Lightweight tender intelligent writing method and system based on structured label mapping

Abstract

The present application belongs to the technical field of tender writing management, and discloses a lightweight tender intelligent writing method and system based on structured label mapping. The method comprises: constructing a source end and demand end label library based on enterprise resources and bidding requirements; obtaining a bidding document template and performing paragraph filtering and reconstruction to identify the to-be-filled area, performing label matching reasoning on the legal paragraph, and outputting the candidate matching result, generating the corresponding pre-filled placeholder in the to-be-filled area combined with the double-end label library, and obtaining the preliminary pre-filled bidding document; presenting it to the interactive interface to perform multi-dimensional verification, responding to the in-situ instruction of manual expansion, performing adaptive supplement and incremental precipitation of the expansion label, and performing compliance text content binding through the prompt word engineering configuration to obtain the full-amount label mapping routing table; based on the table, performing full-amount multi-modal label entity replacement and cross-document technology segment embedding and merging, and outputting the target bidding document. The present application realizes high layout stability, low algorithm power consumption and strong compliance of tender collaborative preparation.

Description

Technical Field

[0001] This invention relates to the field of tender document writing and management technology, specifically to a lightweight intelligent tender document writing method and system based on structured tag mapping. Background Technology

[0002] With the advancement of information technology, bidding has become the mainstream business operation model in fields such as engineering construction, goods procurement, and service outsourcing. The accuracy and standardization of bid document preparation directly determine the compliance of the bid and the final probability of winning the bid. However, the traditional bid document preparation process relies heavily on manual, paragraph-by-paragraph filling and page-by-page verification of unstructured text, which has obvious defects in terms of technical implementation and engineering operation. On the one hand, a large amount of repetitive basic enterprise information, qualification documents, historical performance, and personnel files lack standardized technical indexes, leading to repeated manual copying and pasting operations, resulting in extremely low automation and efficiency in bid document preparation. On the other hand, when manually filling in unstructured forms, visual fatigue or logical oversights can easily lead to omissions, incorrect entries, inconsistent formatting, and multiple inconsistencies in underlying data information. Under the strict compliance review of bidding, these textual and formatting flaws can easily trigger negative evaluations, leading to a serious risk of bid rejection.

[0003] In recent years, with the development of natural language processing technology and general artificial intelligence large language models (hereinafter referred to as large language models), some pure AI end-to-end automatic tender document generation solutions have emerged in the industry. However, due to the heterogeneous characteristics of tender documents, such as long texts, high noise, and diverse layout formats, existing pure AI generation solutions have serious shortcomings in document semantic parsing and layout control. They are prone to outputting tender documents with uncontrollable content, biased contextual logic, and disordered layout formats, which cannot meet the hard requirements of certainty and compliance in bidding scenarios. More importantly, because existing technologies lack a pre-processing mechanism for complex documents, they usually need to input massive amounts of unstructured text into large models for multiple rounds of contextual interaction. This not only leads to extremely large computational resource consumption and token window consumption, significantly increasing the system's computational deployment cost, but also significantly reduces the stability of AI generation results due to the distraction of attention in long texts.

[0004] In summary, the current field of intelligent tender document preparation still lacks a professional writing system that can deeply integrate standardized corporate asset information, structured features of tender documents, and AI semantic filling, resulting in low resource consumption, high determinism, and high reusability. This has led to tender document preparation technology remaining in a bottleneck state of inefficiency, error-proneness, high cost, and wasted computing power for a long time. Summary of the Invention

[0005] The purpose of this application is to provide a lightweight intelligent tender document writing method and system based on structured tag mapping to solve the above-mentioned technical problems.

[0006] Firstly, this application proposes a lightweight intelligent tender document writing method based on structured tag mapping, which includes the following steps: S1. Based on the core resource information of the enterprise and the core demand information of the bidding project, perform structured parsing and standardized tag processing, and map and construct the source tag library and the demand tag library respectively; S2. Obtain the tender document template document, perform paragraph filtering and reconstruction on the tender document template document to identify the areas to be filled in the tender document template document and retain the legal paragraphs; calculate the text hash value of the legal paragraphs to determine whether they hit the preset cache, and output the candidate matching results based on the hit results or the results of tag matching inference for the legal paragraphs that do not hit the preset cache; based on the candidate matching results, combined with the source tag library and the demand tag library, generate the corresponding pre-filled placeholders in the areas to be filled to obtain the preliminary pre-filled tender document; S3. Submit the preliminary pre-filled tender documents to the human interaction interface to perform multi-dimensional verification, respond to the human in-situ processing instructions to perform adaptive supplementation and incremental accumulation of extended tags for matching deviations, and perform compliant text content binding on extended tags through prompt word engineering configuration to obtain a full tag mapping routing table containing full compliant text bindings. S4. Based on the full tag mapping routing table, retrieve the corresponding actual business content from the source tag library or the demand tag library, perform full multimodal tag entity replacement and cross-document technical fragment embedding and merging, and output compliant target tender documents.

[0007] In the aforementioned technical solution, by constructing a decoupled standardized tag library for both the source and demand sides, the process of parsing and writing heavy documents is transformed into lightweight multi-dimensional tag routing mapping and on-demand rendering. In the computational process, a traffic control mechanism combining static caching interception and dynamic large language model deep inference significantly reduces the frequency of unnecessary cloud interface calls and eliminates the computational cost of large models processing long paragraphs of text. Finally, the introduction of reverse replacement and multimodal isolation mechanisms fundamentally solves the problem of cursor index misalignment caused by text length expansion, which is easily generated in traditional long text automatic typesetting, achieving high-concurrency automated intelligent output of highly compliant tender documents with zero page layout corruption.

[0008] Furthermore, in step S1, the core resource information of the enterprise includes basic enterprise information, qualification documents, technical solutions, performance certificates and personnel information; the source tag library is divided into basic information tag group, qualification document tag group, technical solution tag group, performance tag group and personnel tag group according to information type, and each tag is associated with a unique identifier and update timestamp to achieve dynamic incremental updates.

[0009] In the above technical solution, by classifying the core assets of an enterprise into five groups and assigning them globally unique asset identifiers, a strongly typed data routing boundary is established at the underlying level. Incremental tracking is performed on each independent tag entity using timestamps, so as to realize the dynamic incremental update of the local asset dictionary in a high-concurrency, non-full-refresh manner, which ensures the timeliness of the source tag library during high-frequency calls and the consistency of the underlying data.

[0010] Furthermore, in step S1, the core requirements information of the bidding project includes the bidding project name, bidding project number, bidding agency, purchaser, bidding requirements, evaluation criteria, and bid deadline; after constructing the demand-side tag library, the program automatically establishes a multi-dimensional semantic association network between the demand-side tag library and the source-side tag library.

[0011] In the above technical solution, by performing tagging extraction on the core hard indicators of heterogeneous bidding texts, a multi-dimensional semantic association network between the demand side and the source side assets is automatically connected and assembled. In the subsequent rule reasoning pre-stage, semantic alignment of cross-document features is achieved, which greatly shortens the feature search path during subsequent adaptive matching and enhances the system's robustness against heterogeneous text noise.

[0012] Furthermore, in step S2, paragraph filtering and reconstruction are performed on the tender document template document, including: traversing all paragraphs of the tender document template document, removing and filtering all invalid empty paragraphs or blank lines in the document; and checking whether the remaining paragraphs have the characteristics of the fill area or the business license, skipping non-fill area paragraphs that do not have the characteristics of the fill area and the business license, and using them as context references for subsequent paragraphs to be filled. The judgment rules used when checking whether the retained paragraphs have the characteristics of an infill area or a business license include: Determine if a paragraph contains a blank slice, a slice with underlined characters, or a slice with an underline style. Determine whether a paragraph ends with a full-width or half-width colon; Determine whether a paragraph contains preset guiding keywords; Determine whether a paragraph contains time markers in chronological order, and whether the time markers are separated only by spaces or underscores. And determine whether a paragraph contains cross-modal related feature words.

[0013] In the aforementioned technical solution, a low-energy-consumption static document structure pre-filter is constructed before entering the heavy-duty AI recognition chain by utilizing multiple deterministic format, symbol, keyword, and temporal arrangement features. This not only removes visual blank noise that contributes nothing to entity mapping in situ, but more importantly, it retains pure non-filling area text that "lacks filling area features but contains core semantics" as contextual background data. This provides high-fidelity environmental awareness context for subsequent deep inference in large models while significantly reducing the volume of data input to large models, achieving a balance between high accuracy and low computational cost.

[0014] Furthermore, step S2 involves determining whether a preset cache has been hit and performing tag matching inference, including: Step 1: Perform text slicing preprocessing on the valid paragraphs, and calculate a unique SHA-256 hash value based on the preprocessed paragraph text content; Step 2: Use the SHA-256 hash value as the query key to retrieve the matching library that is the preset cache; Step 3: If the result is a match, the tag mapping relationship recorded in the matching library is read directly to generate the pre-filled tag placeholder result for the corresponding paragraph and the interface call is skipped; if the result is a mismatch, the text data of the corresponding paragraph is sent to the large language model for tag matching inference to output the large model inference matching result. Step 4: Summarize the pre-filled label placeholder results and the large model inference matching results to output candidate matching results in a unified manner; The pairing library is a pre-built database cache table that stores cached records. Each record in the cache contains at least: SHA-256 hash value, normalized text, original text for display, label mapping structured array, and number of filling areas. The label mapping structured array supports two positioning types: region filling positioning type and location positioning type. The steps for building the pairing library include: uploading a docx template document in a known format and extracting paragraphs with fill areas; performing segmentation preprocessing on the paragraphs; normalizing the entire text of the paragraphs and calculating the SHA-256 hash value; receiving configuration instructions through the management interface to assign corresponding business tags to each fill area to build a tag mapping structured array; and finally saving the records containing the SHA-256 hash value, normalized text, original text for display, tag mapping structured array, and number of fill areas to the pairing library.

[0015] The aforementioned technical solution integrates highly deterministic cryptographic SHA-256 hash representation technology with a multi-dimensional hybrid location array for storage. This establishes a static defense at the software architecture's entry point. When similar or identical tender document segments are input, high-speed cache interception based on the primary key is achieved at the microsecond level, directly reading the mapping relationship in memory and performing in-situ filling. This process completely eliminates high-latency unstructured large-scale model interface calls in the cloud, resulting in an exponential improvement in system throughput, while simultaneously ensuring highly deterministic reproduction of the basic format at the physical layer.

[0016] Furthermore, the large language model performs label matching inference through the following two rounds of controlled recognition steps: Based on the slice positions after text slicing preprocessing, dynamically construct in memory a slice with... <rn>The paragraph text is marked with a dedicated slice index, and the plain text content of the adjacent paragraph above it is appended in place before each paragraph as an enhanced contextual reference. After all unprocessed paragraphs are constructed, the data is fully delivered, and a structured output constraint mechanism is simultaneously enabled when calling the API to ensure that the model output conforms to the JSON format. Recognition is performed through two rounds of controlled calls: the first round controls the large language model based on the input... <rn>The paragraph text with dedicated slice index tags performs a binary classification task to determine the existence of fill-in-the-blank entries. Invalid paragraphs that do not require labeling are filtered out using logical filtering. The filtered, simplified paragraph data is then input into the second round of calls, controlling the large language model to perform a multi-dimensional semantic matching task, matching each fill-in-the-blank area accordingly. <rn>The sequence number slice is used to match the corresponding target business tag name, and the candidate matching results are output uniformly.

[0017] In the above technical solution, dynamic timing tags are applied to heterogeneous text fragments in memory. <rn>Furthermore, dual-sided context enhancement perfectly eliminates the temporal and spatial localization errors caused by concurrent input of multiple paragraphs. By decoupling and cascading the originally complex open label recognition into two rounds of controlled recognition tasks: the first round uses a lightweight binary classification task to quickly intercept and remove invalid text at the edges, greatly reducing the amount of core context data in the second round; the second round, combined with the underlying protocol-level JSON strong constraint output, completely blocks the character illusion and structural collapse problems inherent in large language models when performing multi-dimensional semantic matching, and the regularity and parsing rate of the output results reach the industrial-grade lossless standard.

[0018] Furthermore, before generating corresponding pre-filled placeholders in the area to be filled based on the candidate matching results, the method also includes performing boundary validity checks on the candidate matching results. The boundary validity checks include the following filtering steps: Paragraph boundary filtering steps: Extract the target paragraph number from the candidate matching results and remove out-of-bounds candidate data that exceeds the current legal paragraph number range; Slice boundary filtering steps: Extract the slice index number from the candidate matching results and remove out-of-bounds slice data that exceed the total number of slices in the corresponding legal paragraph; Tag compliance filtering steps: Detect tag names in candidate matching results, remove invalid phantom data with empty tag names in situ, and compare with the preset business tag whitelist to filter and remove illegal tags; Any abnormal interference data that triggers any filtering step will be discarded and synchronously output to the exception logger for evidence preservation.

[0019] In the above technical solution, by parsing data packets into paragraph boundaries, slice boundaries, and a business whitelist for three-layer hard-core algebraic range assertion, serious bugs such as memory overflows, array overflows, and null pointer crashes caused by large model probability sampling can be discarded and isolated in situ before physical replacement rendering. This achieves the output of highly stable and physically secure streaming data even under complex weakly supervised heterogeneous data streams.

[0020] Furthermore, the replacement logic for generating corresponding pre-filled placeholders in the unfilled area in step S2 includes the following steps: The collaborative replacement control steps involve initiating a paragraph-level reverse index compensation sorting mechanism for the matching list after all entries have passed the boundary validity check. For multimodal tag isolation operations that change the paragraph structure, the operations are executed in reverse order from the end of the document. For slice-level text replacement within a paragraph, the operations are executed in forward order to eliminate the risk of index offset errors in the preceding document due to text replacement or length changes. The in-place text placeholder step for slices involves using the tag names in the corresponding source tag library or demand tag library to perform in-place text placeholders while preserving the original layout style and format for blank or underlined slices. The boundary appending and deduplication steps are as follows: for paragraphs ending with a colon, the corresponding tag name is directly appended after the original text; while for slices with existing tag placeholders, the process is skipped to prevent duplicate replacement. The cross-modal content isolation step automatically inserts cross-modal content on a new line below the current paragraph for specific multimodal image or document tags, generating a preliminary pre-filled tender document.

[0021] In the aforementioned technical solution, a dual-layer, cross-scale collaborative typesetting control algorithm—combining macroscopic "reverse order from the end of the document" and microscopic "forward order of paragraph slices"—effectively overcomes the cursor collapse and misalignment problems caused by the expansion or reduction of the number of preceding characters when traditional word processing engines perform physical replacement of non-uniform-length text. By combining colon appending, deduplication skipping, and multimodal line-level physical isolation design, the system can securely embed high-density mixed text and image assets in a single-batch lossless rendering mode, while fully preserving the original document formatting, significantly improving rendering robustness under large-scale concurrent replacement.

[0022] Furthermore, in step S3, in response to the manual in-situ processing instruction, adaptive supplementation of extended labels is performed to address the matching deviation, specifically including: The preliminary pre-filled bid documents are output and submitted to the human interaction interface, where humans perform multi-dimensional verification and checks on the bid documents. During the manual review process, the accuracy, completeness and compliance of the label filling are manually checked, the semantic matching degree between the label and the original area to be filled is confirmed, and the enterprise asset information or project core requirement information carried by the label is reviewed to see if it meets the requirements of the tender documents. If manual verification reveals filling errors, matching deviations, or missing regions, manual intervention can initiate on-site manual processing commands through the interactive interface to manually replace labels, correct deviations, or adaptively supplement missing labels, thereby correcting the matching decision boundary at the interactive level.

[0023] In the above technical solution, by structurally mapping all tag information onto the interactive panel, users can quickly and in-situ correct and supplement only the "tag metadata layer" without physically reading and writing large Word documents. This reduces the memory load when interacting with heavy heterogeneous documents and dynamically corrects matching dead zones caused by extreme long-tail distribution in pure machine model inference with extremely low human intervention overhead.

[0024] Furthermore, the adaptive supplementation and incremental accumulation of extended tags in step S3 also includes in-situ derivation and adaptive expansion steps for extended tags in unmatched regions: In response to human-input interactive commands, at the human-computer interaction level, it is determined by humans whether the current unmatched area can reuse existing tags in the source tag library or the demand tag library; If it is determined manually that the existing tags do not meet the specific requirements of this tender, dynamic AI-generated tags will be generated in response to the manual in-situ derivation instruction. The dynamic AI-generated tags will be encapsulated using a unified feature expression format that includes a specific generation domain prefix identifier. Based on the interactive prompts, the newly added AI-generated tags are annotated with semantic dimensions to clarify the content requirements of the unstructured content to be filled in corresponding to the extended tags. The dynamic AI-generated tags are then used as incremental data to be accumulated and updated to the corresponding source tag library or demand tag library.

[0025] The aforementioned technical solution provides a self-evolving tag meta-ecosystem adaptive expansion protocol. When encountering entirely new or suddenly emerging highly customized tender document requirements, it allows for in-situ online derivation of unstructured extended tags wrapped with unified domain prefix identifiers, which are then hot-deposited back into the local knowledge base as incremental high-dimensional spatial data. This avoids the format rigidity of the enterprise knowledge base caused by changes in bidding industry standards, enabling the underlying graph database to maintain structural stability while possessing unlimited dynamic semantic generalization capabilities and self-growing asset pool capabilities.

[0026] Furthermore, the full multimodal tag entity replacement and cross-document technology fragment embedding and merging in step S4 includes the following steps: Retrieve all pre-set and supplemented tags and their corresponding actual business content from the full tag mapping routing table, and automatically replace all tag placeholders in the tender document template document in place through document rendering; Once all tags have been replaced and rendered, the file structure is saved and a complete technical compliance target tender document is output. During the replacement process, a split replacement is performed based on the cross-modal attributes of the tag: For qualification document tags, automatically retrieve the corresponding scanned copy or electronic file fragment of the enterprise document and insert it in situ into the target document paragraph in the form of cross-modal image or structured object; For technical solution tags, the system automatically extracts the corresponding standard technical document content or compliant long text paragraphs generated by the large language model, and performs seamless embedding and merging of cross-document technical fragments.

[0027] In the above technical solution, a highly scalable "multimodal rendering execution offloading bus" is established, which delivers the final confirmation results of the underlying full tag routing mapping relationship to the physical layer for rendering in one go. For high-resolution rasterized image objects and large-scale continuous long text unstructured technical chapters, the system can perform adaptive file stream pipeline splicing and node-level fusion embedding across the underlying XML tree of technical documents. This design completely eliminates the huge I / O performance bottleneck caused by repeatedly opening, closing, and saving documents in traditional tender document synthesis, and realizes integrated, lossless, low-latency extreme rendering and merging assembly of complex heterogeneous documents with graphics and text.

[0028] Secondly, this application proposes a lightweight intelligent tender document writing system based on structured tag mapping, which includes: The dual-end tag mapping construction module is configured to perform structured parsing and standardized tag processing based on various core resource information of enterprises and core demand information of bidding projects, and to map and build source tag library and demand tag library respectively; The unfilled area identification and filling module is configured to acquire the tender document template document, perform paragraph filtering and reconstruction on the tender document template document to identify the unfilled areas in the tender document template document and retain the legal paragraphs; calculate the text hash value of the legal paragraphs to determine whether they hit the preset cache, and output the candidate matching results based on the hit results or the results of tag matching inference for legal paragraphs that do not hit the preset cache; based on the candidate matching results, combined with the source tag library and the demand tag library, generate corresponding pre-filled placeholders in the unfilled areas to obtain the preliminary pre-filled tender document; The interactive verification tag accumulation module is configured to present the preliminary pre-filled tender documents to the human interactive interface to perform multi-dimensional verification, respond to human in-situ processing instructions to perform adaptive supplementation and incremental accumulation of extended tags for matching deviations, and perform compliant text content binding on extended tags through prompt word engineering configuration to obtain a full tag mapping routing table containing full compliant text bindings. The multimodal replacement tender document generation module is configured based on a full tag mapping routing table. It retrieves the corresponding actual business content from the source tag library or the demand tag library, performs full multimodal tag entity replacement and cross-document technical fragment embedding and merging, and outputs a compliant target tender document.

[0029] Compared with the prior art, the beneficial results of the present invention are as follows: (1) This application overcomes the inherent weaknesses in layout stability of traditional tender document writing and pure AI tender document generation by deeply integrating label standardization and lightweight structure. Traditional manual writing is inefficient and prone to omissions, while pure AI generation is prone to serious collapse of the tender document format due to character fluctuations in the output of long texts from large models. This invention classifies and sorts heterogeneous enterprise core resource information and bidding requirement information into standardized tags at the front end, and connects a multi-dimensional semantic association network at the bottom layer, which greatly shortens the feature search path; at the same time, it performs lightweight reconstruction and standardized structural transformation of the tender document template at the document level, and transforms unstructured text into an intermediate skeleton JSON file with accurate physical coordinates and feature tags without loss. This "placeholder" architecture design, combined with the two-layer collaborative replacement control through paragraph-level reverse and slice-level forward order, can absolutely maintain the original template layout style format when replacing full multimodal tag entities and embedding and merging cross-document technical fragments, ensuring the robust stability of the final output target tender document in terms of physical format layout and text-image continuity.

[0030] (2) This invention uses lightweight intermediate files, inputting only highly purified core slice features and enhanced contextual reference text spliced ​​in situ for specific paragraphs into the large model, and employs a system prompt word engineering that includes multi-dimensional constraints such as structural definitions, whitelists, and exclusion rules for strong constraints. Accurate identification is achieved through two rounds of controlled invocation. The first round performs a binary classification task based on the existence of fill-in-the-blank entries to eliminate invalid paragraphs, and the second round performs a multi-dimensional semantic matching task, thereby efficiently completing the business tag name matching task for the sequence slices corresponding to the full fill-in area in controlled inference.

[0031] (3) At the interface feedback front-end, this invention establishes a four-fold full-volume boundary legality verification filtering network consisting of paragraph boundary filtering, slice boundary filtering, tag compliance filtering, and business whitelist. This network accurately intercepts and discards abnormal data online, ensuring the high robustness of the pipeline. Simultaneously, the tag mapping relationship in the preliminary pre-filled bid document, which contains the full tag mapping relationship, is presented to the human-computer interaction interface, supporting manual tag replacement, deviation correction, and logical supplementation through expert experience. For long-tail and unmatched special fill-in areas, this invention supports dynamic AI-generated class tags that are encapsulated in a unified feature expression format containing a specific generation domain prefix identifier and are labeled semantically at runtime. Through prompt word engineering, the large model is controlled to generate highly compliant text content that conforms to the formatting specifications and engineering industry standards. After manual review and correction, entity-level relationship binding is performed in the underlying database, which is then used as incremental data and automatically updated to the tag library. This design not only enables the judgment rules and decision boundaries to have dynamic evolution and adaptive expansion capabilities, but also ensures that the final generated bid document content has 100% legal and technical compliance in terms of business indicator response. Attached Figure Description

[0032] The accompanying drawings are included to provide a further understanding of the embodiments and are incorporated in and constitute a part of this specification. The drawings illustrate embodiments and, together with the description, serve to explain the principles of the invention. Many anticipated advantages of the embodiments and other embodiments of the invention will be readily recognized as they become better understood through reference to the following detailed description. Elements in the drawings are not necessarily to scale. The same reference numerals refer to corresponding similar parts.

[0033] Figure 1 This is a flowchart of a lightweight intelligent tender document writing method based on structured tag mapping according to an embodiment of this application; Figure 2 This is a flowchart of the steps of a lightweight intelligent tender document writing method based on structured tag mapping according to a specific embodiment of this application; Figure 3 This is a flowchart of the structured parsing and intelligent filling process for the area to be filled in the tender document according to a specific embodiment of this application; Figure 4 This is a framework diagram of a lightweight intelligent tender document writing system based on structured tag mapping, according to an embodiment of this application. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail 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.

[0035] refer to Figure 1 , Figure 1 A flowchart illustrating a lightweight intelligent tender document writing method based on structured tag mapping, according to an embodiment of this application, is shown in the figure. The method includes the following steps: S1. Based on the enterprise's various core resource information and the core demand information of bidding projects, perform structured parsing and standardized tag processing, and map and construct source-side tag library and demand-side tag library respectively.

[0036] In some specific implementations, the construction of the source-end tag library first involves classifying, organizing, and standardizing the tagging of various core resource information of the enterprise. During this process, all core enterprise information is uniformly formatted into pre-defined tags of {{tag name}}. Core enterprise information broadly includes key assets such as basic enterprise information, qualification documents, technical solutions, performance records, and personnel information. To achieve cross-modal data carrying, qualification document tags accurately correspond to scanned copies or electronic files of relevant documents, while technical solution tags directly correspond to specific engineering and technical documents. This establishes a standardized enterprise tag library at the underlying level that can be categorized, archived, and dynamically updated, ensuring that the tag pool is highly consistent with the actual information of the enterprise.

[0037] In some specific implementations, the construction of the demand-side tag library manifests as the tagged parsing of core demand information for bidding projects. First, specific bidding documents are uploaded. A structured parsing program is then used to perform deep semantic analysis on these documents, extracting the core bidding information with high precision. This information is then encapsulated using a unified tag system with the format {{tag name}}, forming a highly targeted bidding tag library. The core bidding information includes, but is not limited to, key business and technical response points such as the bidding project name, bidding project number, bidding agency, purchaser, bidding requirements, evaluation criteria, and bid deadline. Specifically, the structured parsing program can utilize existing open-source PDF conversion tools (such as PyMuPDF or PDFPlumber) with preset regular expressions, or call existing multimodal document layout parsing models (such as LayoutLM) and optical character recognition (OCR) components to perform document layout analysis, paragraph text extraction, and key-value pair feature alignment on the input heterogeneous bidding documents, thereby achieving standardized extraction of the aforementioned key business and technical response points.

[0038] S2. Obtain the tender document template document, perform paragraph filtering and reconstruction on the tender document template document to identify the areas to be filled in the tender document template document and retain the legal paragraphs; calculate the text hash value of the legal paragraphs to determine whether they hit the preset cache, and output the candidate matching results based on the hit results or the results of tag matching inference for the legal paragraphs that do not hit the preset cache; based on the candidate matching results, combined with the source tag library and the demand tag library, generate the corresponding pre-filled placeholders in the areas to be filled to obtain the preliminary pre-filled tender document.

[0039] In some specific embodiments, step S2 involves performing paragraph filtering and reconstruction on the tender document template document, including: traversing all paragraphs of the tender document template document, removing and filtering all invalid empty paragraphs or blank lines in the document; and checking whether the remaining paragraphs have fill area characteristics or business license characteristics, skipping non-fill area paragraphs that do not have fill area characteristics and business license characteristics, and using them as context references for subsequent paragraphs to be filled. The judgment rules used when checking whether the retained paragraphs have the characteristics of an infill area or a business license include: Determine if a paragraph contains a blank slice, a slice with underlined characters, or a slice with an underline style. Determine whether a paragraph ends with a full-width or half-width colon; Determine whether a paragraph contains preset guiding keywords; Determine whether a paragraph contains time markers in chronological order, and whether the time markers are separated only by spaces or underscores. And determine whether a paragraph contains cross-modal related feature words.

[0040] In some specific embodiments, step S2, which determines whether a preset cache has been hit and performs tag matching inference, includes: Step 1: Perform text slicing preprocessing on the valid paragraphs, and calculate a unique SHA-256 hash value based on the preprocessed paragraph text content; Step 2: Use the SHA-256 hash value as the query key to retrieve the matching library that is the preset cache; Specifically, the pairing library is a pre-built database cache table that stores cached records. Each record in the cache contains at least: SHA-256 hash value, normalized text, original text for display, a label mapping structured array, and the number of fill areas. The label mapping structured array supports two positioning types: region filling positioning type and location positioning type. The construction steps of the pairing library include: uploading a docx template document of a known format and extracting paragraphs with fill areas; performing slicing preprocessing on the paragraphs; normalizing the entire text of the paragraphs and calculating the SHA-256 hash value; receiving configuration instructions through the management interface to assign corresponding business tags to each fill area to construct the label mapping structured array; and finally saving the record containing the SHA-256 hash value, normalized text, original text for display, label mapping structured array, and number of fill areas to the database cache table.

[0041] Step 3: If the result is a match, the tag mapping relationship recorded in the matching library is read directly to generate the pre-filled tag placeholder result for the corresponding paragraph and the interface call is skipped; if the result is a mismatch, the text data of the corresponding paragraph is sent to the large language model for tag matching inference to output the large model inference matching result. Step 4: Summarize the pre-filled label placeholder results and the large model inference matching results to output candidate matching results in a unified manner; The large language model performs label matching inference through the following two rounds of controlled recognition steps: Based on the slice positions after text slicing preprocessing, dynamically construct in memory a slice with... <rn>The paragraph text is marked with a dedicated slice index, and the plain text content of the adjacent paragraph above it is appended in place before each paragraph as an enhanced contextual reference. After all unprocessed paragraphs are constructed, the data is fully delivered, and a structured output constraint mechanism is simultaneously enabled when calling the API to ensure that the model output conforms to the JSON format. Recognition is performed through two rounds of controlled calls: the first round controls the large language model based on the input... <rn>The paragraph text with dedicated slice index tags performs a binary classification task to determine the existence of fill-in-the-blank entries. Invalid paragraphs that do not require labeling are filtered out using logical filtering. The filtered, simplified paragraph data is then input into the second round of calls, controlling the large language model to perform a multi-dimensional semantic matching task, matching each fill-in-the-blank area accordingly. <rn>The sequence number slice is used to match the corresponding target business tag name, and the candidate matching results are output uniformly.

[0042] In some specific embodiments, before generating corresponding pre-filled placeholders in the area to be filled based on the candidate matching results, the method further includes performing boundary validity checks on the candidate matching results. The boundary validity checks include the following filtering steps: Paragraph boundary filtering steps: Extract the target paragraph number from the candidate matching results and remove out-of-bounds candidate data that exceeds the current legal paragraph number range; Slice boundary filtering steps: Extract the slice index number from the candidate matching results and remove out-of-bounds slice data that exceed the total number of slices in the corresponding legal paragraph; Tag compliance filtering steps: Detect tag names in candidate matching results, remove invalid phantom data with empty tag names in situ, and compare with the preset business tag whitelist to filter and remove illegal tags; Any abnormal interference data that triggers any filtering step will be discarded and synchronously output to the exception logger for evidence preservation.

[0043] In some specific embodiments, the replacement logic for generating corresponding pre-filled placeholders in the unfilled area in step S2 includes the following steps: The collaborative replacement control steps involve initiating a paragraph-level reverse index compensation sorting mechanism for the matching list after all entries have passed the boundary validity check. For multimodal tag isolation operations that change the paragraph structure, the operations are executed in reverse order from the end of the document. For slice-level text replacement within a paragraph, the operations are executed in forward order to eliminate the risk of index offset errors in the preceding document due to text replacement or length changes. The in-place text placeholder step for slices involves using the tag names in the corresponding source tag library or demand tag library to perform in-place text placeholders while preserving the original layout style and format for blank or underlined slices. The boundary appending and deduplication steps are as follows: for paragraphs ending with a colon, the corresponding tag name is directly appended after the original text; while for slices with existing tag placeholders, the process is skipped to prevent duplicate replacement. The cross-modal content isolation step automatically inserts cross-modal content on a new line below the current paragraph for specific multimodal image or document tags, generating a preliminary pre-filled tender document.

[0044] S3. Submit the preliminary pre-filled tender documents to the human interaction interface to perform multi-dimensional verification. Respond to the human in-situ processing instructions to perform adaptive supplementation and incremental accumulation of extended tags for matching deviations. And perform compliant text content binding on extended tags through prompt word engineering configuration to obtain a full tag mapping routing table containing full compliant text bindings.

[0045] In some specific embodiments, a preliminary pre-filled tender document is output and presented to a human-interactive interface, where a human performs multi-dimensional verification and checks on the preliminary pre-filled tender document. During this review process, the human focuses on verifying the accuracy, completeness, and compliance of the filled tags to confirm the semantic matching degree between the current tags and the original areas to be filled, and to review whether the enterprise asset information or core project requirement information carried by the tags fully meets the requirements of the tender document. If the human finds filling errors, matching deviations, or area omissions during the verification, the human initiates manual on-site processing of the matching deviations through the interactive interface, manually performing tag replacement, deviation correction, or adaptive logical supplementation of missing tags. This dynamically corrects the matching decision boundary at the interactive level, and encapsulates the dynamically corrected full tag mapping relationship to generate the corresponding full tag mapping routing table.

[0046] In some specific embodiments, the focus is on blank or unmatched areas that have not been intelligently filled with tags after review. Responding to human-input interactive commands, at the human-computer interaction level, it is determined whether existing tags in the source or demand-side tag library can be reused for the current unmatched area. If the human determines that none of the existing tags meet the specific filling requirements of this tender, in response to the human's in-situ derivation command, a dynamic AI-generated tag is generated as an extended tag. The dynamic AI-generated tag is encapsulated using a unified feature expression format containing a specific generation domain prefix. During this adaptive supplementation process, the human performs pre-semantic dimension annotation on the newly added dynamic AI-generated tag according to interactive prompts, explicitly defining the content requirements of the unstructured content to be filled corresponding to the extended tag. Subsequently, this dynamic AI-generated tag with the new semantic annotation is used as incremental data for extended tags and updated in the source or demand-side tag library, completing the adaptive expansion of the tag ecosystem.

[0047] Specifically, the newly derived dynamic AI-generated tags undergo a process of configuring prompt words, rendering with a large language model, and binding the final content review. For each dynamic AI-generated tag, a custom AI-generated prompt word is written and configured manually. This prompt word not only includes constraints on the input data structure but also explicitly limits the target technical content requirements, formatting specifications, current engineering industry standards, and key points of deep adaptation with the current tender documents. The configured prompt word is then input into the large language model interface, where the large language model automatically generates highly compliant target tag content based on the defined context. After the large language model automatically generates the text content, the generated unstructured content is presented again to the interactive interface for final compliance review and text detail correction by a human. When manual verification confirms the accuracy and triggers the final binding instruction, the underlying database responds to the binding instruction, performs entity-level relationship binding between the verified compliant plain text or complex graphic content and the corresponding dynamically generated AI-generated tags, and updates the full tag mapping routing table in place to obtain a full tag mapping routing table containing the binding of all compliant text, thus preparing a complete data alignment network for subsequent full multimodal tag entity replacement.

[0048] S4. Based on the full tag mapping routing table, retrieve the corresponding actual business content from the source tag library or the demand tag library, perform full multimodal tag entity replacement and cross-document technical fragment embedding and merging, and output compliant target tender documents.

[0049] In some specific embodiments, automated multimodal entity replacement and cross-document fragment merging rendering based on a full tag mapping routing table yield the final target tender document. After obtaining all pre-set and supplemented tags and their corresponding actual content in the full tag mapping routing table, the built-in document rendering program automatically replaces all tag placeholders in the tender document template with their corresponding actual business content. During this full multimodal tag entity replacement and cross-document technical fragment embedding process, the distribution and replacement logic is executed according to the cross-modal attributes of the tags: for qualification document tags, the corresponding scanned documents or electronic fragments from the source tag library or demand tag library are automatically retrieved and inserted into the document paragraphs in place as cross-modal images or structured objects; for technical solution tags, the corresponding standard technical document content or compliant long text paragraphs generated by the controlled large model are automatically extracted, and seamless embedding and merging of cross-document technical fragments are performed. After all tags in the document have been replaced and rendered with high precision, the document structure is automatically saved and a compliant target tender document is output, thus completely closing the loop and completing the entire automated collaborative writing process for the key core content of the tender document.

[0050] Example 1 refer to Figure 2 , Figure 2 A flowchart illustrating the steps of a lightweight tender document intelligent writing method based on structured tag mapping according to Embodiment 1 of this application is shown. As shown in the figure, the writing method includes: Step 201: Construction of Enterprise Information and Bidding Information Tags. First, the core asset information of engineering enterprises is categorized and sorted from multiple dimensions to construct a standardized source tag library. The source tag library is specifically divided into 5 core tag groups: The first group is the basic information tag group, which includes, but is not limited to, {{enterprise name}} (e.g., XX Construction Engineering Co., Ltd.), {{unified social credit code}} (e.g., 91XXXXXXXXXXXXXXX), and {{registered address}} (e.g., XX Province, XX City, XX District, XX Road, XX Number); the second group is the qualification document tag group, which includes, but is not limited to, {{enterprise business license}} (originally associated with the corresponding scanned copy of the business license) and {{Grade A General Contracting Qualification for Building Construction}} (originally associated with the corresponding qualification certificate). The first group is a book scan); the second group is a technical solution tag group, which includes, but is not limited to, {{construction engineering technical solutions}} (corresponding construction technical documents in situ) and {{green construction implementation plan}} (corresponding green construction documents in situ); the third group is a performance tag group, which includes, but is not limited to, {{2023 XX municipal engineering bidding performance}} (corresponding winning bid notice scans and project introductions in situ); the fifth group is a personnel tag group, which includes, but is not limited to, {{project manager qualifications}} (corresponding professional title certificates and registered construction engineer certificates scans in situ). To ensure absolute synchronization of the underlying data source, each tag in each tag group is associated with a unique asset identifier (such as QY-JC-001) and a dynamically updated timestamp at the underlying level. Dynamic incremental updates are triggered through a timed maintenance or dedicated maintenance mechanism to ensure that the content information carried by the tag entity remains highly consistent with the latest actual operation and production situation of the enterprise. While building the source tag library, the input unstructured or heterogeneous format bidding documents are obtained through the upload module. Here, a bidding document for a municipal road construction project in PDF format is used as an example. The system invokes built-in document structuring parsing or multimodal semantic parsing components to accurately extract core bidding information through algorithms and perform tagging transformation, thereby mapping and constructing a highly targeted demand-side tag library. The demand-side tag library specifically includes {{Bidding Project Name}} (e.g., XX City XX District Municipal Road Upgrade and Reconstruction Project), {{Bidding Project Number}} (e.g., ZC-2026XXXX), {{Bidding Agency}} (e.g., XX Bidding Agency Co., Ltd.), {{Purchaser}} (e.g., XX City XX District Urban Management Bureau), {{Bidding Requirements}} (e.g., possessing a Class A or higher qualification for general contracting of building construction, project duration 180 days), and {{Evaluation Criteria}} (e.g., technical solution 40 points, performance 30 points, price quotation 30 points).

[0051] Step 202, Identification and filling of the designated areas in the tender documents. Combined with... Figure 3 , Figure 3 This document illustrates a flowchart of the structured parsing and intelligent filling process for the fill-in area of ​​a tender document according to a specific embodiment of this application. As shown in the figure, the tender document fill-in area identification and filling process includes: Step 1: Upload the tender documents and perform pre-paragraph filtering and noise reduction.

[0052] The process retrieves the uploaded Word template document of the tender document, traverses all paragraphs, and reads each paragraph sequentially. It uses an empty paragraph skip operator to directly remove and filter all invalid empty paragraphs or lines, thereby minimizing the amount of data to be processed and reducing subsequent computational resource consumption. Specific checks include: whether paragraphs contain blank segments, segments with underlined characters, or segments with an underline style; whether paragraphs end with a colon (both full-width and half-width colons are included in the judgment criteria); whether paragraphs contain preset guiding keywords such as "to be filled in by the bidder," "fill in," or "full name"; whether paragraphs contain "year," "month," and "day" time identifiers in chronological order, with only spaces or underscores between the identifiers; and whether paragraphs simultaneously contain cross-modal association feature words such as "business license" and "copy / attached / attached." Specifically, cross-modal association feature words are mapped and bound to subsequent image assets through preset heterogeneous text pointers, indicating that the current text paragraph possesses the physical routing attribute to trigger the insertion of subsequent cross-modal images. For paragraphs that are not to be filled but do not have the above characteristics, skip them without further processing and use them as contextual references for subsequent paragraphs to be filled.

[0053] Step 2: Perform slice preprocessing on the retained segments.

[0054] For valid paragraphs obtained through feature screening, the text slicing preprocessing operator "execute preprocessing" is initiated. Text slicing preprocessing is a general standardized step that is called both during subsequent data import into the pairing library and at runtime. It is responsible for organizing and reconstructing the slice structures in Word documents that are caused by non-standard typesetting into a unified underlying representation format, so as to ensure that exactly the same hash feature segmentation results are obtained in two processing runs. The preprocessing includes three stages: 1. Identify and merge the two characters "day" and "date" that are split into different slices due to typesetting (for example, "day space date" is merged into "date"), and clear invalid blank slices remaining from typesetting; 2. Underline standardization stage: For the frequently occurring scenarios in documents where consecutive underscore characters "___" or full-width underscore characters "__" are used to simulate fill-in-the-blank positions, these non-standard characters are automatically and uniformly replaced with standard spaces, and the corresponding slices are set with the native underline style attribute of Word, so that the pseudo-underlines are consistent with the native real underline styles, ensuring that no branch judgment distinction is required for subsequent extraction processing; 3. Adjacent merging stage: Merge adjacent blank slices and underline slices in the memory, so as to reduce the number of redundant slices. Throughout the entire preprocessing workflow, only the text content and style features of the slices are modified, without destroying or altering the original topological structure of the paragraphs.

[0055] Step three: Calculate the hash and perform static hit interception of the pairing library.

[0056] After the text slicing preprocessing is completed, the preprocessed paragraph text content is used to call a preset cryptographic hash function to dynamically calculate a unique SHA-256 hash value, which ensures that the slice structure is consistent with that at the time when the pairing library data is put into storage, preventing the same paragraph from generating different hash values that would result in cache miss. Next, the program determines "whether the hash value query hits the pairing library": if it hits, it indicates that there is a historical writing record for this paragraph, the interception mechanism is triggered, the "label mapping relationship" is directly read to fill the paragraph, thereby skipping the subsequent AI recognition call, which greatly saves the Token window overhead of the large model in engineering and makes the model's attention more concentrated; if it does not hit, the current paragraph is determined to be a brand new heterogeneous paragraph to be filled, and the missed paragraph is routed to the subsequent AI inference link.

[0057] In some specific embodiments, the pairing library is a constructed database cache table used to store cached records of the mapping relationship between the SHA-256 hash values ​​of paragraph normalized text and business tags. Each record in the cache contains: hash value, normalized text, original text for display, tag mapping structured array, and number of fill areas; wherein, the tag mapping structured array supports two positioning types: region filling positioning type and location positioning type. Specifically, the construction process of the pairing library is as follows: upload a docx template document of known format, extract paragraphs with fill areas, perform segmentation preprocessing on the paragraphs, normalize the entire text of the paragraphs and calculate the SHA-256 hash value; then, receive configuration instructions for assigning corresponding business tags to each fill area through the management interface; finally, save the records containing hash values, normalized text, original text for display, tag mapping structured array, and number of fill areas to the database. During system runtime, a normalized hash value is calculated for each paragraph to be filled. If a record with the same hash value is found in the matching library, the cached tag mapping structured array is directly read and in-situ filling is performed, skipping subsequent AI recognition calls. The SHA-256 hash value calculation can employ conventional one-way hash digest techniques in the computer science field. Alternatively, the normalized text string can be converted into a fixed-length 256-bit binary data stream or a corresponding 64-bit hexadecimal string by calling the secure hash algorithm interface in a standard cryptographic library (such as OpenSSL, Java Cryptography Extensions JCE, or the native Hashlib module).

[0058] Step four: Build the tagged text and perform two rounds of controlled AI recognition.

[0059] For segments that do not match the library, a new segment with the specified position is dynamically constructed in memory based on the preprocessed slice location. <rn>The text of each paragraph is marked with a unique slice index, and the plain text of the preceding adjacent paragraph is appended to the beginning of each paragraph as an enhanced "contextual reference." After all paragraphs are constructed, the AI ​​recognition process is initiated. To ensure that the output of the large model fully conforms to the standard JSON format, a structured output constraint mechanism is enabled simultaneously when calling the API. The AI ​​recognition process consists of two controlled calls: the first round will include... <rn>The structured data of paragraphs with dedicated slice index tags is input into the large language model, which is then instructed to perform a binary classification task of fill-in-the-blank existence. This allows the model to determine which paragraphs do indeed need to be tagged, and to filter out paragraphs that do not need to be tagged through logic on the front end of the interface. Subsequently, the paragraph data that has been filtered and purified in the first round is input into the second round of AI recognition, which controls the large model to focus on performing accurate multi-dimensional semantic matching, matching the corresponding target business tag name for each fill-in area in each paragraph.

[0060] Specifically, when a matching library is not found, the structured serialized text data dynamically assembled in memory for the heterogeneous paragraphs to be identified by AI is shown in the following example data structure format: [Above: Supplier Name of Consultation Response:] P3 <r0> date:< / r0> <r1>< / r1> P4 <r0> Tender No.:< / r0> <r1>< / r1> <r2>< / r2> [Previous text: Project number:] P6 <r0> According to your opinion< / r0> <r1>< / r1> <r2> Tender Announcement for Project Procurement Services< / r2> <r3>< / r3> <r4> (Tender No.), Signatory Representative< / r4> <r5>< / r5> <r6> (Full Name, Position) is duly authorized and represents the supplier.< / r6> <r7>< / r7> <r8> (Supplier name, address) Please submit the following documents:< / r8> In the example data structure above, the bracketed [Previous text: ...] field represents the plain text features of the previous valid adjacent paragraph, dynamically retrieved, and is used as a contextual reference during inference to eliminate spatial semantic drift in the large model; the P3, P4, and P6 string fields represent the absolute paragraph number of the currently processed paragraph in the original tender document template document; and the format like <r0>to <r8>Furthermore, paired XML-tagged fields are mapped in situ and represented as dedicated slice index tags arranged in physical order within the paragraph after preprocessing and noise reduction. After the above-mentioned serialized structured text data of all paragraphs to be processed is constructed, the data is fully submitted to the large language model interface to start the AI ​​recognition process. To ensure that the output of the large model can fully converge and conform to the standard JSON format, a structured output constraint mechanism is enabled simultaneously when calling the interface. The AI ​​recognition is specifically divided into two rounds of controlled calls: the first round will use the above-mentioned XML-tagged fields to generate the structured output data. <rn>The structured data of paragraphs with dedicated slice index tags is input into the large language model. The large model is instructed to perform only a binary classification task based on the existence of fill-in-the-blanks, allowing it to determine which paragraphs truly need labels. Invalid paragraphs that do not need labels are then filtered out directly on the front end via the interface. Subsequently, the simplified paragraph data after the first round of filtering and purification is input into the second round of AI recognition, controlling the large model to focus on performing a high-precision multi-dimensional semantic matching task for each fill-in area. <rn>The sequence number slice accurately matches the corresponding target business tag name, and finally outputs the candidate matching results in the form of a standard strongly typed JSON array.

[0061] Step 5: Perform boundary verification and reverse index compensation filling post-processing.

[0062] After the large language model returns a JSON array of candidate padding values, to ensure absolute stability of tag padding and global layout formatting, multiple boundary range and whitelist checks are automatically performed for filtering: First, the paragraph numbers (P0 / P1...) in the returned results are checked to see if they are within the range of valid paragraph numbers currently being processed; then, the slice indexes in the returned results (i.e.,...) are checked... <rn>The system checks whether the sequence number is within the cumulative number of segments of the current valid paragraph; then it detects and removes invalid phantom data with empty tag names, and compares and verifies whether the target tag is within the preset business tag whitelist. For any abnormal interference data that does not conform to the above validity range rules, it directly discards the data and outputs it to the exception log recorder for filling and tracing. Finally, for the matching list that has passed the boundary validity check, it is merged and grouped by paragraph, and the paragraph-level reverse index compensation sorting mechanism is activated. For multimodal tag isolation operations that may change the paragraph structure (such as starting a new line for a business license), it is strictly executed in reverse order from the end of the document; for segment-level text replacement within a paragraph, it is executed in forward order, thereby completely eliminating the risk of index offset disorder of the preceding document position caused by text replacement and length changes. In the replacement logic, for blank or underlined slices, the corresponding tag names from the source tag library or demand tag library are used to perform in-place text placeholders while perfectly preserving the original layout style and format; for paragraphs ending with a colon, the corresponding tag name is directly appended after the original text; and for slices with existing tag placeholders, the replacement is skipped to prevent duplicate replacement; for specific multimodal image or document tags such as business licenses, cross-modal insertion is supported by automatically starting a new line below the current paragraph, thereby safely and efficiently generating standardized tender document templates with accurate pre-filled tags.

[0063] Step 203: Manually review and correct the inserted tags. Output the tender document with the pre-filled tag text placeholders and submit it to the manual interface for multi-dimensional verification. During this review, the manual review focuses on the accuracy, completeness, and compliance of the tag filling to confirm the semantic match between the current tag and the original area to be filled, and to examine whether the enterprise asset information or core project requirement information carried by the tag fully meets the mandatory requirements of the tender document. If the manual review finds filling errors, matching deviations, or missing area signals, the manual reviewer will initiate on-site manual processing of the matching deviations through the interactive interface. For example, during manual review, it was discovered that the "Construction Technical Solution: []" area in the tender document template was intelligently filled with the tag "{{Construction Technical Solution for Building Engineering}}". However, based on expert experience, it was determined that the existing technical document corresponding to this tag belonged to conventional building engineering and did not fully adapt to and respond to the specific bidding requirements of this municipal road upgrade and renovation project, thus constituting a matching deviation. In this case, the manual staff directly performed the in-situ replacement of the tag through the interactive interface, replacing the tag with the more environmentally adaptable tag "{{Green Construction Implementation Plan}}", and corrected the content of the technical document associated with the tag in situ, ensuring that the supplemented solution fully complies with all the mandatory requirements for green construction in this tender.

[0064] Step 204, Supplementing Unmatched Area Tags. The adaptive supplementation and incremental accumulation of extended tags focus on blank or unmatched areas that have not yet been intelligently filled with tags after review. When manually scanning the full text through the review interface, it is discovered that special areas such as "After-sales service commitment: ______" in the tender document do not match any tags in the existing tag library. First, responding to the manually input search command, the human-computer interaction layer determines whether the current unmatched area can reuse existing tags from the source tag library or the demand tag library. If the human determines that the existing tags cannot meet the specific filling requirements of this tender document due to a lack of customized features, responding to the human-computer interaction in-situ derivation command, the extended tag derivation mechanism is activated. Dynamic AI-generated tags are generated as extended tags in the current area. These dynamic AI-generated tags are encapsulated using a unified feature expression format that includes a specific generation domain prefix. Specifically, this manifests as in-situ derivation and generation of a new unified format tag {{AI:After-sales service commitment}}. During this adaptive supplementation process, based on the system's field definition prompts, the operator performs preliminary semantic dimension annotation on the newly added AI-generated tag. The content requirements for the unstructured content to be filled in corresponding to this extended tag are explicitly defined in the background. For example, its semantic description is annotated as "In conjunction with this municipal road project, write an after-sales service commitment that meets the bidding requirements, including warranty period, repair response time, personnel configuration, etc." Subsequently, the system's data routing component responds to the submission command, accumulating the dynamically generated AI tag with the new semantic annotation as incremental data and updating it to the tag library, completing the adaptive expansion and structural closure of the tag ecosystem.

[0065] Step 205, AI-generated tag content editing. The compliant text content binding process involves customizing the prompt word engineering configuration, content model rendering, and final content review for the newly derived and accumulated dynamic tags of the {{AI:After-sales Service Commitment}} type in the tag library. Manually, through the configuration panel, a dedicated AI-generated prompt word engineering is written and configured for this dynamic AI-generated tag. The prompt words not only include constraints on the input data structure but also clearly limit and standardize the target technical content requirements, formatting specifications, industry standards, and key points of deep adaptation with the current bidding documents. The specific prompt word engineering logic written and locked here is: "In conjunction with the XX City XX District Municipal Road Upgrade and Reconstruction Project, write an after-sales service commitment, requiring a warranty period of no less than 2 years, a repair response time of no more than 2 hours, a professional repair team, clear repair procedures and division of responsibilities, conforming to the industry standards for after-sales service in municipal engineering, with standardized language and specific content, avoiding empty rhetoric." Next, the configured prompt words are input into the large language model interface. The large model automatically generates highly compliant target tag content that meets the constraints of the above indicators based on the defined bidding requirements context. After the large language model completes the automatic text generation, the generated unstructured text content is presented and rendered to the interactive interface again. The generated content undergoes a final compliance review and correction of textual detail deviations by a human. Once the human confirms that the content fully complies with the bidding document requirements and is error-free, the final binding command is triggered. The underlying relational database responds to the binding command, binding the verified compliant after-sales service commitment content with the corresponding {{AI:After-sales Service Commitment}} tag at the entity level primary and foreign key relationship, thereby obtaining a full tag mapping routing table containing all compliant text bindings.

[0066] Step 206, Tag Replacement and Tender Document Generation. The final target tender document generation is an automated multimodal entity replacement and fragment merging rendering process based on the aforementioned full tag mapping routing table. Using a high-performance document rendering program, all pre-set and supplemented tags and their corresponding actual business content are automatically extracted from the full tag mapping routing table, and all tag placeholders in the tender document template are replaced in their original positions with their corresponding actual business content. During this full-scale multimodal tag entity replacement and fragment embedding process, the program automatically executes the split-replacement logic based on the cross-modal attributes of the tags: For plain text or business field tags, the program automatically replaces {{Bidding Project Name}} in place with the string "XX City XX District Municipal Road Upgrading and Reconstruction Project", and replaces {{Company Name}} in place with the string "XX Construction Engineering Co., Ltd."; For qualification document type multimodal tags, the program automatically retrieves the corresponding scanned copy or electronic fragment of the enterprise document from the enterprise's standard asset library, specifically replacing the {{General Contracting Level 1 Qualification for Building Construction}} tag in place with the corresponding scanned image of the qualification certificate and securely inserting it into the document; For customized derived tags, the program extracts the bound AI-generated text and replaces the {{AI:After-sales Service Commitment}} tag in place with the aforementioned commitment content that has been manually reviewed and corrected. Once all tags in the document have been replaced and rendered with high precision, the program activates the formatting operator to automatically unify the global file format, standardizing the entire document to a preset standard layout (e.g., Song typeface, size 4, line spacing 3 mm, top left margin 25 mm, bottom right margin 15 mm), and automatically numbering pages. Simultaneously, a detailed tag replacement log is generated at the underlying level for compliance traceability. Finally, the system automatically performs structured file saving and outputs a complete tender document, securely and efficiently closing the loop and completing the entire automated collaborative writing process for the key core content of the tender document.

[0067] Furthermore, in order to verify the technological advancement and resource-saving advantages of the lightweight tender document intelligent writing method based on structured tag mapping proposed in this application in practical engineering applications, a multi-algorithm comparative performance test was performed on the same tender document skeleton (such as the credit information section .docx) containing complex heterogeneous text. The test results are detailed in Table 1.

[0068] Table 1. Comparison of performance metrics for different extraction schemes on the same test file.

[0069] Under the same benchmark test environment, the pure text extraction scheme [0], the scheme without optimization of the method of this application [1], the scheme of this application [2], and the general large model direct extraction scheme [3] were used to conduct a multi-index quantitative evaluation of the identification and filling of the area to be filled. As shown in Table 1, although the pure text extraction scheme [0] consumes less tokens, its accurate insertion number is directly zero, which is completely unable to cope with the complex tender document layout. Although the scheme without optimization of the method of this application [1] identifies a large number of areas to be filled, its single run consumes a word throughput of up to 428976 tokens, and the running time is extended to 168.6s, resulting in extremely high and uncontrollable computing power consumption and communication delay. The general large model direct extraction scheme [3] lacks the pre-lightweight reconstruction and standardized JSON constraints, and its intermediate reasoning process produces serious illusions, resulting in a sharp reduction in the number of matching tags to 17, and the recognition completeness is seriously insufficient. In contrast, when using the scheme of this application [2], based on the accurate identification of 146 areas to be filled, the absolute convergence alignment of the number of matching tags and the number of precise insertions was achieved (the number of precise insertions reached 64 times), proving the strong interception effect of the pre-boundary legality verification filtering network on the illusion of large models. More significantly, under the strong constraint of the single-round multi-task matching mechanism, the token consumption of this application scheme in a single run was greatly reduced to 30863 tokens. Compared with the scheme [1] which was not optimized by the method of this application, its token consumption dropped by more than 90%, and the overall running time was significantly compressed to 107.1s. This proves that while ensuring the high-precision in-situ filling of the tender and the stability of the page inheritance, the method of this application truly achieves the core invention effect of "lightweight" with low computing resource consumption, low computing power consumption, and high response concurrency, building a very high technical barrier for competitors.

[0070] Continue to refer to Figure 4 As an implementation of the above method, in a second aspect, this application provides an embodiment of a framework diagram 400 of a lightweight intelligent tender document writing system based on structured tag mapping. This system embodiment is similar to... Figure 1 Corresponding to the illustrated method embodiment, this system can be specifically applied to various electronic devices. The system 400 includes a dual-end tag mapping construction module 401, a region-to-be-filled identification and filling module 402, an interactive verification tag deposition module 403, and a multimodal replacement tender generation module 404, all interconnected. The dual-end tag mapping construction module 401 is configured to perform structured parsing and standardized tag processing based on various core resource information of enterprises and core demand information of bidding projects, and to map and construct source-end tag library and demand-end tag library respectively. The unfilled area identification and filling module 402 is configured to obtain the tender document template document, perform paragraph filtering and reconstruction on the tender document template document to identify the unfilled areas in the tender document template document and retain the legal paragraphs; calculate the text hash value of the legal paragraphs to determine whether they hit the preset cache, and output the candidate matching results based on the hit results or the results of tag matching inference for legal paragraphs that do not hit the preset cache; based on the candidate matching results, combined with the source tag library and the demand tag library, generate corresponding pre-filled placeholders in the unfilled areas to obtain the preliminary pre-filled tender document; The interactive verification tag accumulation module 403 is configured to present the preliminary pre-filled tender document to the human interactive interface to perform multi-dimensional verification, respond to the human in-situ processing instructions to perform adaptive supplementation and incremental accumulation of extended tags for matching deviations, and perform compliant text content binding on extended tags through prompt word engineering configuration to obtain a full tag mapping routing table containing full compliant text bindings. The multimodal replacement tender document generation module 404 is configured to retrieve the corresponding actual business content from the source tag library or the demand tag library based on the full tag mapping routing table, perform full multimodal tag entity replacement and cross-document technical fragment embedding and merging, and output a compliant target tender document.

[0071] Although the principles of the present invention have been described in detail above with reference to preferred embodiments, those skilled in the art should understand that the above embodiments are merely illustrative explanations of the implementation of the present invention and are not intended to limit the scope of the present invention. The details in the embodiments do not constitute a limitation on the scope of the present invention. Any obvious changes, such as equivalent transformations or simple substitutions, based on the technical solutions of the present invention without departing from the spirit and scope of the present invention fall within the protection scope of the present invention.< / rn> < / rn> < / rn> < / rn> < / rn> < / rn> < / rn> < / rn> < / rn> < / rn> < / rn> < / rn>

Claims

1. A lightweight intelligent tender document writing method based on structured tag mapping, characterized in that, The method includes: S1. Based on the core resource information of the enterprise and the core demand information of the bidding project, perform structured parsing and standardized tag processing, and map and construct the source tag library and the demand tag library respectively; S2. Obtain the tender document template document, perform paragraph filtering and reconstruction on the tender document template document to identify the areas to be filled in the tender document template document and retain the legal paragraphs; calculate the text hash value of the legal paragraphs to determine whether they hit the preset cache, and output the candidate matching result based on the hit result or the result of tag matching reasoning for the legal paragraphs that do not hit the preset cache; based on the candidate matching result, combined with the source tag library and the demand tag library, generate corresponding pre-filled placeholders in the areas to be filled to obtain the preliminary pre-filled tender document; S3. The preliminary pre-filled tender document is submitted to the human interaction interface to perform multi-dimensional verification. In response to the human in-situ processing instruction, adaptive supplementation and incremental accumulation of extended tags are performed for matching deviations. The extended tags are bound to compliant text content through prompt word engineering configuration to obtain a full tag mapping routing table containing full compliant text bindings. S4. Based on the full tag mapping routing table, retrieve the corresponding actual business content from the source tag library or the demand tag library, perform full multimodal tag entity replacement and cross-document technical fragment embedding and merging, and output a compliant target tender document; Specifically, step S3, which involves responding to the manual in-situ processing instruction and adaptively supplementing the extended labels to address the matching deviation, includes: The preliminary pre-filled bid document is output and submitted to the human interaction interface, where a human performs multi-dimensional verification and checks on the preliminary pre-filled bid document. During the manual review process, the accuracy, completeness and compliance of the label filling are manually checked, the semantic matching degree between the label and the original area to be filled is confirmed, and the enterprise asset information or project core requirement information carried by the label is reviewed to see if it meets the requirements of the tender documents. If manual verification reveals filling errors, matching deviations, or missing regions, manual in-situ processing commands can be initiated through the interactive interface to manually replace labels, correct deviations, or adaptively supplement missing labels to correct the matching decision boundary at the interactive level. The full multimodal tag entity replacement and cross-document technology fragment embedding and merging in step S4 includes the following steps: Obtain all pre-set and supplemented tags and their corresponding actual business content from the full tag mapping routing table, and automatically replace all tag placeholders in the tender document template document in place through document rendering; Once all tags have been replaced and rendered, the file structure is saved and the complete technical compliance target tender document is output. During the replacement process, a split replacement is performed based on the cross-modal attributes of the tag: For qualification document tags, automatically retrieve the corresponding scanned copy or electronic file fragment of the enterprise document and insert it in situ into the target document paragraph in the form of cross-modal image or structured object; For technical solution tags, the system automatically extracts the corresponding standard technical document content or compliant long text paragraphs generated by the large language model, and performs seamless embedding and merging of cross-document technical fragments.

2. The lightweight intelligent tender document writing method based on structured tag mapping according to claim 1, characterized in that, The core resource information of the enterprise mentioned in step S1 includes basic enterprise information, qualification documents, technical solutions, performance certificates, and personnel information; the source tag library is divided into basic information tag group, qualification document tag group, technical solution tag group, performance tag group, and personnel tag group according to information type, and each tag is associated with a unique identifier and update timestamp to achieve dynamic incremental updates; the core requirement information of the bidding project includes the bidding project name, bidding project number, bidding agency, purchaser, bidding requirements, evaluation criteria, and bid deadline.

3. The lightweight intelligent tender document writing method based on structured tag mapping according to claim 1, characterized in that, In step S2, paragraph filtering and reconstruction of the tender document template document are performed, including: traversing all paragraphs of the tender document template document, removing and filtering all invalid empty paragraphs or blank lines in the document; and checking whether the remaining paragraphs have the characteristics of the fill area or the characteristics of the business license, skipping non-fill area paragraphs that do not have the characteristics of the fill area and the characteristics of the business license, and using them as context references for subsequent paragraphs to be filled. The judgment rules used when checking whether the retained paragraphs have the characteristics of an infill area or a business license include: Determine if a paragraph contains a blank slice, a slice with underlined characters, or a slice with an underline style. Determine whether a paragraph ends with a full-width or half-width colon; Determine whether a paragraph contains preset guiding keywords; Determine whether the paragraph contains time markers in chronological order, and whether the time markers are separated only by spaces or underscores; And determine whether a paragraph contains cross-modal related feature words.

4. The lightweight intelligent tender document writing method based on structured tag mapping according to claim 1, characterized in that, The S2 step, which determines whether a preset cache has been hit and performs tag matching inference, includes: Step 1: Perform text slicing preprocessing on the legal paragraphs, and calculate a unique SHA-256 hash value based on the preprocessed paragraph text content; Step 2: Use the SHA-256 hash value as the query key to retrieve the matching library that is the preset cache; Step 3: If the determination result is a match, the tag mapping relationship recorded in the matching library is directly read to generate the pre-filled tag placeholder result for the corresponding paragraph and the interface call is skipped; if the determination result is a mismatch, the text data of the corresponding paragraph is submitted to the large language model for tag matching inference to output the large model inference matching result. Step 4: Summarize the pre-filled label placeholder results and the large model inference matching results to output the candidate matching results in a unified manner; The pairing library is a pre-built database cache table that stores cache records. Each record in the cache records contains at least: SHA-256 hash value, normalized text, original text for display, label mapping structured array, and number of filling areas. The label mapping structured array supports two positioning types: region filling positioning type and location positioning type. The steps for constructing the pairing library include: uploading a docx template document in a known format and extracting paragraphs with fill areas; performing segmentation preprocessing on the paragraphs; normalizing the entire text of the paragraphs and calculating the SHA-256 hash value; receiving configuration instructions through the management interface to assign corresponding business tags to each fill area to construct the tag mapping structured array; and finally saving the record containing the SHA-256 hash value, the normalized text, the original text for display, the tag mapping structured array, and the number of fill areas to the pairing library.

5. The lightweight intelligent tender document writing method based on structured tag mapping according to claim 4, characterized in that, The large language model performs label matching inference, which includes the following two rounds of controlled recognition steps: Based on the slice positions after the text slice preprocessing, a dynamic structure with [text slice positions] is constructed in memory. <rn> The paragraph text is marked with a dedicated slice index, and the plain text content of the adjacent paragraph above it is appended in place before each paragraph as an enhanced contextual reference.< / rn> After all unprocessed paragraphs are constructed, the data is fully delivered, and a structured output constraint mechanism is simultaneously enabled when calling the interface to ensure that the model output conforms to the JSON format. Identification is performed through two rounds of controlled calls: the first round controls the large language model based on the input... <rn>The paragraph text with dedicated slice index tags performs a binary classification task based on the existence of fill-in-the-blank entries. Invalid paragraphs that do not require labeling are filtered out using logical filtering. Then, the filtered, simplified paragraph data is input into the second round of calls, controlling the large language model to perform a multi-dimensional semantic matching task, matching each fill-in-the-blank area accordingly. <rn> The sequence number slice is matched to the corresponding target business tag name, and the candidate matching results are output uniformly.< / rn> < / rn> 6. The lightweight intelligent tender document writing method based on structured tag mapping according to claim 4, characterized in that, Before generating corresponding pre-filled placeholders in the area to be filled based on the candidate matching results, the method further includes performing boundary validity verification on the candidate matching results, the boundary validity verification including the following filtering steps: Paragraph boundary filtering step: Extract the target paragraph number from the candidate matching results and remove out-of-bounds candidate data that exceeds the range of the current legal paragraph numbers; Slice boundary filtering step: Extract the slice index number from the candidate matching results and remove out-of-bounds slice data that exceed the total number of corresponding legal segment slices; Tag compliance filtering steps: Detect the tag names in the candidate matching results, remove invalid phantom data with empty tag names in situ, and compare with the preset business tag whitelist to filter and remove illegal tags; Any abnormal interference data that triggers any of the filtering steps described above will be discarded and simultaneously output to the abnormal log recorder for evidence storage.

7. The lightweight intelligent tender document writing method based on structured tag mapping according to claim 1, characterized in that, The replacement logic for generating the corresponding pre-filled placeholder in the unfilled area in step S2 includes the following steps: The collaborative replacement control steps involve initiating a paragraph-level reverse index compensation sorting mechanism for the matching list after all entries have passed the boundary validity check. For multimodal tag isolation operations that change the paragraph structure, the operations are executed in reverse order from the end of the document. For slice-level text replacement within a paragraph, the operations are executed in forward order to eliminate the risk of index offset errors in the preceding document due to text replacement or length changes. The in-situ text placeholder step for slices involves using the tag names in the corresponding source tag library or demand tag library to perform in-situ text placeholders while preserving the original layout style and format for blank or underlined slices. The boundary appending and deduplication steps are as follows: for paragraphs ending with a colon, the corresponding tag name is directly appended after the original text; while for slices with existing tag placeholders, the process is skipped to prevent duplicate replacement. The cross-modal content isolation step involves automatically inserting cross-modal tags on a new line below the current paragraph for multimodal images or document tags, generating the preliminary pre-filled tender document.

8. The lightweight intelligent tender document writing method based on structured tag mapping according to claim 1, characterized in that, The adaptive supplementation and incremental deposition of extended tags in step S3 also includes in-situ derivation and adaptive expansion steps for extended tags in unmatched regions: In response to human-input interactive commands, at the human-computer interaction level, it is determined by humans whether the current unmatched area can reuse the existing tags in the source tag library or the demand tag library; If the existing tags are determined by the manual process to be unsuitable for the requirements of this tender, dynamic AI-generated tags will be generated in response to the manual in-situ derivation instruction. The dynamic AI-generated tags will be encapsulated using a unified feature expression format that includes a generation domain prefix identifier. Based on the interactive prompts, the human staff performs preliminary semantic dimension annotation on the newly added AI-generated tags, clarifies the content requirements of the unstructured content to be filled corresponding to the extended tags, and saves the dynamic AI-generated tags as incremental data and updates them to the corresponding source tag library or demand tag library.

Images