been violated or is just within the compliance red line; the presence of a component greater than zero indicates a substantial breach of the corresponding dimension's value (e.g., exceeding the project deadline or budget). If the system determines the existence of overflow components greater than zero, it directly intercepts the output process, deeming the preliminary business tender as failing compliance verification. At this point, the system extracts these overflow components greater than zero, serializes and encapsulates them along with their corresponding business feature tags, generating structured feature misalignment information. This feature misalignment information serves as a negative feedback beacon, directly outputting to trigger the subsequent reverse update and recall process of the retrieval damping coefficient. This processing eliminates the ambiguity of natural language comparison, simplifies the underlying logic judgment process of computer-executed compliance verification, and provides numerical basis for the reverse and quantitative update of the retrieval damping coefficient in subsequent steps.

[0049] Conversely, if the system determines that there are no overflow components greater than zero in the multidimensional boundary overflow vector, it indicates that all actual business characteristics in the preliminary business bid have safely fallen within the compliant value range. At this point, the system determines that the preliminary business bid has passed the bid compliance verification. In response to this pass determination, the current preliminary business bid is directly solidified as the final compliant business bid for external output, and the processing lifecycle of the current bid generation request is normally terminated.

[0050] It should be added that the specific acquisition method and construction process of the preset rule relationship table are as follows: Through the data synchronization interface, the system reads the enterprise's internal standard business specification documents and historical core business indicator ledgers. The system executes an information extraction script to parse and extract different project categories (e.g., procurement, engineering construction) and the rigid compliance requirements constrained by each project category, such as maximum budget and minimum warranty period. Subsequently, the system initializes a two-dimensional table structure in the relational database, defining the extracted project category as a unique retrieval primary key, and converting the corresponding rigid compliance requirements into numerical compliance boundary indicators, which are then entered as data column fields under this primary key. After completing the mapping and data writing, the system solidifies it into a preset rule relationship table.

[0051] Specifically, in specific embodiments of the present invention, such as Figure 4As shown, obtaining the multidimensional boundary overflow vector includes: performing semantic text parsing on the feature field to be verified, extracting the non-numerical text of the feature field to be verified, removing the non-numerical text, and converting the removed feature field to be verified into a standard numerical scalar; extracting compliance boundary indicators for the compliance numerical range, and aligning the standard numerical scalar with the compliance boundary indicators in the same dimension; if the current compliance boundary indicator is a lower limit control, subtracting the standard numerical scalar from the compliance boundary indicator to generate a positive overflow value; if the current compliance boundary indicator is an upper limit control, subtracting the compliance boundary indicator from the standard numerical scalar to generate a positive overflow value; and reorganizing the positive overflow values ​​of each dimension into an array to construct the multidimensional boundary overflow vector.

[0052] In a specific embodiment of this invention, a text parsing script or regular expression is invoked to scan the feature field to be validated character by character. The system identifies and extracts non-numerical text such as quantifiers, unit symbols, or modifiers attached to the field, and removes these non-numerical texts from memory. After data cleaning, the system performs strong type conversion on the remaining pure number strings, converting them into standard numerical scalars that can be directly used in arithmetic operations at the computer's underlying level.

[0053] Then, the corresponding compliance value range is read, and the compliance boundary indicators as the benchmark rules are extracted. Based on the business attribute identifier of each feature, the standard numerical scalars obtained from the above transformation are mapped one by one with the corresponding compliance boundary indicators to complete the alignment in the same dimension.

[0054] The system then reads the control type flag for the current dimension and executes differentiated subtraction instructions for different control types to generate a positive overflow value reflecting the extent of the out-of-bounds movement. Specifically: If the current compliance boundary indicator is read as a lower limit control (i.e., the actual core business indicator is required not to be lower than a certain threshold), the system will use the compliance boundary indicator as the minuend and subtract the standard numerical scalar.

[0055] If the current compliance boundary indicator is read as an upper limit control (i.e., the actual core business indicator is required not to exceed a certain threshold), the system will subtract the compliance boundary indicator from the standard numerical scalar.

[0056] By switching the minuend and subtrahend according to the constraint flag, the two opposite out-of-bounds phenomena of being below the lower limit or above the upper limit are transformed into an absolute difference greater than zero, generating a positive overflow value.

[0057] Finally, the system allocates a contiguous address space in memory, extracts the calculated positive overflow values ​​of each data dimension, and performs array reassembly and splicing. The resulting multidimensional array structure after splicing and encapsulation is defined as a multidimensional boundary overflow vector.

[0058] Specifically, in a specific embodiment of the present invention, the flexible coupling calculation of vector data and metadata to obtain a candidate bid set includes: calculating the vector distance between the bid generation request and the vector data to obtain semantic similarity; extracting the target boundary features from the bid generation request and calculating the absolute numerical difference between the target boundary features and the corresponding features in the metadata to obtain the boundary offset; performing a product exponential operation on the retrieval damping coefficient and the boundary offset to generate a similarity decay value; subtracting the similarity decay value from the semantic similarity to obtain the comprehensive matching degree; sorting the comprehensive matching degrees of all vector data in descending order to obtain a comprehensive matching sequence, and constructing a candidate bid set based on the comprehensive matching sequence.

[0059] In a specific embodiment of the present invention, the tender document generation request received from the front end is converted into a high-dimensional request vector. This vector can be calculated by using vector calculation algorithms such as cosine similarity, Euclidean distance, or inner product matching, and the distance can be measured with historical samples in the vector database to output the semantic similarity at the text level.

[0060] The parsing engine is invoked to extract the target boundary features from the request. The system then compares the extracted values ​​with the corresponding features of historical samples in the metadata database, directly calculating the absolute difference between the two and defining it as the boundary offset. This offset objectively quantifies the extent to which historical samples deviate from the actual requirements of the current project.

[0061] Read the retrieval damping coefficient corresponding to the current iteration, multiply it by the calculated boundary offset, and directly use the product as the exponent to generate the similarity decay value. The specific calculation formula is as follows:

[0062] In the formula, This represents the similarity decay value. Indicates the retrieval damping coefficient. Indicates the boundary offset.

[0063] Under this mechanism, when the boundary offset of a historical sample is zero (i.e., fully compliant), the similarity decay value naturally returns to zero, without penalizing its similarity. However, once the core business indicator features it contains exceed the numerical limit, even if the historical sample is highly similar to the literal semantics of the current request, the deviation will be non-linearly amplified by the above exponential formula with the base e.

[0064] The system uses semantic similarity as the minuend and subtracts the generated similarity decay value to obtain the comprehensive matching degree of a single historical data point. Based on this comprehensive matching degree, the system performs descending sorting of the global data and extracts the front-end sequence according to a preset recall threshold, thus completing the construction of the candidate bid set.

[0065] It is important to note that the calculation and acquisition process for the preset recall threshold is as follows: The system obtains the maximum context window capacity when the downstream large language model performs the splicing task as the upper limit of resource constraints, and extracts historical compliance verification logs to statistically analyze the frequency of core business indicator feature misalignment events under different data recall scales. Based on the statistical results, the system calculates the mapping relationship between the recall base and the feature misalignment frequency, and extracts the critical recall value where the misalignment frequency is lower than the warning standard and close to the above-mentioned upper limit of resource constraints, and marks it as the preset recall threshold.

[0066] Specifically, in a specific embodiment of the present invention, outputting a compliant business tender document includes: parsing feature misalignment information and extracting a multidimensional boundary overflow vector; setting a reverse iteration step size and performing a scalar multiplication operation on the multidimensional boundary overflow vector and the reverse iteration step size to generate a damping increment vector; performing a vector summation of the current retrieval damping coefficient and the damping increment vector to obtain a retrieval damping optimization coefficient; responding to the retrieval damping optimization coefficient, re-performing the elastic coupling calculation to obtain a set of corrected candidate tender documents; inputting the set of corrected candidate tender documents into a large language model for secondary tender document text generation, and outputting a compliant business tender document.

[0067] In an embodiment of the present invention, feature misalignment information is received when the aforementioned verification fails. The system calls a parsing script to read the data encapsulation format of this information and directly extracts a multi-dimensional boundary overflow vector representing the degree of business indicator out-of-bounds behavior.

[0068] The system retrieves a preset reverse iteration step size from memory. Specifically, within the numerical range of (0, 1), the system uses a grid search or Bayesian optimization algorithm to find a numerical solution that minimizes the number of correction rounds for each core business indicator without generating reverse overshoot (i.e., suppressing oscillations). This numerical solution is then instantiated as the preset reverse iteration step size. Next, the extracted multidimensional boundary overflow vector is multiplied by the reverse iteration step size to generate a damping increment vector. This damping increment vector represents the penalty weight that needs to be added in the current round; the more severe the feature dimension's out-of-bounds error, the larger its corresponding increment value in the vector.

[0069] The system extracts the currently active damping coefficients from memory and performs a vector summation operation on them and the calculated damping increment vector. Specifically: ; Indicates the retrieval damping coefficient. This represents the retrieval damping optimization coefficient. This represents the damping increment vector.

[0070] By accumulating the values, the system obtains the retrieval damping optimization coefficient after targeted penalty update. The retrieval damping optimization coefficient is the updated comprehensive deduction weight, which is used to forcibly deduct the matching score of historical samples with similar core business indicator violation characteristics in subsequent retrieval matching.

[0071] In response to the optimized retrieval damping coefficient, the system sends it back to the data recall node as the latest penalty parameter and re-executes the aforementioned flexible coupling calculation process. Based on this, the system recalculates the similarity decay value of each historical data and re-sorts and truncates the sequence to obtain a corrected candidate bid set stripped of out-of-bounds interference terms.

[0072] Finally, the system inputs this revised candidate bid set as contextual corpus into the downstream pre-trained large language model. The system triggers the large language model's inference and generation instructions, performs secondary bid text generation, and finally outputs compliant business bids.

[0073] It should be added that the closed-loop correction scheme defined in this solution is mainly applied to scenarios where key indicators exceed limits due to historical corpus contamination when large language models generate professional and compliant text. Specifically, in real bidding and tendering business, a large number of historical tender documents have reference value in terms of literal semantics, but their accompanying values ​​such as budget and schedule are often limited by the project environment at that time, which conflicts with the access boundary characteristics of the current project. If the system only uses conventional one-way retrieval augmented generation (RAG) technology, it is easy to recall historical texts that are semantically perfect but have serious numerical violations, causing the large model to have a compliance illusion.

[0074] To verify the actual technical effectiveness of the closed-loop correction scheme described in this invention, such as... Figure 5As shown, a comparative simulation experiment was constructed for the procurement scenario of industrial automation machinery. The experiment was conducted in a corpus containing 10,000 historical test tender documents. The core high-value data access boundary of the current equipment R&D project was set as follows: the total system budget is controlled within 5 million yuan, and the control latency of the robotic arm must be less than 15 milliseconds. In the control group, the system only used the conventional one-way retrieval augmented generation (RAG) technology. Although the large language model recalled the historically retrieved equipment solutions with highly relevant semantics, it was limited by the pollution of historical high-priced corpora and generated non-compliant text with a budget of 8 million yuan and a latency of 20 milliseconds, producing a typical compliance illusion. In the experimental group, the system entered the closed-loop control loop defined by this invention, analyzed the feedback feature misalignment information and extracted the multi-dimensional boundary overflow vector; then, the system called the preset reverse iteration step size (e.g., 0.45) that had been optimized and solidified in the (0,1) interval through the Bayesian optimization algorithm in the early stage, multiplied it with the overflow vector to generate a damping increment vector, and summed it with the current retrieval damping coefficient execution vector to obtain a retrieval damping optimization coefficient with a high penalty weight. In response to the optimized coefficients, the system re-executes the elastic coupling calculation, penalizing and forcibly stripping historical interference data with high budget and high latency characteristics at the data recall node. Finally, the system inputs the obtained clean and corrected candidate bid set into the large language model. The secondary generated business bids not only perfectly match the industrial automation scenario in their underlying technical logic, but also accurately and simultaneously converge within the rigid compliance red lines of 5 million yuan and 15 milliseconds, fully demonstrating that this mechanism can effectively block historical data contamination and completely eliminate numerical out-of-bounds fluctuations and compliance illusions.

[0075] Specifically, in a specific embodiment of the present invention, the multidimensional boundary overflow vector is multiplied by the reverse iteration step size to generate a damping increment vector. This includes: extracting the overflow components of each dimension in the multidimensional boundary overflow vector, and retrieving historical compliance sample data from metadata based on the current dimension; extracting statistical features from the historical compliance sample data to construct a historical compliance sample distribution matrix, which contains the mean and variance of the historical compliance sample data; calculating the statistical deviation of the overflow components relative to the historical compliance sample distribution matrix to obtain a dynamic deviation scalar; inputting the dynamic deviation scalar into a preset nonlinear activation function for normalization mapping to generate a dynamic severity level weight, and multiplying the reverse iteration step size by the dynamic severity level weight to generate a dynamic iteration step size vector; and performing a Hadamard product operation on the multidimensional boundary overflow vector and the dynamic iteration step size vector to generate a damping increment vector.

[0076] In this embodiment of the invention, the system extracts the overflow components of each dimension in the multidimensional boundary overflow vector. Based on the business attribute identifier corresponding to the current dimension, the system retrieves and reads historical compliance sample data that has passed compliance verification in the underlying metadata repository. The system calls statistical analysis operators to calculate the mathematical expectation and data dispersion of the aforementioned historical compliance sample data in the current dimension and instantiates and constructs a historical compliance sample distribution matrix in memory.

[0077] The system incorporates the currently generated overflow component into the aforementioned distribution matrix and calculates its statistical deviation from the historical normal distribution. The specific deviation calculation employs either the standard deviation mapping or the Mahalanobis distance algorithm. For example, the system calculates the absolute difference between the overflow component and the mean, divides it by the standard deviation, and outputs a dynamic deviation scalar. This dynamic deviation scalar objectively characterizes the rarity of the current out-of-bounds error deviating from the historical normal fluctuation range.

[0078] Subsequently, the calculated dynamic deviation scalar is input into a preset nonlinear activation function (e.g., the Sigmoid or Tanh function) for normalization mapping. Through the nonlinear compression properties of this activation function, the system forces potentially divergent deviation scalars to converge to the open interval (0, 1), generating smooth dynamic severity level weights. Next, the system performs a scalar multiplication operation between the initially preset base backward iteration step size (scalar) and this dynamic severity level weight, thereby generating an independent dynamic iteration step size for each indicator dimension. The system then concatenates the dynamic iteration step sizes of each dimension sequentially to construct a dynamic iteration step size vector.

[0079] The original multidimensional boundary overflow vector is combined with the newly generated dynamic iterative step size vector using the Hadamard product (i.e., multiplying corresponding elements one by one). This ultimately generates the damping increment vector.

[0080] Specifically, in this embodiment of the invention, before re-executing the elastic coupling calculation, a convergence blocking verification step is included, which specifically involves: counting the number of verification iterations of the current tender document generation request and determining whether the number of verification iterations is greater than a preset convergence blocking threshold; if it is not greater, then in response to the retrieval damping optimization coefficient, triggering the re-executing of the elastic coupling calculation; if it is greater, then interrupting the verification iteration, extracting the candidate tender document set corresponding to the maximum comprehensive matching degree in each round of verification iterations already executed, and outputting an alarm business tender document. The alarm business tender document indicates that the candidate tender document set still cannot automatically converge to the tender document compliance verification standard after extreme iterations, and there is feature misalignment information that requires manual review and intervention.

[0081] In a specific embodiment of the present invention, the system directly counts the number of verification iterations that have been executed for the current tender document generation request and determines whether the number of verification iterations is greater than a preset convergence blocking threshold.

[0082] If the value is not greater than the specified value, it indicates that the system is still within a reasonable algorithm convergence period. The system responds to the aforementioned retrieval damping optimization coefficient, triggers the re-execution of the elastic coupling calculation node, and continues the next round of correction.

[0083] If the value is greater than the limit, it indicates a conflict in the core business indicator constraints, and the system immediately interrupts the verification iteration. At this time, the system extracts the candidate bid set with the maximum comprehensive matching degree from each round of verification iterations already executed. The system outputs this set as an alarm business bid, which directly indicates that the candidate bid set still cannot automatically converge to the bid compliance verification standard after extreme iterations, and explicitly includes feature misalignment information that requires manual review and intervention.

[0084] As another embodiment of the present invention, the system completes closed-loop correction and achieves complete iterative exit before triggering the convergence blocking threshold. Specifically, after the system completes the secondary tender document generation for any round, the generated tender document is transferred back to the compliance verification node, the core business indicators in the text are extracted, and boundary comparison processing is performed again.

[0085] The system counts the current number of verification iterations and determines that it is not greater than the preset convergence blocking threshold. Simultaneously, the system determines that the feedback feature misalignment information is empty during verification, meaning that all overflow components in the multidimensional boundary overflow vector calculated by parsing are zero. The combined results indicate that after previous rounds of retrieval damping optimization and corpus cleaning, the latest generated tender document text meets the compliance verification standards.

[0086] In response to the judgment command that the multidimensional boundary overflow vector is zero, the system actively interrupts the closed-loop control loop and exits the verification iteration. The system then suspends the subsequent damping coefficient update and re-execution of the elastic coupling calculation, directly solidifies and encapsulates the text generated in the current round, and outputs it as the final compliant business tender document.

[0087] It is important to note that feature misalignment information refers to a structured negative feedback data packet. When the initial business tender fails compliance verification, this feature misalignment information is formed by serializing and encapsulating the positive-zero overflow component caused by the actual generated feature value exceeding the limit, and the business feature tag corresponding to the overflow component. It is used to trigger and quantify the reverse update of the underlying retrieval damping coefficient.

[0088] This invention also provides a multi-source heterogeneous data end-to-end closed-loop processing system, comprising: a heterogeneous data parsing module for receiving and parsing multi-source heterogeneous data to obtain semantic features and access boundary features; a feature transformation and storage module for converting semantic features into vector data for storage and converting access boundary features into metadata for storage; an initial damping determination module for collecting historical verification iterations and determining the retrieval damping coefficient, wherein the retrieval damping coefficient corresponds one-to-one with the historical verification iterations; and an elastic coupling calculation module for responding to a tender document generation request based on the retrieval damping coefficient. The system performs flexible coupling calculations on vector data and metadata to obtain a candidate bid set. The retrieval damping coefficient is used to quantify the degree of decay of retrieval similarity by the admission boundary features. The compliance misalignment extraction module generates preliminary business bids based on the candidate bid set and performs bid compliance verification on the preliminary business bids. If the bid compliance verification fails, the feature misalignment information between the preliminary business bids and the candidate bid set is extracted. The damping reverse update module updates the retrieval damping coefficient in reverse according to the feature misalignment information, and re-obtains the candidate bid set based on the updated retrieval damping coefficient, outputting compliant business bids.

[0089] Specifically, in this embodiment of the invention, the elastic coupling calculation module includes: a semantic similarity calculation unit, used to calculate the vector distance between the tender document generation request and the vector data to obtain semantic similarity; a boundary offset calculation unit, used to extract the target boundary features in the tender document generation request and calculate the absolute numerical difference between the target boundary features and the corresponding features in the metadata to obtain the boundary offset; a similarity decay generation unit, used to perform a product exponential operation on the retrieval damping coefficient and the boundary offset to generate a similarity decay value; a comprehensive matching calculation unit, used to subtract the similarity decay value from the semantic similarity to obtain the comprehensive matching degree; and a candidate set construction unit, used to sort the comprehensive matching degrees of all vector data in descending order to obtain a comprehensive matching sequence and construct a candidate tender document set based on the comprehensive matching sequence.

[0090] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for closed-loop processing of multi-source heterogeneous data across the entire data link, characterized in that, include: S1 receives multi-source heterogeneous data, parses the multi-source heterogeneous data, and obtains semantic features and access boundary features; S2 converts the semantic features into vector data for storage, and converts the admission boundary features into metadata for storage; S3 collects the historical verification iteration count and determines the retrieval damping coefficient, wherein the retrieval damping coefficient corresponds one-to-one with the historical verification iteration count; S4 responds to the tender document generation request and performs elastic coupling calculation on the vector data and the metadata based on the retrieval damping coefficient to obtain a candidate tender document set. The retrieval damping coefficient is used to quantify the degree of attenuation of the retrieval similarity by the admission boundary features. S5 generates a preliminary business bid based on the candidate bid set, and performs a bid compliance verification on the preliminary business bid. If the bid compliance verification fails, the feature misalignment information between the preliminary business bid and the candidate bid set is extracted. S6 updates the retrieval damping coefficient in reverse based on the feature misalignment information, and re-obtains the candidate tender set based on the updated retrieval damping coefficient, and outputs compliant business tenders.

2. The multi-source heterogeneous data end-to-end closed-loop processing method according to claim 1, characterized in that, The multi-source heterogeneous data is analyzed to obtain semantic features and admission boundary features, including: The multi-source heterogeneous data is format-normalized to obtain the text stream to be processed; Retrieve the delimiter of the text stream to be processed, and divide the text stream to be processed based on the delimiter to obtain the basic text sequence; Based on a preset window length, the basic text sequence is overlapped and truncated to generate overlapping text blocks. There is a preset number of overlapping characters between adjacent overlapping text blocks. Perform high-dimensional space mapping on the overlapping text blocks to obtain the semantic features; The overlapping text blocks are subjected to entity recognition and context parsing to determine candidate entity phrases. The candidate entity phrases are then mapped and matched with the business dictionary to obtain the core elements for admission. The core access elements are converted using timestamps and numerical standards to generate access boundary features.

3. The multi-source heterogeneous data end-to-end closed-loop processing method according to claim 1, characterized in that, The preliminary business tender documents are subject to compliance verification, including: Parse the preliminary business tender document to obtain the feature fields to be verified; Based on the tender document generation request, collect project category tags, traverse the preset rule relationship table according to the project category tags, and locate the target compliance rule row; Extract the compliance boundary indicators of the target compliance rule row and construct the compliance value range; The feature field to be verified is compared with the compliant numerical range by boundary comparison to obtain a multi-dimensional boundary overflow vector, and it is determined whether there is an overflow component greater than zero in the multi-dimensional boundary overflow vector. If it exists, it is determined that the preliminary business tender document has failed the tender document compliance verification, the overflow component with a value greater than zero is obtained, and feature misalignment information is generated.

4. The multi-source heterogeneous data end-to-end closed-loop processing method according to claim 3, characterized in that, Obtain the multidimensional boundary overflow vector, including: Semantic text parsing is performed on the feature field to be verified to extract the non-numerical text of the feature field to be verified, and the non-numerical text is removed. The removed feature field to be verified is then converted into a standard numerical scalar. Extract the compliance boundary indicators of the compliance numerical range, and align the standard numerical scalar with the compliance boundary indicators in the same dimension; If the current compliance boundary indicator is a lower limit control, then the standard numerical scalar value is subtracted from the compliance boundary indicator to generate a positive overflow value; If the current compliance boundary indicator is an upper limit control, then the standard numerical scalar is subtracted from the compliance boundary indicator to generate a positive overflow value; The positive overflow values ​​of each dimension are reorganized into arrays to construct a multidimensional boundary overflow vector.

5. The multi-source heterogeneous data end-to-end closed-loop processing method according to claim 1, characterized in that, The vector data and the metadata are elastically coupled to obtain a set of candidate bids, including: Calculate the vector distance between the tender document generation request and the vector data to obtain semantic similarity; Extract the target boundary features from the tender document generation request, and calculate the absolute numerical difference between the target boundary features and the corresponding features in the metadata to obtain the boundary offset; The similarity decay value is generated by multiplying the retrieval damping coefficient with the boundary offset and performing an exponential operation on the product. The overall matching degree is obtained by subtracting the similarity decay value from the semantic similarity. The comprehensive matching degree of all the vector data is sorted in descending order to obtain a comprehensive matching sequence, and a candidate bid set is constructed based on the comprehensive matching sequence.

6. The multi-source heterogeneous data end-to-end closed-loop processing method according to claim 5, characterized in that, Output compliant business tender documents, including: Analyze the feature misalignment information and extract the multidimensional boundary overflow vector; Set the reverse iteration step size, and perform a scalar multiplication operation between the multidimensional boundary overflow vector and the reverse iteration step size to generate a damping increment vector. The current search damping coefficient is summed with the damping increment vector to obtain the search damping optimization coefficient. In response to the aforementioned retrieval damping optimization coefficient, the elastic coupling calculation is re-executed to obtain a revised candidate tender set; The modified candidate bid set is input into a large language model to generate a second bid text, and a compliant business bid is output.

7. The multi-source heterogeneous data end-to-end closed-loop processing method according to claim 6, characterized in that, The multidimensional boundary overflow vector is multiplied by the reverse iteration step size to generate a damping increment vector, including: Extract the overflow components of each dimension in the multidimensional boundary overflow vector, and based on the current dimension, retrieve historical compliance sample data that has passed compliance verification from the metadata. Statistical features are extracted from the historical compliance sample data to construct a historical compliance sample distribution matrix, which includes the mean and variance of the historical compliance sample data. Calculate the statistical deviation of the overflow component relative to the historical compliant sample distribution matrix to obtain the dynamic deviation scalar; The dynamic deviation scalar input is normalized and mapped using a preset nonlinear activation function to generate dynamic severity level weights. The reverse iteration step size is then multiplied by the dynamic severity level weights to generate a dynamic iteration step size vector. The multidimensional boundary overflow vector and the dynamic iteration step size vector are subjected to a Hadamard product operation to generate a damping increment vector.

8. The multi-source heterogeneous data end-to-end closed-loop processing method according to claim 6, characterized in that, Before re-performing the elastic coupling calculation, a convergence blocking verification step is also included, specifically: Count the number of verification iterations for the current tender document generation request, and determine whether the number of verification iterations is greater than a preset convergence blocking threshold; If it is not greater than, then in response to the retrieval damping optimization coefficient, the re-execution of the elastic coupling calculation is triggered; If the value is greater than the limit, the verification iteration is interrupted, and the candidate bid set corresponding to the maximum comprehensive matching degree in each round of verification iteration is extracted. An alarm business bid is then output. The alarm business bid indicates that the candidate bid set still cannot automatically converge to the bid compliance verification standard after extreme iterations, and there is feature misalignment information that requires manual review and intervention.

9. A multi-source heterogeneous data end-to-end closed-loop processing system, characterized in that, This system is used to execute the multi-source heterogeneous data end-to-end closed-loop processing method according to any one of claims 1 to 8, the system comprising: The heterogeneous data parsing module is used to receive multi-source heterogeneous data, parse the multi-source heterogeneous data, and obtain semantic features and access boundary features. The feature transformation and storage module is used to transform the semantic features into vector data for storage, and to transform the admission boundary features into metadata for storage. The initial damping determination module collects the number of historical verification iterations and determines the retrieval damping coefficient, wherein the retrieval damping coefficient corresponds one-to-one with the number of historical verification iterations; The elastic coupling calculation module, in response to the tender document generation request, performs elastic coupling calculation on the vector data and the metadata based on the retrieval damping coefficient to obtain a set of candidate tender documents. The retrieval damping coefficient is used to quantify the degree of attenuation of the retrieval similarity by the admission boundary features. The compliance misalignment extraction module generates preliminary business bids based on the candidate bid set and performs bid compliance verification on the preliminary business bids. If the bid compliance verification fails, it extracts the feature misalignment information between the preliminary business bids and the candidate bid set. The damping reverse update module updates the retrieval damping coefficient in reverse according to the feature misalignment information, and re-obtains the candidate bid set based on the updated retrieval damping coefficient, and outputs compliant business bids.

10. The multi-source heterogeneous data end-to-end closed-loop processing system according to claim 9, characterized in that, The elastic coupling calculation module includes: A semantic similarity calculation unit is used to calculate the vector distance between the tender document generation request and the vector data to obtain semantic similarity. The boundary offset calculation unit is used to extract the target boundary features from the tender document generation request, and calculate the absolute numerical difference between the target boundary features and the corresponding features in the metadata to obtain the boundary offset. A similarity attenuation generation unit is used to perform a product exponential operation on the retrieval damping coefficient and the boundary offset to generate a similarity attenuation value. A comprehensive matching calculation unit is used to obtain a comprehensive matching degree by subtracting the similarity decay value from the semantic similarity; The candidate set construction unit is used to sort the comprehensive matching degree of all the vector data in descending order, obtain the comprehensive matching sequence, and construct a candidate bid set based on the comprehensive matching sequence.

Images