Intelligent analysis and recommendation method for bidding documents, computer device and readable storage medium

Abstract

This application discloses a method, computer device, and readable storage medium for intelligent analysis and recommendation of tender documents, belonging to the field of artificial intelligence and natural language processing technology. The method includes the following steps: constructing a knowledge graph based on acquired target content related to tendering and bidding, and calling an embedding model to generate text vectors; storing the knowledge graph in a graph database and the text vectors in a vector database; responding to a user's query request, extracting low-level keywords and high-level keywords based on the query request; wherein the query request carries at least the tender information of the tender project; performing graph vector collaborative retrieval in the graph database and the vector database based on the low-level keywords and the high-level keywords to obtain a multi-source fusion context; and generating a tender recommendation scheme based on the multi-source fusion context. This application can improve the accuracy of tendering and bidding analysis.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence and natural language processing technology, specifically relating to a method for intelligent analysis and recommendation of tender documents, computer equipment, and readable storage medium. Background Technology

[0002] In the fields of engineering construction and government procurement, bidding is a crucial step for companies to win projects. Bidding specialists need to analyze massive amounts of bidding documents, which are often hundreds of pages long, complex in structure, and contain numerous technical specifications, double-page tables, and scanned images, all with rigorous logic.

[0003] Currently, the industry has introduced retrieval enhancement generation technology based on vector databases in an attempt to improve efficiency. However, this technology typically involves simply splitting long documents into isolated text fragments for flattened retrieval. This method severs the inherent logical connections within the document. For example, it cannot effectively link "qualification requirements" scattered in different locations with their corresponding "proof formats." This leads to systems used for bidding analysis being prone to information gaps, logical breaks, and even factual "illusions" when answering complex questions, making it difficult to meet the stringent accuracy requirements of bidding analysis and limiting its practicality. Summary of the Invention

[0004] The purpose of this application is to provide a method for intelligent analysis and recommendation of tender documents, a computer device, and a readable storage medium, which can improve the accuracy of tender analysis.

[0005] To solve the above-mentioned technical problems, this application is implemented as follows: In a first aspect, embodiments of this application provide a method for intelligent analysis and recommendation of tender documents, the method comprising: Based on the acquired target content related to bidding, a knowledge graph is constructed, and an embedding model is invoked to generate text vectors; The knowledge graph is stored in a graph database, and the text vectors are stored in a vector database; In response to a user's query request, low-level keywords and high-level keywords are extracted based on the query request; wherein, the query request carries at least the bidding information of the bidding project; Based on the low-level keywords and the high-level keywords, graph vector collaborative retrieval is performed in the graph database and the vector database to obtain a multi-source fusion context; Based on the multi-source fusion context, a bidding recommendation scheme is generated.

[0006] Secondly, embodiments of this application provide a tender document intelligent analysis and recommendation device, the tender document intelligent analysis and recommendation device comprising: The first generation module is used to construct a knowledge graph based on the acquired target content related to bidding and tendering, and to call the embedding model to generate text vectors; A storage module is used to store the knowledge graph in a graph database and the text vectors in a vector database; The extraction module is used to respond to a user's query request and extract low-level keywords and high-level keywords based on the query request; wherein, the query request carries at least the bidding information of the bidding project; The retrieval module is used to perform graph-vector collaborative retrieval in the graph database and the vector database based on the low-level keywords and the high-level keywords to obtain a multi-source fusion context. The second generation module is used to generate a bidding recommendation scheme based on the multi-source fusion context.

[0007] Thirdly, embodiments of this application provide a computer device including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the method described in the first aspect.

[0008] Fourthly, embodiments of this application provide a computer-readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first aspect.

[0009] Fifthly, embodiments of this application also provide a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in the first aspect.

[0010] In this embodiment, a knowledge graph is constructed based on the acquired target content related to bidding, and an embedding model is invoked to generate text vectors. The knowledge graph is stored in a graph database, and the text vectors are stored in a vector database. In response to a user's query request, low-level keywords and high-level keywords are extracted based on the query request. The query request carries at least the bidding information of the bidding project. Based on the low-level keywords and the high-level keywords, graph vector collaborative retrieval is performed in the graph database and the vector database to obtain a multi-source fusion context. Based on the multi-source fusion context, a bidding recommendation scheme is generated. This addresses the "illusion" and logical discontinuity problems in long document reasoning by employing graph-enhanced text indexing and a two-layer retrieval paradigm. Compared to simple vector retrieval, it utilizes the graph structure to capture entity relationships across paragraphs, enabling the retrieval of specific details while grasping the overall bidding intent, significantly reducing model illusions and thus improving the accuracy of bidding analysis. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating the intelligent tender analysis and recommendation method provided in some embodiments of this application; Figure 2 This is a schematic diagram of the overall process of the intelligent tender analysis and recommendation method provided in some embodiments of this application; Figure 3 This is a flowchart of the multimodal text parsing and knowledge construction module of the tender document intelligent analysis and recommendation method provided in some embodiments of this application; Figure 4 This is a schematic diagram of the two-layer retrieval and decision reasoning logic of the intelligent tender analysis and recommendation method provided in some embodiments of this application; Figure 5 This is a structural block diagram of a tender document intelligent analysis and recommendation device provided in some embodiments of this application; Figure 6 These are internal structural diagrams of a computer device provided in some embodiments of this application. Detailed Implementation

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0014] The intelligent tender document analysis and recommendation method provided in this application will be described in detail below with reference to the accompanying drawings, through specific embodiments and application scenarios.

[0015] Reference Figure 1 The method includes steps 102-110. Wherein: Step 102: Based on the obtained target content related to bidding, construct a knowledge graph and call the embedding model to generate text vectors.

[0016] In some embodiments, the target content is multimodal content, such as text, images, and tables; it may be stored in, but is not limited to, tender documents and technical documents in the field, and its storage format may include, but is not limited to, PDF, Word, and Excel. This target content can be read by an automated script tool running on a local server.

[0017] In some embodiments, constructing a knowledge graph based on the acquired target content related to bidding includes: Based on the images and tables obtained from the target content related to bidding, visual text information is generated by calling the Vision-Language Models (VLM).

[0018] Based on context-preserving technology, the visual text information is backfilled into the corresponding coordinate position in the text stream of the target content to obtain text data that is aligned in position and semantics.

[0019] The Large Language Model (LLM) is invoked to extract triples from the text data; wherein the triples include a head entity, a relation, and a tail entity.

[0020] A knowledge graph is constructed based on the triples.

[0021] VLM can be deployed locally to perform deep feature extraction and content parsing of target content, that is, to perform offline visual analysis on images and tables in the target content, thereby extracting text information and structural features.

[0022] Among them, the context position preservation technology is used to accurately backfill the parsed visual text information into the text stream corresponding to the original document coordinates, keeping its logical connection with the context unchanged. This transforms multimodal heterogeneous documents into plain text data that is strictly aligned in both position and semantics, providing a standardized input stream for subsequent knowledge graph construction.

[0023] It's important to note that traditional techniques lack the ability to analyze multimodal heterogeneous data. This is because target content is typically a heterogeneous document combining PDF, Word, and Excel formats, containing numerous complex tables carrying key parameters and scanned images. Traditional techniques mostly rely on Optical Character Recognition (OCR) to extract plain text, often losing information about the row and column structure of tables and the spatial location information in images. This results in the inability to extract useful information when analyzing equipment parameter tables (a type of table) or qualification scans (a type of image) using LLM, significantly reducing the recall rate due to structural errors. Specifically: Traditional techniques use OCR to extract plain text from images or tables, resulting in the loss of the original layout structure and implicit semantic content. For example, in a device parameter table, OCR might read the first row's title and the third row's values ​​together, completely disrupting the table's row and column relationships and turning it into a jumble of illogical characters. Faced with this fragmented and disordered information, LLM (Layered Analytical Processing) cannot accurately identify the original key parameters from the misplaced characters, meaning it cannot extract useful information. Furthermore, "reduced recall" means that although the correct answer is in the document, because it's fragmented, LLM cannot find it completely, resulting in the final analysis report missing crucial content that should have been discovered.

[0024] This embodiment utilizes context-preserving technology to maintain the logical connection between visual text information and the surrounding text of corresponding images or tables. By leveraging the image-text analysis capabilities of VLM to add logical relationship descriptions between blank text, the logical relationship between images and tables within the target content is preserved. Specifically, unlike traditional OCR which only recognizes characters, VLM performs deep analysis of images and tables in the original document, extracting textual information while preserving structural features such as the row and column distribution of tables. Furthermore, for pure image content, VLM automatically analyzes the semantic relationships within the image. For example, a red curved arrow in the center of a road in an image will be given a corresponding description after VLM analysis to summarize the meaning of the image. Based on this, the combination of context-preserving technology and VLM ensures that there are no issues with word order disorder or unclear references during subsequent LLM reading, thereby improving the accuracy of extracting triples from text data.

[0025] It should be noted that since the reading of target content and the deployment of VLM can both be performed locally, the entire process does not require data to be uploaded to the cloud, thus achieving physical isolation and privacy security of the data. Specifically, addressing the pain point of highly sensitive bidding data, this embodiment adopts a localized deployment scheme from VLM parsing, Embedding vectorization to LLM inference, using local computing power to replace cloud APIs, eliminating the risk of leakage of core business secrets, and ensuring system stability in offline or intranet environments.

[0026] In some embodiments, when invoking an LLM, the text data is first sliced ​​and then input into the LLM along with the corresponding prompts. The LLM then performs deep semantic analysis on the text data and extracts entities and relationships driven by the LLM. This allows for the parsing of standardized <head entity, relation, tail entity> triples from unstructured text. Entities (head and tail entities) include: company qualifications, personnel names, specific performance names, amounts, and performance attributes (e.g., general contracting / subcontracting / specialized contracting). Relationships include: the relationship between personnel and projects (e.g., who is the project manager), the relationship between projects and attributes (e.g., whether the project is a general contracting or subcontracting), and the relationship between personnel and certificates (what level of certificate they hold).

[0027] For example, an unstructured text might read: "December 5, 2024, a tunnel project in a certain province, electromechanical engineering section 1, 123,455,432 yuan, Class I highway general contracting. Project Manager: Gao, Chief Engineer: Ding. Already registered in the Ministry of Transport system." The ternary set could be: <Tunnel project in a certain province, electromechanical engineering section 1, amount: 123,455,432 yuan>, <Tunnel project in a certain province, electromechanical engineering section 1, type: Class I highway general contracting>, <Gao, serving as project manager, tunnel project in a certain province, electromechanical engineering section 1>, <Ding, serving as technical head, tunnel project in a certain province, electromechanical engineering section 1>, <Tunnel project in a certain province, electromechanical engineering section 1, entered in the Ministry of Transport system>. This automatically constructs a domain knowledge graph reflecting the logical structure of the tender document.

[0028] The Prompt is a pre-designed set of structured prompt templates designed to address the complex logic implicit in the correspondence between "qualification requirements" and "scoring criteria" in tender documents. These Prompts contain targeted instructions designed to guide the large language model to accurately focus on the key elements of the tender document, thereby accurately extracting entities and relationships. For example, a Prompt might include: "You are a senior highway electromechanical engineering cost estimator and data analyst. Your task is to extract core entities and relationships from unstructured text or tabular data to build a knowledge base of the company's bidding capabilities." Prompts can also include other targeted instructions, such as: "Convert all amounts to 'ten thousand yuan' units."

[0029] In some embodiments, a deduplication mechanism is introduced, that is, synonymous or duplicate entities are automatically merged during the entity extraction stage, thereby compressing the size of the knowledge graph and avoiding the infinite expansion of the knowledge base.

[0030] In some embodiments, the invocation of the embedding model to generate text vectors includes: Generate summary descriptions for the head entity, the relation, and the tail entity, respectively.

[0031] The head entity, the relation, the tail entity, and the corresponding summary description are mapped to structured key-value pairs; wherein the key of the key-value pair is the head entity, the relation, or the tail entity, and the value of the key-value pair is the summary description corresponding to the key.

[0032] The key is converted into a high-dimensional vector representation by calling an embedding model (such as Sentence-Transformers) to obtain a text vector.

[0033] The summary descriptions can be generated using LLM to explain the context of entities and relationships, thereby enriching the semantic information of the knowledge graph. In some embodiments, to improve the efficiency of knowledge graph construction, summary descriptions of some core entities and core relationships among all entities and relationships can be generated.

[0034] Step 104: Store the knowledge graph in a graph database and the text vector in a vector database.

[0035] It is understandable that graph databases and vector databases are used to form the underlying index structure for graph-vector collaborative retrieval. This underlying index structure not only preserves the topological relationships of knowledge, but also has the semantic matching capability of vectors.

[0036] In some embodiments, the method further includes: In response to the acquisition of new content related to bidding, the graph database and the vector database are updated based on an incremental update mechanism.

[0037] This embodiment takes into account the timeliness of bidding information and supports incremental data updates. Specifically, for the construction and updating of the knowledge graph, it provides dual input channels: "complete bid documents" and "individual technical solutions." Uploading new bid documents or technical solutions through these channels only extracts entities from the newly added content and mounts them onto the existing knowledge graph, without needing to reconstruct the global index. This enables real-time management and maintenance of Retrieval-augmented Generation (RAG) knowledge source files.

[0038] Step 106: In response to the user's query request, extract low-level keywords and high-level keywords based on the query request; wherein, the query request carries at least the bidding information of the bidding project.

[0039] Step 108: Based on the low-level keywords and the high-level keywords, perform graph vector collaborative retrieval in the graph database and the vector database to obtain a multi-source fusion context.

[0040] The bidding information may include, but is not limited to, bidding announcements and bidding documents.

[0041] Low-level and high-level keywords can be extracted using LLM.

[0042] In some embodiments, low-level keywords are keywords related to specific details, such as specific qualification levels, monetary amounts, and place names; high-level keywords are keywords related to abstract themes, such as "intelligent transportation," "internal collaborative projects," and "high priority."

[0043] In the process of graph-vector collaborative retrieval, vector similarity retrieval (i.e., fuzzy retrieval based on semantic similarity) is performed in the vector database using low-level keywords and high-level keywords, and one-hop expansion retrieval of graph structure is performed in the graph database. This enables hybrid retrieval in both the vector and graph databases, so as to simultaneously capture the specific entity details and global semantic background in the tender document, thereby solving the problem of taking things out of context in traditional flat retrieval.

[0044] In some embodiments, the step of performing graph vector collaborative retrieval in the graph database and the vector database based on the low-level keywords and the high-level keywords to obtain a multi-source fusion context includes: Based on the low-level keywords, a first target text vector is retrieved from the first text vectors converted from the head entity or the tail entity in the vector database, and based on the high-level keywords, a second target text vector is retrieved from the second text vectors converted from the relation in the vector database.

[0045] Determine the first target head entity or the first target tail entity corresponding to the first target text vector, and determine the first target relationship corresponding to the second target text vector.

[0046] From the graph database, determine the first neighboring entity of the first target head entity or the first target tail entity, and the second neighboring entity of the first target relationship.

[0047] Based on the key-value pairs, target summary descriptions corresponding to the first target head entity or first target tail entity, the first target relationship, the first neighboring entity, and the second neighboring entity are determined respectively.

[0048] By concatenating all the target summary descriptions and combining them with the first target head entity or the first target tail entity, the first target relationship, the first neighboring entity, and the second neighboring entity, a multi-source fusion context is obtained.

[0049] In other words, the multi-source fusion context includes entities, relationships, and a concatenated summary description of the target.

[0050] Step 110: Generate a bidding recommendation scheme based on the multi-source fusion context.

[0051] In some embodiments, generating a bidding recommendation scheme based on the multi-source fusion context includes: Based on the multi-source fusion context and preset weighted prompt words, LLM is invoked to generate a bidding recommendation scheme; The weighting prompts are constructed based on hard constraints and soft scoring terms. The hard constraints and soft scoring terms are used to generate a comprehensive score for each bidding project. The hard constraints are also used to perform pre-verification on the soft scoring terms. The bidding recommendation scheme is determined based on the top N bidding projects in terms of comprehensive score, where N is a positive integer.

[0052] In some embodiments, the weighting prompts are determined based on expert experience, which defines a weighting model for "business evaluation," "technical evaluation," and "price evaluation." The LLM (Local Management Model) generates scores for these three items—"business evaluation," "technical evaluation," and "price evaluation"—through a Chain of Thought (CoT) inference, thereby completing a comprehensive evaluation of the bidding project. This embodiment refers to this scoring process as soft scoring, implemented using the soft scoring items in the weighting prompts.

[0053] In some embodiments, a pre-verification using hard constraints is performed before soft scoring. This pre-verification, or hard constraint, refers to determining whether a veto item is valid. The veto items include five dimensions: enterprise qualifications, mandatory online performance verification, financial thresholds, credit requirements, and personnel deficiencies. If any dimension meets the criteria, the project is deemed vetoed, the overall score for the relevant bidding project is 0, the bid recommendation result is marked as "not recommended," and no further soft scoring is required. If none of the dimensions meet the criteria, then subsequent soft scoring is performed.

[0054] Among them, enterprise qualifications may include, but are not limited to: Class A qualification for special contracting of highway traffic engineering (highway electromechanical engineering), Class A qualification for professional contracting of electronic and intelligent engineering, Class A qualification for professional contracting of fire protection facilities engineering, Class A qualification for general contracting of electromechanical engineering construction, a valid safety production license, and a valid business license of the enterprise.

[0055] Among them, the mandatory online performance verification may include, but is not limited to: completion time (e.g., the past 3 or 5 years), completion method (e.g., at least 1 independently completed highway electromechanical project), contract amount (e.g., ≥ 20 million yuan), project quality (e.g., qualified), and independent general contracting requirements (e.g., requiring independent general contracting, subcontracting or participating projects are considered as not meeting the requirements).

[0056] Among them, the financial thresholds may include, but are not limited to: registered capital (e.g., ≥ 100 million yuan), average operating revenue (e.g., ≥ 100 million yuan in 2022–2024), current ratio or current / quick ratio in the past three years (e.g., ≥ 1), and working capital or bank credit certificate ≥ the amount required by the tender.

[0057] Among these, credit requirements may include, but are not limited to: not being disqualified from bidding, not being ordered to suspend business or go bankrupt, not being listed in the list of seriously dishonest enterprises, and having the latest credit rating (such as C or above).

[0058] Among these, personnel deficiencies may include, but are not limited to: Project Manager: A first-class construction engineer in mechanical and electrical engineering, holding a Class B safety certificate, and having experience as a project manager for highway mechanical and electrical engineering projects; Project Manager has no ongoing project conflicts; Project Chief Engineer has relevant work experience.

[0059] In some embodiments, the comprehensive score The CoT inference is obtained through the inference chain, and the CoT inference is implemented by the following formula (1): (1) Among them, the As the veto coefficient, in the event that the hard constraint triggers a veto to pass, the... Otherwise, the above 1; the aforementioned The above and stated The scores are weighted scores from three dimensions: business evaluation, technical evaluation, and price evaluation.

[0060] In this application, the weights for the three dimensions—business evaluation, technical evaluation, and price evaluation—can be 0.4, 0.3, and 0.3, respectively, without limitation. In other words, taking a percentage system as an example, the highest possible score for business evaluation could be 40, for technical evaluation 30, and for price evaluation 30.

[0061] in, The weighted score for business evaluation is derived by weighting three sub-items: similar performance of the enterprise, key personnel of the project, and enterprise credit and reputation. Specifically, it can be achieved through the following formula (2): (2) in, To correct the coefficients, the algorithm is driven by expert-preset logic in the weighting prompts. When the LLM identifies a project as belonging to a specific strategic collaborative project, such as an internal group collaboration project, This allows the score for that dimension to be the full score, while the scores for regular items are not the full score, such as... ; For similar performance, the score can be calculated by summing the number of main contract performances and the amount of each contract over the past 5 years. The score for key personnel allocation can be obtained by weighting the qualifications and performance achievement rates of project managers, chief engineers, and other key positions. The credit and reputation score of an enterprise can be determined by a comprehensive assessment based on the credit rating of the relevant competent authorities, such as the provincial transportation authorities, and the enterprise's safety and quality records over the past three years.

[0062] in, The technical evaluation score can be quantified by the semantic matching degree of the technical documents, which can be achieved through the following formula (3): (3) in, Full marks can be preset for experts in the three sub-items of construction organization design, completeness of technical solutions and risk protection measures; It can generate a matching coefficient for LLM based on the retrieved solution details and bidding requirements.

[0063] in, To evaluate the bid price, the deviation rate of the bid price from the benchmark price can be used as a criterion. The step function scoring can be achieved using the following formula (4): (4) If the deviation rate cannot be obtained, the LLM can estimate it based on the budget amount and the reasonableness criteria in the weighted prompts. The reasonableness criteria serve as an "alternative logic" for bid evaluation when hard calculation parameters such as the "deviation rate" are unavailable. This criterion guides the LLM to compare and evaluate whether the bid price is within a scientific profit range and market competition by analyzing the "budget amount" or "project funding source" extracted from the bidding announcement. Specifically, it covers logical verification of whether the bid price matches the resource allocation of the technical solution, a prediction of whether the funding scale can cover implementation costs, and a weighted consideration of the cost advantages unique to collaborative projects within the group. This criterion ensures that the system can still provide a business-supported score based on expert-defined reasonableness principles even with incomplete data.

[0064] In some embodiments, because traditional intelligent recommendation systems are often "black box" systems, only providing a "recommend bidding" or "not recommended" conclusion, they lack visual supporting evidence. Users cannot intuitively see the model's judgment criteria, leading to insufficient trust in the AI ​​system and making it difficult to truly assist human decision-making. Therefore, the method further includes: The bidding recommendation scheme is visualized through a three-level linkage; wherein the three-level linkage includes a result matrix, perspective scoring, and text rendering.

[0065] In this embodiment, the front-end is built based on React (Hooks) and Ant Design to create a single-page application that carries the following core functions: ① Global Situation Awareness: In the workbench module, ECharts renders the "Enterprise Knowledge Brain" in real time, dynamically displaying the total number of entities extracted from the backend and the network of relationships. Users can interactively adjust the number of visualization nodes and relationships to intuitively monitor the status of the knowledge base.

[0066] ② Intelligent Filtering and Dynamic Analysis: The front-end can support multiple file formats such as Excel, Word, and PDF through the Dragger component. Built-in dynamic API routing logic intelligently switches back-end interfaces based on the uploaded file type.

[0067] ③ Dimensional Penetration Display: The recommended bidding solutions are displayed through a three-level linkage. The first level is a results matrix, showing the ranking and core advantage tags such as "internal collaboration" and "high amount"; the second level is a perspective scoring, which breaks down the comprehensive score into three dimensions: business, technology, and pricing, showing the specific scoring basis; the third level is intelligent text rendering, with built-in lightweight Markdown parsing logic that automatically identifies business keywords and applies bolding, hierarchical styles, etc., to improve the readability of the analysis report.

[0068] ④ Automated Report Generation: Supports converting AI scoring details into Blob URLs, or exporting single or batch analysis reports in Markdown format via a backend interface. During the generation process, the frontend automatically removes consecutive spaces and line breaks, and intelligently handles filename encoding to ensure that the exported file structure is standardized.

[0069] It is understood that this embodiment provides interpretable decision support through visualization. Specifically, addressing the problem of existing LLM recommendations lacking a basis, it integrates expert weighting prior experience with front-end visualization technology, breaking down abstract AI scores into specific score items in dimensions such as business, technology, and pricing, and providing an intuitive graph-based display, making the recommendation results highly interpretable, thereby effectively assisting bidding personnel in making scientific decisions.

[0070] This embodiment constructs a knowledge graph based on the acquired bidding-related target content and generates text vectors by calling an embedding model. The knowledge graph is stored in a graph database, and the text vectors are stored in a vector database. In response to a user's query request, low-level and high-level keywords are extracted based on the query request. The query request at least carries bidding information for the bidding project. Based on the low-level and high-level keywords, graph vector collaborative retrieval is performed in the graph database and the vector database to obtain a multi-source fusion context. Based on the multi-source fusion context, a bidding recommendation scheme is generated. This solves the "illusion" and logical discontinuity problems in long document reasoning. It employs graph-enhanced text indexing and a two-layer retrieval paradigm. Compared to simple vector retrieval, it utilizes the graph structure to capture entity relationships across paragraphs, enabling the retrieval of specific details while grasping the overall bidding intent, significantly reducing model illusions and improving the accuracy of bidding analysis.

[0071] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0072] For ease of understanding, a specific embodiment is used as an example, including the following steps S1-S5: S1: Utilizing VLM and context-preserving techniques, a multimodal document parsing and alignment module is constructed, namely... Figure 2 Multimodal document parsing on the left and Figure 3 A multimodal document parsing and alignment module; The construction steps of the multimodal document parsing and alignment module are as follows: First, run an automated script tool on the local server to read the original tender documents or domain technical documents; for complex images and tables contained in the documents, load a visual encoding model to perform deep feature extraction and content parsing; then, use context position preservation technology to accurately backfill the parsed visual text information into the text stream corresponding to the original document coordinates, keeping its logical connection with the context unchanged, thereby transforming the multimodal heterogeneous document into plain text data with strict alignment in both position and semantics, providing a standardized input stream for subsequent graph construction; S2: Based on a graph-enhanced text indexing strategy, Prompt Engineering is designed, utilizing LLM-driven entity relation extraction to construct a domain knowledge building module for the tender document. Figure 2 The knowledge construction and optimization part of the middle and Figure 3 The tender document domain knowledge construction module; The construction method of the knowledge construction module is as follows: First, the plain text data obtained in S1 is sliced; then, a special Prompt for the tender document domain is designed and applied to perform LLM-driven entity and relation extraction, parsing entity relation triples of type <head entity, relation, tail entity> from the unstructured text, thereby constructing a knowledge graph for the tender document domain; further, entity descriptions are generated through LLM and mapped together with relation edges into structured key-value pairs; finally, the embedding model Sentence-Transformers is called to transform the extracted text into high-dimensional vector representations, and the entities, relation graph structures and vector data are stored in the local graph database and vector database respectively, forming a graph-vector collaborative underlying index structure; S3: Introduces a deduplication mechanism and incremental update logic to achieve dynamic maintenance and optimization of the knowledge base, i.e. Figure 2 The knowledge base optimization part; The construction method of the knowledge base optimization module is as follows: In the index construction stage, a graph deduplication mechanism is introduced to automatically identify and merge semantically repetitive entity nodes and compress the graph size to ensure the simplicity and efficiency of the knowledge base; at the same time, incremental data updates are supported. When new tender documents or technical solutions are uploaded, only the newly added data is indexed locally and mounted to the existing graph. Real-time throughput and fusion of new knowledge can be achieved without reconstructing the global index. S4: Employing a two-layer retrieval paradigm, combining low-level keywords with high-level abstract topics, a hybrid retrieval reasoning module is constructed, namely... Figure 2 and Figure 4 The mixed retrieval reasoning part; The operating logic of the hybrid retrieval reasoning module is as follows: In response to the user's query request, the system simultaneously extracts low-level specific keywords and high-level abstract topics; it uses a vector library to perform fuzzy search based on semantic similarity, and at the same time uses a graph library to perform one-hop expansion based on graph structure; this process can simultaneously capture the "specific entity details" and "global semantic background" in the tender document, solving the problem of taking things out of context in traditional flat retrieval. It should be noted that during the front-end search, only the "tender announcement" or "tender document" is uploaded (i.e., only the file is uploaded). Analysis instructions are given to the LLM through Prompt keywords. These instructions are implemented by adding relevant constraints to the Prompt keywords on the back-end. Essentially, the file to be searched is converted into a text, which is then set as a "user question." Prompt keyword constraints are added to the user question, and the search is performed as a single "user question."

[0073] The following is a feasible example of the retrieval process: final_prompt = f"""The following is a list of tender documents / announcements to be analyzed (please treat all as the latest tender announcements for analysis; do not refuse to score based on date or nature of announcement, as this may include **Party A's tender announcements** or **Party B's bidding proposals**): {context_content}---Task---As the company's chief bidding operations analyst, please select the **most worthwhile {ai_stable_count} projects** from the above list based on your local knowledge base (historical tenders, qualifications, personnel performance). {SCORING_LOGIC}---Important Filtering Instructions---1. **Internal Collaboration Priority**: Projects with titles or content containing the words "CCCC," "subcontracting," "internal," or "system" are high-value opportunities for internal collaboration within the group. **Regardless of the amount or whether it is a subcontracting, please be sure to keep them in the recommended list** and do not filter them. 2. **Regular Filtering**: In addition to retaining the above collaborative projects, further filter for regular projects with large amounts and matching qualifications. 3. **Abnormal Data Handling**: If the title contains "ResultSnapshot" or "Not Found", please ignore this item. ---Output Requirements--- 1. **Must** output only one standard JSON array. 2. JSON format example: [{{"id": "ITEM_1","score": 95,"reason": "Passed the veto item, qualification fully matched.","analysis": "**Project Name**: S5 line of a certain province...\\n**Decision Recommendation**: Recommended to register...\\n## Reason for Recommendation: Veto item review (veto)\\n**Enterprise Qualification**: Possesses first-level qualification...\\n...","tags": ["Strongly Recommended", "Qualification Matched"]}}] 3. The score is an integer between 0 and 100, and must be strictly scored according to the "three-layer funnel".4. **The content of the `analysis` field must strictly adhere to the following format (preserving newline characters):** ****Project Name**: (Extract title or filename) **Decision Recommendation**: Recommend registration / Try with caution / Abandon (brief reasons) ## Recommendation Reasons: Veto Item Review (One-vote veto) **Company Qualification**: (Detailed analysis of whether it possesses Level 1 qualification, etc.) **Mandatory Online Performance Check**: (Detailed analysis of years, amount, independent general contracting requirements) **Financial Threshold**: (Detailed analysis of funds, revenue, etc.) **Reputation Requirements**: (Detailed analysis of credit rating, blacklist status) **Personnel Weaknesses**: (Detailed analysis of Level 1 construction engineer certificate, performance requirements) ## Deep Matching Analysis: **Business Evaluation**: (Analysis of owner relationship, regional advantages) **Technical Evaluation**: (Analysis of scheme difficulty, personnel experience) **Quotation Evaluation**: (Analysis of budget profit margin) 5. **IMPORTANT**: Ensure the output is in valid JSON format. Newline characters within the string must be escaped as \\n. 6. **Special Requirements for the `summary` Field:** Please ignore garbled characters and invalid characters in the original text, extract the project name, core budget, key qualification requirements, and project objectives, and generate a coherent summary (approximately 300 words). 7. Please strictly sort by score from highest to lowest and return the first {ai_stable_count} results. S5: The retrieved multi-source fusion context is fused with pre-defined expert prior experience. A full-dimensional qualification matching and pre-defined weighted scoring are performed through an LLM decision chain, driving an interactive visualization engine to output the decision results—that is, decision generation and recommendation. Figure 2 and Figure 4 The section shown is about expert experience-driven reasoning, decision-making, and interaction.

[0074] The above-mentioned intelligent analysis and recommendation method for tender documents based on graph structure retrieval enhancement includes two core aspects in step S5: decision reasoning and visualization. (1) At the decision-making reasoning level, the system constructs a weighted Prompt based on expert prior experience, driving the LLM to perform logical reasoning on the bidding project. Based on the rich context containing graph structure information retrieved by S4, the model automatically compares the enterprise qualifications with the bidding requirements, generates a logically consistent scoring report through chain of thought reasoning, and finally outputs the Top N bidding recommendation schemes.

[0075] (2) At the visualization level, the S5 decision output layer integrates a core architecture and visualization engine, which can render the "enterprise knowledge brain" in real time and support the display of in-depth analysis results through three-level linkage: The first level is the results matrix: intuitively displaying project rankings, matching progress bars, and core advantage tags such as "internal collaboration" and "high amount"; The second level is the perspective scoring model: the comprehensive score is broken down into three core dimensions: business, technology, and pricing. The scoring criteria for each item are displayed in detail through sub-tables, so as to make the score traceable. The third level is intelligent text rendering: It has built-in lightweight Markdown parsing logic to process the AI-generated in-depth analysis reports, automatically identify business keywords and apply bolding, hierarchical styles, etc., to improve the readability of the analysis reports.

[0076] Based on the same inventive concept, this application also provides a tender document intelligent analysis and recommendation device for implementing the tender document intelligent analysis and recommendation method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the tender document intelligent analysis and recommendation device provided below can be found in the limitations of the tender document intelligent analysis and recommendation method described above, and will not be repeated here.

[0077] In one exemplary embodiment, such as Figure 5 As shown, a tender document intelligent analysis and recommendation device is provided, including a first generation module 100, a storage module 200, an extraction module 300, a retrieval module 400, and a second generation module 500, wherein: The first generation module 100 is used to construct a knowledge graph based on the acquired target content related to bidding and to call the embedding model to generate text vectors.

[0078] The storage module 200 is used to store the knowledge graph in a graph database and the text vector in a vector database.

[0079] The extraction module 300 is used to respond to a user's query request and extract low-level keywords and high-level keywords based on the query request; wherein, the query request carries at least the bidding information of the bidding project.

[0080] The retrieval module 400 is used to perform graph vector collaborative retrieval in the graph database and the vector database based on the low-level keywords and the high-level keywords to obtain a multi-source fusion context.

[0081] The second generation module 500 is used to generate a bidding recommendation scheme based on the multi-source fusion context.

[0082] In some embodiments, the first generation module 100 is specifically used for: Based on the images and tables obtained from the target content related to bidding, the Visual Language Model (VLM) is invoked to generate visual text information.

[0083] Based on context-preserving technology, the visual text information is backfilled into the corresponding coordinate position in the text stream of the target content to obtain text data that is aligned in position and semantics.

[0084] The Large Language Model (LLM) is invoked to extract triples from the text data; wherein the triples include a head entity, a relation, and a tail entity.

[0085] A knowledge graph is constructed based on the triples.

[0086] In some embodiments, the first generation module 100 is further configured to: Generate summary descriptions for the head entity, the relation, and the tail entity, respectively.

[0087] The head entity, the relation, the tail entity, and the corresponding summary description are mapped to structured key-value pairs; wherein the key of the key-value pair is the head entity, the relation, or the tail entity, and the value of the key-value pair is the summary description corresponding to the key.

[0088] The embedding model is invoked to convert the key into a high-dimensional vector representation, resulting in a text vector.

[0089] In some embodiments, the retrieval module 400 is specifically used for: Based on the low-level keywords, a first target text vector is retrieved from the first text vectors converted from the head entity or the tail entity in the vector database, and based on the high-level keywords, a second target text vector is retrieved from the second text vectors converted from the relation in the vector database.

[0090] Determine the first target head entity or the first target tail entity corresponding to the first target text vector, and determine the first target relationship corresponding to the second target text vector.

[0091] From the graph database, determine the first neighboring entity of the first target head entity or the first target tail entity, and the second neighboring entity of the first target relationship.

[0092] Based on the key-value pairs, target summary descriptions corresponding to the first target head entity or first target tail entity, the first target relationship, the first neighboring entity, and the second neighboring entity are determined respectively.

[0093] By concatenating all the target summary descriptions and combining them with the first target head entity or the first target tail entity, the first target relationship, the first neighboring entity, and the second neighboring entity, a multi-source fusion context is obtained.

[0094] In some embodiments, the apparatus further includes: The update module is used to update the graph database and the vector database based on an incremental update mechanism in response to new content related to bidding.

[0095] In some embodiments, the apparatus further includes: The third generation module is used to generate a bidding recommendation scheme by calling LLM based on the multi-source fusion context and preset weighted prompt words.

[0096] The weighting prompts are constructed based on hard constraints and soft scoring terms. The hard constraints and soft scoring terms are used to generate a comprehensive score for each bidding project. The hard constraints are also used to perform pre-verification on the soft scoring terms. The bidding recommendation scheme is determined based on the top N bidding projects in terms of comprehensive score, where N is a positive integer.

[0097] In some embodiments, the comprehensive score The CoT inference is obtained through inference chain CoT inference, which is implemented by the following formula: Among them, the As the veto coefficient, in the event that the hard constraint triggers a veto to pass, the... Otherwise, the above 1; the aforementioned The above and stated The scores are weighted scores from three dimensions: business evaluation, technical evaluation, and price evaluation.

[0098] In some embodiments, the apparatus further includes: The display module is used to visualize the bidding recommendation scheme through a three-level linkage; wherein, the three-level linkage includes a result matrix, perspective scoring, and text rendering.

[0099] Each module in the aforementioned intelligent tender document analysis and recommendation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0100] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 6As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a tender document intelligent analysis and recommendation method.

[0101] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0102] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0103] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0104] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for intelligent analysis and recommendation of tender documents, characterized in that, The intelligent analysis and recommendation method for tender documents includes: Based on the acquired target content related to bidding, a knowledge graph is constructed, and an embedding model is invoked to generate text vectors; The knowledge graph is stored in a graph database, and the text vectors are stored in a vector database; In response to a user's query request, low-level keywords and high-level keywords are extracted based on the query request; wherein, the query request carries at least the bidding information of the bidding project; Based on the low-level keywords and the high-level keywords, graph vector collaborative retrieval is performed in the graph database and the vector database to obtain a multi-source fusion context; Based on the multi-source fusion context, a bidding recommendation scheme is generated.

2. The intelligent analysis and recommendation method for tender documents according to claim 1, characterized in that, The process of constructing a knowledge graph based on the acquired target content related to bidding and tendering includes: Based on the images and tables in the target content related to bidding, the Visual Language Model (VLM) is invoked to generate visual text information. Based on context-preserving technology, the visual text information is backfilled into the corresponding coordinate position in the text stream of the target content to obtain text data that is aligned in position and semantics. The Large Language Model (LLM) is invoked to extract triples from the text data; wherein the triples include a head entity, a relation, and a tail entity; A knowledge graph is constructed based on the triples.

3. The intelligent analysis and recommendation method for tender documents according to claim 2, characterized in that, The process of calling the embedding model to generate text vectors includes: Generate summary descriptions for the head entity, the relation, and the tail entity, respectively; The head entity, the relation, the tail entity, and the corresponding summary description are mapped to structured key-value pairs; wherein the key of the key-value pair is the head entity, the relation, or the tail entity, and the value of the key-value pair is the summary description corresponding to the key; The embedding model is invoked to convert the key into a high-dimensional vector representation, resulting in a text vector.

4. The intelligent analysis and recommendation method for tender documents according to claim 3, characterized in that, The process involves performing a graph-vector collaborative retrieval based on the low-level keywords and the high-level keywords in the graph database and the vector database to obtain a multi-source fusion context, including: Based on the low-level keywords, a first target text vector is retrieved from the first text vector converted from the head entity or the tail entity in the vector database; and based on the high-level keywords, a second target text vector is retrieved from the second text vector converted from the relation in the vector database. Determine the first target head entity or the first target tail entity corresponding to the first target text vector, and determine the first target relationship corresponding to the second target text vector; From the graph database, determine the first neighboring entity of the first target head entity or the first target tail entity, and the second neighboring entity of the first target relationship; Based on the key-value pairs, the target summary descriptions corresponding to the first target head entity or the first target tail entity, the first target relationship, the first neighboring entity, and the second neighboring entity are determined respectively. By concatenating all the target summary descriptions and combining them with the first target head entity or the first target tail entity, the first target relationship, the first neighboring entity, and the second neighboring entity, a multi-source fusion context is obtained.

5. The intelligent analysis and recommendation method for tender documents according to claim 1, characterized in that, The method further includes: In response to the acquisition of new content related to bidding, the graph database and the vector database are updated based on an incremental update mechanism.

6. The intelligent analysis and recommendation method for tender documents according to claim 1, characterized in that, The step of generating a bidding recommendation scheme based on the multi-source fusion context includes: Based on the multi-source fusion context and preset weighted prompt words, LLM is invoked to generate a bidding recommendation scheme; The weighting prompts are constructed based on hard constraints and soft scoring terms. The hard constraints and soft scoring terms are used to generate a comprehensive score for each bidding project. The hard constraints are also used to perform pre-verification on the soft scoring terms. The bidding recommendation scheme is determined based on the top N bidding projects in terms of comprehensive score, where N is a positive integer.

7. The intelligent analysis and recommendation method for tender documents according to claim 6, characterized in that, The overall score The CoT inference is obtained through inference chain CoT inference, which is implemented by the following formula: Among them, the As the veto coefficient, in the event that the hard constraint triggers a veto to pass, the... Otherwise, the above 1; the aforementioned The above and stated The scores are weighted scores from three dimensions: business evaluation, technical evaluation, and price evaluation.

8. The intelligent analysis and recommendation method for tender documents according to claim 1, characterized in that, The method further includes: The bidding recommendation scheme is visualized through a three-level linkage; wherein the three-level linkage includes a result matrix, perspective scoring, and text rendering.

9. A computer device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the tender intelligent analysis and recommendation method as described in any one of claims 1-8.

10. A readable storage medium, characterized in that, The program or instructions are stored on the readable storage medium, and when the program or instructions are executed by a processor, they implement the steps of the tender intelligent analysis and recommendation method as described in any one of claims 1-8.

Images