A Method and System for Screening Bidding Information Based on Intelligent Agents
By employing a dual-path heterogeneous parallel inference architecture for intelligent agents, and combining the confidence fusion of integrated tree models and large language models, the problem of high false positive and false negative rates in the screening of bidding information in existing technologies is solved, thus achieving efficient screening of bidding information.
Patent Information
- Application Number
- CN202611139743.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-30
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-07-30
AI Technical Summary
In existing technologies, single-type reasoning models struggle to balance accurate matching of structured fields with deep semantic understanding in screening bidding information, resulting in high false positive and false negative rates.
A top-down nine-layer decoupled architecture of intelligent agents is adopted. Through a dual-path heterogeneous parallel inference architecture, the bidding announcement data is parsed into structured business data and text semantic data, which are respectively input into the ensemble tree model and the large language model to generate confidence scores. Conflict detection and fusion are performed through Dempster-Shafer evidence theory, and finally, the screening results are generated based on business rules.
It achieves dual coverage of accurate matching of structured fields in bidding information and deep semantic understanding, reducing the false positive rate and the false negative rate.
Smart Images

Figure CN122654878B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and system for screening bidding information based on intelligent agents. Background Technology
[0002] Automatic tender screening methods based on computer systems are widely used in business opportunity mining. Existing technologies typically employ a single type of inference model to process tender data. For example, they utilize ensemble learning models based on gradient boosting decision trees to quickly classify structured business fields, or pre-trained large language models to perform semantic understanding and matching evaluation of tender announcement text. However, because existing technologies rely on a single model path for tender screening, the accuracy is insufficient. Specifically, while a single ensemble tree model can efficiently handle multi-dimensional structured features, it struggles to understand the complex business logic and implicit requirements in tender announcements; while a single large language model, although possessing strong text understanding capabilities, struggles to accurately fit the multi-dimensional quantitative evaluation standards specific to enterprises, and the inference results lack interpretable deterministic constraints. Neither approach can simultaneously achieve accurate matching of structured fields and deep semantic understanding, resulting in persistently high false positive and false negative rates in tender screening. Summary of the Invention
[0003] This invention provides a method and system for screening bidding information based on intelligent agents, in order to solve the problems of high false positive rate and high false negative rate in the existing bidding information screening technology.
[0004] In a first aspect, embodiments of this application provide a method for filtering bidding information based on an intelligent agent, the method being executed by an intelligent agent, including: Obtain bidding announcement data, extract text content and parse structured fields from the bidding announcement data to obtain structured business data and text semantic data; The structured business data is input into the ensemble tree model of the first inference path to generate a first confidence score; at the same time, historical cases similar to the text semantic data are retrieved from the preset vectorized domain knowledge base, and the text semantic data and the historical cases are concatenated and input into the large language model to generate a second confidence score. Calculate the conflict coefficient between the first confidence level and the second confidence level. Only when the conflict coefficient is lower than a preset conflict threshold, perform evidence theory fusion on the first confidence level and the second confidence level to obtain a comprehensive confidence level. The comprehensive confidence level is forcibly corrected based on pre-configured business rules, and a filtering result label for the bidding announcement data is generated based on the corrected confidence level. The filtering result label includes at least one of: recommended, pending review, or filtered. The intelligent agent adopts a top-down nine-layer decoupled architecture, including: The data storage layer includes a raw database, a feature library, a model library, a vector database, and a relational database, which provides vertical data reading and persistence support for other layers. The data acquisition layer is equipped with website tag crawling tools, an incremental synchronization mechanism, and a data deduplication engine; The data preprocessing layer is configured with text cleaning subunits, structure extraction subunits, normalization subunits, and anomaly detection subunits; The feature engineering layer is configured with a label system construction subunit, a word embedding subunit, and a feature cross selection subunit; The machine learning engine layer is configured with multiple integrated tree-based models, model fusion sub-units, and online learning sub-units; The large model semantic understanding layer is configured with a RAG retrieval enhancement subunit, a Prompt engineering subunit, a Few-shot learning subunit, and a semantic similarity calculation unit, and retrieves the historical cases by interacting with the vector database in the data storage layer. The decision fusion layer is configured with a dual-path confidence-weighted fusion subunit, a business rule engine, and a threshold adaptive adjustment subunit. The output and application layer is configured with an automatic classification sub-unit for bidding information, a recommendation scoring and ranking sub-unit, a risk warning sub-unit, and a visual dashboard. The continuous learning layer is configured with a misjudged sample recovery subunit, an active learning sampling subunit, and a model periodic retraining subunit.
[0005] In some embodiments, the method is executed by an agent employing a top-down nine-layer decoupled architecture, including: The data storage layer includes a raw database, a feature library, a model library, a vector database, and a relational database, which provides vertical data reading and persistence support for other layers. The data acquisition layer is equipped with website tag crawling tools, an incremental synchronization mechanism, and a data deduplication engine; The data preprocessing layer is configured with text cleaning subunits, structure extraction subunits, normalization subunits, and anomaly detection subunits; The feature engineering layer is configured with a label system construction subunit, a word embedding subunit, and a feature cross selection subunit; The machine learning engine layer is configured with multiple integrated tree-based models, model fusion sub-units, and online learning sub-units; The large model semantic understanding layer is configured with a RAG retrieval enhancement subunit, a Prompt engineering subunit, and a Few-shot learning subunit, and retrieves the historical cases by interacting with the vector database in the data storage layer. The decision fusion layer is configured with a dual-path confidence-weighted fusion subunit, a business rule engine, and a threshold adaptive adjustment subunit. The output and application layer is configured with an automatic classification sub-unit for bidding information, a recommendation scoring and ranking sub-unit, a risk warning sub-unit, and a visual dashboard. The continuous learning layer is configured with a misjudged sample recovery subunit, an active learning sampling subunit, and a model periodic retraining subunit.
[0006] In some embodiments, the internal and cross-layer connections of the intelligent agent are as follows: In the data acquisition layer, the outputs of the website tag crawling tool and the incremental synchronization mechanism are jointly input to the data deduplication engine, and the data deduplication engine outputs the deduplicated raw data to the data preprocessing layer. In the data preprocessing layer, the raw data is input to the text cleaning subunit, the cleaned text is passed to the structure extraction subunit, the output of the structure extraction subunit is passed to the normalization subunit and the anomaly detection subunit for parallel processing, and the outputs of the normalization subunit and the anomaly detection subunit are converged and output to the feature engineering layer. In the feature engineering layer, the structured field data in the input data is passed to the label system construction subunit, and the unstructured text data in the input data is passed to the word embedding subunit. The outputs of the label system construction subunit and the word embedding subunit are jointly input to the feature cross selection subunit. The fused feature vector output by the feature cross selection subunit is simultaneously passed to the machine learning engine layer and the large model semantic understanding layer, forming a dual-path parallel inference architecture. In the machine learning engine layer, the fused feature vector is simultaneously input to multiple parallel ensemble tree-based models, and the outputs of each ensemble tree-based model converge to the model fusion subunit. The model fusion subunit outputs the generated first confidence score to the decision fusion layer. In the large model semantic understanding layer, the RAG retrieval enhancement subunit calls the vector database in the data storage layer to retrieve the historical cases. The retrieval results, along with the current text data and the reference examples selected by the Few-shot learning subunit, are input to the Prompt engineering subunit to generate input prompts. The input prompts are inferred by the large language model to generate the second confidence score, and the second confidence score is output to the decision fusion layer. In the decision fusion layer, the dual-confidence weighted fusion subunit receives the first confidence level and the second confidence level and performs evidence fusion. The fusion result is passed to the business rule engine for forced correction. The threshold adaptive adjustment subunit dynamically adjusts the decision threshold of the business rule engine based on historical feedback data. The corrected final confidence level is output to the output and application layer. In the output and application layer, the received final confidence level is processed sequentially by the automatic classification subunit and the recommendation scoring and ranking subunit. The processing results, together with the risk features output by the risk warning subunit, are displayed on the visualization dashboard. The user manual feedback data received by the visualization dashboard is output to the continuous learning layer. In the continuous learning layer, the manual feedback data is input to the misjudged sample recovery subunit. After the active learning sampling subunit selects high-value training samples, it triggers the model periodic retraining subunit to perform retraining. After retraining, through the reverse feedback link, it outputs model weight update signals to the online learning subunit and the model fusion subunit of the machine learning engine layer, vector library update signals to the RAG retrieval enhancement subunit of the large model semantic understanding layer, threshold adjustment signals to the threshold adaptive adjustment subunit of the decision fusion layer, and feature weight optimization signals to the feature cross-selection subunit of the feature engineering layer, forming a closed-loop evolution mechanism.
[0007] In some embodiments, the acquisition of bidding announcement data, including text content extraction and multi-dimensional structured field parsing to obtain structured business data and text semantic data, is executed sequentially by the data acquisition layer, data preprocessing layer, and feature engineering layer, specifically including: The data acquisition layer uses the website bidding information crawling tool to periodically crawl the original bidding announcement data of external bidding websites, and uses the data deduplication engine to deduplicate based on content fingerprints; The data preprocessing layer receives the deduplicated data and sequentially calls the text cleaning subunit to remove HTML tags and noisy text, the structured extraction subunit to extract key fields based on named entity recognition, the normalization subunit to perform robust scaling processing on numerical fields, and the anomaly detection subunit to remove abnormal data based on the isolated forest algorithm to obtain standardized data. The feature engineering layer receives the standardized data, calls the tag system construction subunit to construct nine-dimensional structured tags that at least cover industry, personnel size, whether it is a specific type of enterprise, and procurement budget, calls the word embedding subunit to perform semantic encoding on the text content to obtain a text semantic vector, calls the feature cross selection subunit to perform feature cross combination of the nine-dimensional structured tags and the text semantic vector and filter based on model gain, and outputs the structured business data and text semantic data.
[0008] In some embodiments, in the feature engineering layer, the nine-dimensional structured tags include national standard industry, industry related to the target enterprise, customer personnel size, whether it is a specific type of enterprise, establishment year, cooperation status, procurement budget, procurement demand number of people, and bidding qualification requirements; the word embedding subunit uses TF-IDF or BERT model to semantically encode the bidding announcement text, and the customer personnel size is binned to a preset level, and the procurement budget is robustly scaled.
[0009] In some embodiments, the step of inputting the structured business data into the ensemble tree model of the first inference path to generate a first confidence score, which is executed by the machine learning engine layer, specifically includes: The structured business data is input in parallel into the XGBoost classifier, random forest, and LightGBM ensemble tree-based models of the machine learning engine layer; The outputs of the three base models are stacked and fused using the model fusion subunit to generate the first confidence score. The online learning subunit is equipped with the FTRL algorithm, which is used to perform incremental weight updates on the three integrated tree-based models when it receives the update signal from the back feedback link.
[0010] In some embodiments, the step of concatenating the text semantic data with the historical cases and inputting the result into a large language model to generate a second confidence score, which is then executed by the semantic understanding layer of the large model, specifically includes: The RAG retrieval enhancement subunit transforms the text semantic data into query vectors, performs a two-stage retrieval combining BM25 coarse recall and vector cosine fine ranking in the vector database of the data storage layer, and re-ranks the recall results by maximum marginal relevance to obtain the similar historical cases. The Few-shot learning subunit selects the K most relevant samples from the similar historical cases; The Prompt subunit provides a prompt template for the constraint output format. The K samples, the text semantic data, and the prompt template are concatenated and input into the large language model. The large language model outputs the inference result containing the conclusion and confidence in a structured format, from which the second confidence is extracted.
[0011] In some embodiments, the decision fusion layer specifically performs the following operations: The dual-path confidence weighted fusion subunit constructs basic probability allocation functions for the machine learning engine layer and the large model semantic understanding layer, and calculates the conflict coefficient K based on Dempster-Shafer evidence theory. When K ≤ 0.8, the first confidence level and the second confidence level are fused using the Dempster combination rule to obtain the comprehensive confidence level; when K > 0.8, fusion is rejected and the corresponding bidding announcement data is directly marked as pending manual review. The business rule engine has built-in hard filtering rules and weighted adjustment rules to forcibly correct the labeling results after the comprehensive confidence or rejection fusion. The threshold adaptive adjustment subunit dynamically adjusts the upper and lower thresholds used for the three categories based on historical feedback data, where the three categories correspond to recommendation, pending review, and filtering.
[0012] Secondly, embodiments of this application provide a tender information screening system based on an intelligent agent, wherein the intelligent agent-based tender information screening system is equipped with an intelligent agent, including: The data acquisition module is used to acquire bidding announcement data, extract text content from the bidding announcement data and parse multi-dimensional structured fields to obtain structured business data and text semantic data. The dual-path inference module includes a first inference unit and a second inference unit. The first inference unit is used to input the structured business data into an integrated tree model to generate a first confidence level. The second inference unit is used to retrieve historical cases similar to the text semantic data from a preset vectorized domain knowledge base, and then concatenate the text semantic data with the historical cases and input them into a large language model to generate a second confidence level. The decision fusion module is used to calculate the conflict coefficient between the first confidence level and the second confidence level, and only when the conflict coefficient is lower than a preset conflict threshold, the first confidence level and the second confidence level are fused according to evidence theory to obtain a comprehensive confidence level. The rule output module is used to forcibly correct the comprehensive confidence level based on pre-configured business rules, and generate filtering result tags for the bidding announcement data based on the corrected confidence level. The filtering result tags include at least one of recommended, pending review, or filtered. The intelligent agent adopts a top-down nine-layer decoupled architecture, including: The data storage layer includes a raw database, a feature library, a model library, a vector database, and a relational database, which provides vertical data reading and persistence support for other layers. The data acquisition layer is equipped with website tag crawling tools, an incremental synchronization mechanism, and a data deduplication engine; The data preprocessing layer is configured with text cleaning subunits, structure extraction subunits, normalization subunits, and anomaly detection subunits; The feature engineering layer is configured with a label system construction subunit, a word embedding subunit, and a feature cross selection subunit; The machine learning engine layer is configured with multiple integrated tree-based models, model fusion sub-units, and online learning sub-units; The large model semantic understanding layer is configured with a RAG retrieval enhancement subunit, a Prompt engineering subunit, a Few-shot learning subunit, and a semantic similarity calculation unit, and retrieves the historical cases by interacting with the vector database in the data storage layer. The decision fusion layer is configured with a dual-path confidence-weighted fusion subunit, a business rule engine, and a threshold adaptive adjustment subunit. The output and application layer is configured with an automatic classification sub-unit for bidding information, a recommendation scoring and ranking sub-unit, a risk warning sub-unit, and a visual dashboard. The continuous learning layer is configured with a misjudged sample recovery subunit, an active learning sampling subunit, and a model periodic retraining subunit.
[0013] Thirdly, embodiments of this application provide an intelligent agent for executing the agent-based bidding information screening method as described in any one of the embodiments of this application.
[0014] In one scheme provided by the above-mentioned agent-based bidding information filtering method and system, the method is executed by an agent and includes: acquiring bidding announcement data; extracting text content and parsing structured fields from the bidding announcement data to obtain structured business data and text semantic data; inputting the structured business data into the ensemble tree model of the first inference path to generate a first confidence level; simultaneously, retrieving historical cases similar to the text semantic data from a preset vectorized domain knowledge base, concatenating the text semantic data and historical cases, and inputting them into a large language model to generate a second confidence level; calculating the conflict coefficient between the first confidence level and the second confidence level, and only when the conflict coefficient is lower than a preset conflict threshold, fusing the first confidence level and the second confidence level with evidence theory to obtain a comprehensive confidence level; forcibly correcting the comprehensive confidence level based on pre-configured business rules, and generating filtering result labels for the bidding announcement data based on the corrected confidence level, the filtering result labels including at least one of: recommended, pending review, or filtered; wherein, the agent adopts a top-down nine-layer decoupled architecture, including: a data storage layer containing a raw database, a feature library, a model library, a vector database, and a relational layer. The system comprises the following layers: a database layer providing vertical data reading and persistence support for other layers; a data acquisition layer with website tag crawling tools, incremental synchronization mechanisms, and a data deduplication engine; a data preprocessing layer with text cleaning, structured extraction, normalization, and anomaly detection subunits; a feature engineering layer with tag system construction, word embedding, and feature cross-selection subunits; a machine learning engine layer with multiple ensemble tree-based models, model fusion subunits, and online learning subunits; a large-scale model semantic understanding layer with RAG retrieval enhancement, Prompt engineering, Few-shot learning, and semantic similarity calculation subunits, and interacts with the vector database in the data storage layer to retrieve historical cases; a decision fusion layer with dual-path confidence weighted fusion subunits, a business rule engine, and threshold adaptive adjustment subunits; an output and application layer with automatic tag classification, recommendation scoring and ranking, risk warning, and visualization dashboards; and a continuous learning layer with misjudged sample collection, active learning sampling, and periodic model retraining subunits.By simultaneously extracting text content and parsing multi-dimensional structured fields from bidding announcement data, high-dimensional feature data containing both structured business information and textual semantic information is obtained. This structured business data is then input into an integrated tree model path to generate a first confidence level. Simultaneously, similar historical cases are retrieved from a vectorized domain knowledge base and concatenated with the textual semantic data before being input into a large language model path to generate a second confidence level. The conflict coefficient between the two confidence levels is then calculated, and evidence theory fusion is performed only when the conflict is below a preset threshold to obtain a comprehensive confidence level. Based on pre-configured business rules, the comprehensive confidence level is forcibly corrected, and a filtering result label is generated that at least distinguishes between recommended, pending review, or filtered results. Thus, through a dual-path heterogeneous reasoning and evidence theory conflict resolution mechanism, the problem of insufficient filtering accuracy caused by the limited information representation capabilities of a single model path is overcome. This achieves dual coverage of accurate matching of structured fields and deep semantic understanding of bidding announcements, reducing the false positive and false negative rates in bidding information filtering. Attached Figure Description
[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0016] Figure 1 This is a schematic diagram of the information filtering system in one embodiment of the present invention; Figure 2 This is a flowchart illustrating a bidding information filtering method based on an intelligent agent according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an intelligent agent according to one embodiment of the present invention; Figure 4 yes Figure 2 A schematic diagram of the implementation process of step S10; Figure 5 yes Figure 2 A schematic diagram of the implementation process of step S20; Figure 6 yes Figure 2 Another implementation flowchart of step S20; Figure 7 This is a schematic diagram of a bidding information screening system based on intelligent agents; Figure 8 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention. Detailed Implementation
[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0018] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof. It should also be understood that, as used in this specification and the appended claims, the term "and / or" refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.
[0019] Furthermore, in the description of this invention and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.
[0020] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of the invention include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.
[0021] It should be understood that the sequence number of each step in the following embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.
[0022] To illustrate the technical solution of the present invention, specific embodiments are described below.
[0023] Automatic tender screening methods based on computer systems are widely used in business opportunity mining. Existing technologies typically employ a single type of inference model to process tender data. For example, they utilize ensemble learning models based on gradient boosting decision trees to quickly classify structured business fields, or pre-trained large language models to perform semantic understanding and matching evaluation of tender announcement text. However, because existing technologies rely on a single model path for tender screening, the accuracy is insufficient. Specifically, while a single ensemble tree model can efficiently handle multi-dimensional structured features, it struggles to understand the complex business logic and implicit requirements in tender announcements; while a single large language model, although possessing strong text understanding capabilities, struggles to accurately fit the multi-dimensional quantitative evaluation standards specific to enterprises, and the inference results lack interpretable deterministic constraints. Neither approach can simultaneously achieve accurate matching of structured fields and deep semantic understanding, resulting in persistently high false positive and false negative rates in tender screening.
[0024] The fundamental reason for the insufficient accuracy of the aforementioned tender information screening lies in the fact that the tender information screening task itself has dual heterogeneous information processing requirements: on the one hand, it requires precise quantitative matching and rule constraint verification of structured fields such as budget amount, qualification requirements, and industry classification; on the other hand, it requires deep semantic understanding of the natural language descriptions in the tender announcements to identify implicit business needs, eliminate ambiguous expressions, and assess the degree of matching with its own service capabilities. However, in existing technologies, although ensemble tree models can efficiently process high-dimensional discrete features and accurately model numerical fields through binning and cross-referencing, the model's feature input relies on manually predefined structured fields and lacks the ability to model complex semantic relationships in unstructured long texts. When faced with tender announcements that are vaguely worded, contain industry terminology abbreviations, or have multi-layered nested conditions, they are prone to misjudgment due to the lack of semantic information. While large language models possess powerful contextual understanding and knowledge reasoning capabilities, enabling them to capture implicit information from long texts, their output is essentially a probability-driven autoregressive generation result. They lack deterministic constraints on structured fields and interpretable decision boundaries. When faced with hard rules requiring precise matching, such as monetary thresholds, geographical ranges, or exclusion of specific qualifications, they are prone to producing illusionary outputs that contradict business realities. More importantly, when a single model path encounters tags that simultaneously contain strong rule constraints and complex semantics, an irreconcilable structural contradiction exists between the capabilities of the two types of models. Using either path alone will lead to the neglect or misjudgment of certain types of information, resulting in the omission of high-value tags or the incorrect recommendation of low-quality tags, leading to persistently high overall false positive and false negative rates.
[0025] To address the aforementioned issues, this application proposes a method and system for screening bidding information based on intelligent agents. By constructing a dual-path heterogeneous parallel inference architecture, the bidding announcement data is parsed into structured business data and textual semantic data. These are input into an ensemble tree model path for precise matching of quantitative features to generate a first confidence level, and into a large language model path combined with historical cases retrieved from a vectorized domain knowledge base for deep semantic understanding to generate a second confidence level. Then, the Dempster-Shafer evidence theory is used to detect and fuse the confidence levels of the two paths. A comprehensive confidence level is generated only when the conflict coefficient is below a preset threshold. Based on pre-configured business rules, the comprehensive confidence level is forcibly corrected, and a screening result label is output, distinguishing between recommended, pending review, or filtered results. This organically integrates the deterministic constraint capability of the ensemble tree model with the semantic generalization capability of the large language model, systematically solving the technical contradiction that a single model path cannot simultaneously address structured field matching and deep semantic understanding due to limited information representation capabilities, thus reducing the false positive and false negative rates in bidding information screening.
[0026] The agent-based bidding information filtering method provided in this invention can be applied to, for example... Figure 1 The information filtering system shown includes an agent-based bidding information filtering device (here, a virtual device corresponding to the method, and targets controlled, detected, or communicated by the virtual device, such as servers and terminal devices), wherein the terminal devices communicate with the server via a network or bus.
[0027] In practical applications, this information filtering system can be deployed on terminal devices. The terminal devices acquire bidding announcement data, extract text content, and parse multi-dimensional structured fields to obtain structured business data and textual semantic data. The structured business data is then input into a locally deployed ensemble tree model to generate a first confidence score. Simultaneously, historical cases similar to the textual semantic data are retrieved from a local or remote vectorized domain knowledge base. The textual semantic data and historical cases are concatenated and input into a large language model to generate a second confidence score. The conflict coefficient between the first and second confidence scores is calculated, and only the conflict coefficient is considered. When the confidence level is below a preset conflict threshold, the first confidence level and the second confidence level are fused using evidence theory to obtain a comprehensive confidence level. Based on pre-configured business rules, the comprehensive confidence level is forcibly corrected, and based on the corrected confidence level, a filtering result label for the bidding announcement data is generated. The filtering result label includes at least one of recommended, pending review, or filtered. This completes the localized intelligent filtering of bidding information on the terminal device side, reducing the real-time dependence on cloud communication. At the same time, through dual-path heterogeneous reasoning and evidence theory conflict resolution mechanism, it takes into account both accurate matching of structured fields and deep semantic understanding, effectively reducing the false positive rate and false negative rate of bidding information filtering.
[0028] The information filtering system can be a terminal device, which refers to a device that corresponds to the server and provides local services to customers. This terminal device can be installed on, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices.
[0029] The information filtering system can also be a server, which can be implemented using a standalone server or a server cluster consisting of multiple servers.
[0030] In practical applications, this information filtering system can be deployed on a server. The server receives bidding announcement data uploaded by terminal devices or actively collects bidding announcement data from external bidding websites. It extracts text content and parses multi-dimensional structured fields from the bidding announcement data to obtain structured business data and text semantic data. The structured business data is input into the ensemble tree model of the first inference path to generate a first confidence score. Simultaneously, it retrieves historical cases similar to the text semantic data from a pre-set vectorized domain knowledge base. The text semantic data and historical cases are concatenated and input into a large language model to generate a second confidence score. The conflict coefficient between the first and second confidence scores is calculated, and only in conflict... When the coefficient is lower than the preset conflict threshold, the first confidence level and the second confidence level are fused using evidence theory to obtain a comprehensive confidence level. Based on the pre-configured business rules, the comprehensive confidence level is forcibly corrected, and a screening result label for the bidding announcement data is generated according to the corrected confidence level. The screening result label includes at least one of recommended, pending review, or filtered, and the screening result label is sent to the terminal device for display. In this way, the powerful computing power of the server is used to complete dual-path parallel reasoning and evidence fusion decision-making. Through the dual-path heterogeneous reasoning mechanism and evidence theory conflict resolution strategy, dual coverage of accurate matching of structured fields and deep semantic understanding is achieved, reducing the false positive rate and false negative rate of bidding information screening.
[0031] A server can be implemented using a standalone server or a server cluster consisting of multiple servers.
[0032] In one embodiment, such as Figure 2 As shown, a method for filtering bidding information based on intelligent agents is provided, which can be applied to... Figure 1 The following steps are used as an example of the information filtering system in the text: S10-S50.
[0033] S10. Obtain the bidding announcement data, extract the text content and parse the structured fields of the bidding announcement data to obtain structured business data and text semantic data.
[0034] For example, raw tender announcement page data is collected from at least one tender information publishing website. The collected raw tender announcement page data undergoes text extraction to remove non-textual noise such as HTML tags, script code, and advertising information, resulting in a plain text tender announcement. Named entity recognition processing is performed on the tender announcement text to extract key fields such as project name, procuring entity name, budget amount, bid deadline, and contact information. Regular expressions are used to standardize the currency unit in the budget amount and the date format. The extracted numerical fields are binned and normalized. For example, the size of the customer personnel is divided into six tiers according to preset intervals, and the procurement budget amount is robustly scaled and mapped to a unified dimension interval to eliminate numerical scale differences between different fields. Anomaly detection is performed on each tender announcement data point during processing. Anomaly scores are calculated for each data point based on the isolated forest algorithm. Data records with anomaly scores exceeding a preset threshold are marked as abnormal data and removed. The structured fields retained after anomaly detection are constructed into multi-dimensional structured labels. These labels cover at least nine dimensions, including national standard industry classification, customer personnel size level, whether it is a specific type of enterprise, establishment year range, historical cooperation status level, normalized value of procurement budget, procurement demand number level, and bidding qualification requirements. Word embedding processing is performed on the text content of the bidding announcement. A pre-trained BERT model or TF-IDF vectorization method is used to transform the unstructured text into fixed-dimensional text semantic vectors. These text semantic vectors are then combined with the multi-dimensional structured labels through feature cross-combination. After feature filtering based on model gain, the fused structured business data and text semantic data are output.
[0035] S20. Input the structured business data into the ensemble tree model of the first inference path to generate the first confidence level. At the same time, retrieve historical cases similar to the text semantic data from the preset vectorized domain knowledge base, concatenate the text semantic data and historical cases, and input them into the large language model to generate the second confidence level.
[0036] For example, the first inference path is configured with an ensemble tree model based on a stacking fusion strategy. This ensemble tree model includes three base models: an XGBoost classifier, a random forest classifier, and a LightGBM classifier. The three base models receive the input structured business data in parallel. The XGBoost classifier performs forward inference with a hyperparameter configuration of 500 trees, a maximum depth of 8, and a learning rate of 0.05. The random forest classifier performs voting inference with a hyperparameter configuration of 300 trees and a maximum depth of 12. The LightGBM classifier performs gradient inference with a hyperparameter configuration of 600 trees and 63 leaf nodes. The output results of the three base models are converged to the model fusion subunit. The model fusion subunit uses a secondary logistic regression model to perform weighted fusion of the outputs of the three base models to generate a first confidence level representing that the bidding announcement belongs to the target tender information. The second inference path is equipped with a large language model. Before inputting the text semantic data into the large language model, the text semantic data is first converted into query vectors. A two-stage retrieval is performed in the vectorized domain knowledge base. In the first stage, the BM25 algorithm is used to perform coarse keyword recall to obtain a set of candidate cases. In the second stage, the vector cosine similarity algorithm is used to fine-rank the set of candidate cases. The fine-ranked candidate cases are then re-ranked by maximum marginal relevance to select the top K historical cases with diversity. The text descriptions, corresponding screening conclusions, and confidence scores of the selected K historical cases are used as contextual information and concatenated with the text semantic data and a preset prompt template. The concatenated complete input data is input into the large language model. Under the constraints of the prompt template, the large language model outputs the inference results in a structured format, which includes the matching conclusion, confidence score, and reasoning. The second confidence score, which indicates that the tender notice belongs to the target tender, is extracted from the inference results.
[0037] S30. Calculate the conflict coefficient between the first confidence level and the second confidence level. Only when the conflict coefficient is lower than the preset conflict threshold, perform evidence theory fusion on the first confidence level and the second confidence level to obtain the comprehensive confidence level.
[0038] For example, basic probability allocation functions are constructed for the first and second inference paths, respectively. The basic probability allocation function for the first inference path is based on the first confidence level multiplied by the historical accuracy of the first inference path as a weighting factor, while the basic probability allocation function for the second inference path is based on the second confidence level multiplied by the historical accuracy of the second inference path as a weighting factor. The two basic probability allocation functions assign probability qualities to the three hypotheses of "match," "mismatch," and "uncertainty," respectively. The conflict coefficient K between the two pieces of evidence from the first and second inference paths is calculated based on Dempster-Shafer evidence theory. The conflict coefficient K is calculated by multiplying the probability qualities assigned to disjoint hypotheses in the basic probability allocation functions of the first and second inference paths and summing the results. The calculated conflict coefficient K is compared with a preset conflict threshold of 0.8. When the conflict coefficient K is less than or equal to 0.8, it is determined that there is no fundamental conflict between the inference conclusions of the two paths. The first and second confidence levels are orthogonally summed and fused using Dempster's combination rule, and the probability quality assigned to the "match" hypothesis after fusion is used as the comprehensive confidence level. When the conflict coefficient K is greater than 0.8, it is determined that there is a serious discrepancy between the reasoning conclusions of the two paths, the evidence fusion operation is refused, and the bidding announcement data corresponding to the conflict is directly marked as pending manual review, without generating a comprehensive confidence score.
[0039] S40. Based on the pre-configured business rules, the overall confidence level is forcibly corrected, and based on the corrected confidence level, the filtering result labels of the bidding announcement data are generated. The filtering result labels include at least one of the following: recommended, pending review, or filtered.
[0040] For example, the pre-configured business rules include two categories: hard filtering rules and weighted adjustment rules. Hard filtering rules define the logical conditions that force the overall confidence level to the lowest or highest value under specific conditions. Weighted adjustment rules define the adjustment logic that adjusts the overall confidence level upward or downward according to a preset weight coefficient when specific conditions are met. Key fields in the structured business data are matched one by one with the rule conditions in the business rule base. When the purchasing unit attribute field in the bidding announcement data matches the "specific type of enterprise" rule condition, the overall confidence level is multiplied by the preset upward adjustment weight coefficient according to the weighted adjustment rule. When the budget amount field in the bidding announcement data is lower than the preset minimum budget threshold, the overall confidence level is forced to zero according to the hard filtering rule. When the qualification requirement field in the bidding announcement data contains specific excluded qualifications, the overall confidence level is forced to zero according to the hard filtering rule. The revised confidence level is compared with preset three-class classification thresholds. These thresholds include an upper threshold for separating recommended and pending-review categories, and a lower threshold for separating pending-review and filtered categories. The specific values of the upper and lower thresholds are dynamically adjusted by the threshold adaptive adjustment subunit based on historical human feedback data. When the revised confidence level is greater than or equal to the upper threshold, the corresponding bidding announcement data is labeled "Recommended"; when the revised confidence level is less than the upper threshold but greater than the lower threshold, the corresponding bidding announcement data is labeled "Pending Review"; and when the revised confidence level is less than or equal to the lower threshold, the corresponding bidding announcement data is labeled "Filtered".
[0041] In one of the solutions provided by the above-mentioned agent-based bidding information filtering method and system, the method includes: acquiring bidding announcement data; extracting text content and parsing structured fields from the bidding announcement data to obtain structured business data and text semantic data; inputting the structured business data into the ensemble tree model of the first inference path to generate a first confidence level; simultaneously, retrieving historical cases similar to the text semantic data from a preset vectorized domain knowledge base, concatenating the text semantic data and historical cases, and inputting them into a large language model to generate a second confidence level; calculating the conflict coefficient between the first confidence level and the second confidence level, and only when the conflict coefficient is lower than a preset conflict threshold, fusing the first confidence level and the second confidence level with evidence theory to obtain a comprehensive confidence level; forcibly correcting the comprehensive confidence level based on pre-configured business rules, and generating filtering result labels for the bidding announcement data based on the corrected confidence level, wherein the filtering result labels include at least one of: recommended, pending review, or filtered. By simultaneously extracting text content and parsing multi-dimensional structured fields from bidding announcement data, high-dimensional feature data containing both structured business information and textual semantic information is obtained. This structured business data is then input into an integrated tree model path to generate a first confidence level. Simultaneously, similar historical cases are retrieved from a vectorized domain knowledge base and concatenated with the textual semantic data before being input into a large language model path to generate a second confidence level. The conflict coefficient between the two confidence levels is then calculated, and evidence theory fusion is performed only when the conflict is below a preset threshold to obtain a comprehensive confidence level. Based on pre-configured business rules, the comprehensive confidence level is forcibly corrected, and a filtering result label is generated that at least distinguishes between recommended, pending review, or filtered results. Thus, through a dual-path heterogeneous reasoning and evidence theory conflict resolution mechanism, the problem of insufficient filtering accuracy caused by the limited information representation capabilities of a single model path is overcome. This achieves dual coverage of accurate matching of structured fields and deep semantic understanding of bidding announcements, reducing the false positive and false negative rates in bidding information filtering.
[0042] In one embodiment, such as Figure 3 As shown, the method is executed by an agent, which adopts a top-down nine-layer decoupled architecture. The agent includes nine functional layers: data storage layer, data acquisition layer, data preprocessing layer, feature engineering layer, machine learning engine layer, large model semantic understanding layer, decision fusion layer, output and application layer, and continuous learning layer.
[0043] The data storage layer comprises five storage units: a raw database, a feature library, a model library, a vector database, and a relational database. The raw database stores the original bidding announcement page data collected from external bidding websites. The feature library stores feature copies of structured business data and text semantic data generated after processing by the feature engineering layer. The model library stores the training version files of the ensemble tree model and the model hyperparameter configuration. The vector database stores the text semantic vectors of historical bidding announcement cases and the corresponding filtering conclusion labels. The relational database stores logs of user manual feedback records, rule configuration parameters, and filtering result labels. The data storage layer provides data reading and persistence support for the other eight functional layers through vertical call interfaces.
[0044] The data acquisition layer is equipped with a website bidding information crawling tool, an incremental synchronization mechanism, and a data deduplication engine. The website bidding information crawling tool periodically accesses at least one bidding information publishing website and downloads the HTML page data of the bidding announcement list page and details page according to a preset crawling cycle. The incremental synchronization mechanism identifies newly published bidding announcements by comparing the timestamps or announcement numbers of the crawled pages. The data deduplication engine performs deduplication judgment on the crawled bidding announcement data based on the URL hash value, page content fingerprint, and announcement number combined key.
[0045] The data preprocessing layer is configured with a text cleaning subunit, a structured extraction subunit, a normalization subunit, and an anomaly detection subunit. The text cleaning subunit is used to remove non-text content such as HTML tags, JavaScript code, and floating advertisements from the original HTML page data and output the tender announcement text in plain text format. The structured extraction subunit is used to extract structured fields such as budget amount, deadline, and purchasing unit name from the plain text text using a named entity recognition model and regular expressions. The normalization subunit is used to perform robust scaling processing on the extracted numerical fields to eliminate dimensional differences. The anomaly detection subunit is used to calculate the anomaly score of each data record based on the isolated forest algorithm and remove data records with anomaly scores exceeding a preset threshold.
[0046] The feature engineering layer is configured with a tag system construction subunit, a word embedding subunit, and a feature cross-selection subunit. The tag system construction subunit is used to construct structured fields into structured tags covering multiple dimensions such as national standard industry, customer personnel size level, whether it is a specific type of enterprise attribute, establishment year range, historical cooperation status level, normalized value of procurement budget, procurement demand number level, and bidding qualification requirement category. The word embedding subunit is used to convert the bidding announcement text into a fixed-dimensional text semantic vector using a pre-trained BERT model or TF-IDF vectorization method. The feature cross-selection subunit is used to perform feature cross-combination of structured tags and text semantic vectors and select the feature dimension with the highest contribution based on the XGBoost model gain value to output a fused feature vector.
[0047] The machine learning engine layer is configured with three ensemble tree base models: XGBoost classifier, Random Forest classifier, and LightGBM classifier, as well as a model fusion subunit and an online learning subunit. The three ensemble tree base models receive fused feature vectors in parallel and perform forward inference independently. The model fusion subunit is used to stack and fuse the outputs of the three base models using a secondary logistic regression model to generate the first confidence score. The online learning subunit is configured with the FTRL algorithm to incrementally update the weight parameters of the three base models when it receives update signals from the continuous learning layer.
[0048] The large-scale model's semantic understanding layer is configured with a RAG retrieval enhancement subunit, a Prompt engineering subunit, a Few-shot learning subunit, and a semantic similarity calculation unit. The RAG retrieval enhancement subunit uses text semantic vectors as query vectors to perform a two-stage similar case retrieval in the vector database of the data storage layer. The Prompt engineering subunit provides a prompt template that constrains the output format of the large language model and concatenates the retrieved historical cases, current text semantic data, and the prompt template into a complete model input. The Few-shot learning subunit selects the K samples with the highest semantic similarity to the current tender announcement from the retrieved historical cases as inference reference examples for the large language model. The semantic similarity calculation unit calculates the statistical distribution of the semantic similarity values between the K retrieved historical cases and the current tender announcement, and sends this statistical distribution to the decision fusion layer.
[0049] The decision fusion layer is configured with a dual-path confidence weighted fusion subunit, a business rule engine, and a threshold adaptive adjustment subunit. The dual-path confidence weighted fusion subunit is used to perform conflict detection and evidence fusion on the first and second confidence levels based on the Dempster-Shafer evidence theory and output a comprehensive confidence level. The business rule engine is used to forcibly correct the comprehensive confidence level based on preset hard filtering rules and weighted adjustment rules. The threshold adaptive adjustment subunit is used to dynamically adjust the upper and lower thresholds for the three-class classification based on the statistical distribution of historical manual feedback data.
[0050] The output and application layer is configured with an automatic bidding information classification subunit, a recommendation scoring and ranking subunit, a risk warning subunit, and a visualization dashboard. The automatic bidding information classification subunit is used to classify bidding announcements into three categories: recommended, pending review, or filtered, based on the comparison results of the corrected confidence level and the three-classification threshold. The recommendation scoring and ranking subunit is used to calculate the ranking score by comprehensively considering the confidence level value, commercial priority weight, and the freshness factor of the announcement release time, and arrange them in descending order of score. The risk warning subunit is used to generate a warning notification when it is detected that the bid deadline of the bidding announcement is less than the preset emergency threshold or the budget amount exceeds the preset high threshold. The visualization dashboard is used to display key indicators such as the number of bidding information processed daily, the model accuracy curve, and the trend of manual review rate to the operations personnel, and to receive manual correction feedback from the operations personnel on the labels of the screening results.
[0051] The continuous learning layer is configured with a misclassified sample collection subunit, an active learning sampling subunit, and a model periodic retraining subunit. The misclassified sample collection subunit is used to automatically capture manually corrected label samples and record the original prediction confidence, original prediction category, and the manually corrected true label of the sample. The active learning sampling subunit is used to select the most valuable labeled samples from the misclassified sample pool and add them to the training data pool using a hybrid sampling strategy that combines uncertainty sampling and diversity sampling. The model periodic retraining subunit is used to pull samples from the training data pool at a preset period or when data distribution drift is detected to perform full or incremental retraining on the ensemble tree model.
[0052] In this embodiment, the bidding information screening method is constructed as an intelligent agent with a nine-layer decoupled architecture. The functions of data collection, preprocessing, feature engineering, dual-path reasoning, decision fusion, result output, and continuous learning are modularly deployed in layers. Each layer collaborates with the other through a preset data transmission interface and a reverse feedback link. This enables the intelligent agent to automatically trigger model retraining and knowledge base updates through the continuous learning layer when faced with changes in the distribution of bidding announcement data, thereby achieving long-term autonomous improvement in screening accuracy.
[0053] In some embodiments, such as Figure 3 As shown, the internal layers and cross-layer connections of the intelligent agent are as follows: In the data acquisition layer, the outputs of the website tag crawling tool and the incremental synchronization mechanism are input to the data deduplication engine, which then outputs the deduplicated raw data to the data preprocessing layer.
[0054] In the data preprocessing layer, the raw data is input to the text cleaning subunit, the cleaned text is passed to the structure extraction subunit, the output of the structure extraction subunit is passed to the normalization subunit and the anomaly detection subunit for parallel processing, and the outputs of the normalization subunit and the anomaly detection subunit are converged and output to the feature engineering layer.
[0055] In the feature engineering layer, structured field data in the input data is passed to the label system construction subunit, and unstructured text data in the input data is passed to the word embedding subunit. The outputs of the label system construction subunit and the word embedding subunit are input to the feature cross-selection subunit. The fused feature vector output by the feature cross-selection subunit is simultaneously passed to the machine learning engine layer and the large model semantic understanding layer, forming a dual-path parallel inference architecture.
[0056] In the machine learning engine layer, the fused feature vectors are simultaneously input into multiple parallel ensemble tree-based models. The outputs of each ensemble tree-based model converge to the model fusion sub-unit, which then outputs the generated first confidence score to the decision fusion layer.
[0057] In the semantic understanding layer of the large model, the RAG retrieval enhancement subunit calls the vector database in the data storage layer to retrieve historical cases. The retrieval results, along with the current text data and the reference examples selected by the Few-shot learning subunit, are input into the Prompt engineering subunit to generate input prompts. After the input prompts are inferred by the large language model, a second confidence score is generated, and the second confidence score is output to the decision fusion layer.
[0058] In the decision fusion layer, the dual-confidence weighted fusion subunit receives the first confidence level and the second confidence level and performs evidence fusion. The fusion result is passed to the business rule engine for forced correction. The threshold adaptive adjustment subunit dynamically adjusts the decision threshold of the business rule engine based on historical feedback data. The corrected final confidence level is output to the output and application layers.
[0059] In the output and application layer, the received final confidence level is processed sequentially by the automatic classification subunit and the recommendation scoring and ranking subunit. The processing results, together with the risk features output by the risk warning subunit, are displayed on the visualization dashboard. The user feedback data received by the visualization dashboard is output to the continuous learning layer.
[0060] In the continuous learning layer, manually fed-in data is input to the misjudged sample recovery subunit. After the active learning sampling subunit selects high-value training samples, it triggers the model periodic retraining subunit to perform retraining. After retraining is completed, through the reverse feedback link, it outputs model weight update signals to the online learning subunit and model fusion subunit of the machine learning engine layer, outputs vector library update signals to the RAG retrieval enhancement subunit of the large model semantic understanding layer, outputs threshold adjustment signals to the threshold adaptive adjustment subunit of the decision fusion layer, and outputs feature weight optimization signals to the feature cross selection subunit of the feature engineering layer, forming a closed-loop evolution mechanism.
[0061] Specifically, such as Figure 3 As shown, the internal layers and cross-layer connections of this intelligent agent are as follows: The internal connection relationships of the data acquisition layer are as follows: The website bidding information crawling tool initiates page crawling requests to the target bidding information publishing website according to the preset Uniform Resource Locator (URL) scheduling list, obtains the original HTML page data of the bidding announcement list page and detail page, and outputs the crawled page data along with the crawling timestamp and source website identifier; The incremental synchronization mechanism maintains a local cache set of unique identifiers of the crawled announcements. After each round of crawling, it performs a difference operation between the set of unique identifiers of the announcements crawled in this round and the local cache set to identify newly released bidding announcements, and outputs the list of new announcement identifiers; The output of the website bidding information crawling tool and the output of the incremental synchronization mechanism are jointly input to the data deduplication engine. The data deduplication engine extracts the URL hash value and text content fingerprint of each input data and matches them with the stored fingerprint database. The original bidding announcement data that is determined to be non-duplicate is output to the data preprocessing layer through the inter-layer data transfer interface.
[0062] The internal connections of the data preprocessing layer are as follows: Raw tender announcement data received from the data acquisition layer is input to the text cleaning subunit. The text cleaning subunit uses a combination of regular expression matching and DOM tree parsing to remove HTML tags, CSS styles, JavaScript code, and non-text elements from the original HTML page data, outputting only plain text data containing the main text of the tender announcement. The cleaned plain text data is then passed to the structured extraction subunit. This subunit loads a named entity recognition model to perform sequence labeling on the plain text data to identify organizational entities such as the procuring entity name and agency name. Simultaneously, it calls a regular expression template library to perform pattern matching and numerical extraction on key fields such as budget amount, bid deadline, and tender document price, outputting data containing the tender announcement's main text. The structured data of key-value pairs is extracted. The output of the structure extraction subunit is passed to the normalization subunit and the anomaly detection subunit for parallel processing. The normalization subunit performs currency unit unification and robust scaling on the budget amount field to eliminate the difference in the units of the amount values between different announcements. At the same time, the date field is converted to a standard format. The anomaly detection subunit loads a pre-trained isolated forest model to calculate the average path length of each data record on multiple random partitioning trees as an anomaly score and marks data records with anomaly scores exceeding a preset threshold as anomalies. The normalized data output by the normalization subunit and the anomaly-marked data output by the anomaly detection subunit are aggregated. The data records marked as anomalies are removed from the aggregated data stream, and the remaining normalized data is output to the feature engineering layer through the inter-layer data transfer interface.
[0063] The internal connections of the feature engineering layer are as follows: From the standardized data received from the data preprocessing layer, structured field data is passed to the tag system construction subunit, and unstructured text data is passed to the word embedding subunit. The tag system construction subunit reads the purchasing unit name field and calls the enterprise information query interface to obtain supplementary dimension information such as the national standard industry classification code, whether it is a specific type of enterprise attribute identifier, the year of establishment of the enterprise, and historical cooperation records. This supplementary dimension information is then merged with the original fields in the standardized data, such as budget amount, personnel size, and qualification requirements, to construct a structured tag vector containing nine dimensions: national standard industry, industry classification related to the target enterprise, customer personnel size level, whether it is a specific type of enterprise attribute, year of establishment range, cooperation status level, normalized value of the procurement budget, procurement demand personnel level, and bidding qualification requirement category. The word embedding subunit loads the Transformer encoder network of the pre-trained BERT model, and segments and positions the main text of the bidding announcement. After encoding, the text semantic vector is input into the encoder network, and the sentence-level semantic vector output from the last layer of the encoder is taken as the text semantic vector. The nine-dimensional structured label vector output from the label system construction subunit and the text semantic vector output from the word embedding subunit are input into the feature cross selection subunit. The feature cross selection subunit concatenates the feature vectors from the two sources and performs second-order and third-order feature cross operations to generate combined features containing interactive information of structured features and semantic features. The XGBoost model is used to calculate the information gain value of each cross feature dimension on historical training data, and then selects several features with a cumulative contribution rate reaching a preset ratio according to the gain value from high to low as the filtered fusion feature vector. The fusion feature vector output by the feature cross selection subunit is simultaneously passed to the machine learning engine layer and the large model semantic understanding layer, forming a dual-path parallel inference architecture. That is, the same fusion feature vector is copied into two copies and sent to the two parallel inference layers below through independent inter-layer data transfer interfaces.
[0064] The internal connections of the machine learning engine layer are as follows: the fused feature vector received from the feature engineering layer is simultaneously input into three parallel ensemble tree-based models: the XGBoost classifier, the Random Forest classifier, and the LightGBM classifier. The XGBoost classifier performs tree-by-tree inference on the input feature vector based on its internally trained gradient boosting decision tree set, summing the prediction scores of each leaf node and mapping them to a first prediction probability value using the Sigmoid function. The Random Forest classifier performs parallel voting inference on the input feature vector based on its internally trained multiple Bootstrap sampling decision trees, using the voting ratio as a second prediction probability value. The LightGBM classifier, based on its... The internally trained histogram-accelerated decision tree ensemble performs tree-by-tree inference on the input feature vector and accumulates the prediction scores of each leaf node to obtain the third prediction probability value. The first, second, and third prediction probability values output by the three ensemble tree-based models are converged to the model fusion subunit. The model fusion subunit concatenates the three prediction probability values into a three-dimensional feature vector and inputs it into the trained secondary logistic regression model. The secondary logistic regression model uses the learned weight coefficients to perform a weighted linear combination of the three-dimensional feature vector and maps it to the first confidence score through the Sigmoid function. The model fusion subunit outputs the generated first confidence score to the decision fusion layer through the inter-layer data transfer interface.
[0065] The internal connections of the semantic understanding layer of the large model are as follows: The RAG retrieval enhancement subunit receives text semantic vectors as query vectors from the feature engineering layer, calls the similarity retrieval interface of the vector database in the data storage layer, and calculates the cosine similarity between the query vector and the text semantic vectors of each historical bidding announcement case stored in the vector database. The vector database returns a set of candidate historical cases with the highest cosine similarity ranking, along with the corresponding case text descriptions and screening conclusion labels. The retrieved set of candidate historical cases, together with the text semantic data and structured field summary of the current bidding announcement, are input into the Few-shot learning subunit. The Few-shot learning subunit selects the top K cases from high to low semantic similarity between the candidate historical cases and the current bidding announcement, and organizes the text descriptions and screening conclusions of each case according to a preset example format template. The formatted reference example text is then output. The retrieval results of the RAG retrieval enhancement subunit and the reference example selected by the Few-shot learning subunit are input into the Prompt engineering subunit. The Prompt engineering subunit concatenates the prompt template that constrains the output format of the large language model, the reference example text, and the text semantic data and structured field summary of the current tender announcement in a preset order to form a complete model input prompt. This model input prompt is input into the large language model. The large language model performs autoregressive generation token by token in its Transformer decoder network and outputs an inference result string that conforms to the JSON format constraints in the prompt template and contains the matching conclusion field and the confidence score field. The second confidence score is parsed from this inference result string and output to the decision fusion layer through the inter-layer data transfer interface.
[0066] The internal connections of the decision fusion layer are as follows: The dual-path confidence-weighted fusion subunit simultaneously receives the first confidence score output from the machine learning engine layer and the second confidence score output from the large model semantic understanding layer. The dual-path confidence-weighted fusion subunit constructs basic probability assignment functions for the first and second inference paths respectively, and performs orthogonal summation on the two basic probability assignment functions based on the Dempster combination rule of Dempster-Shafer evidence theory to generate the fused probability assignment. A comprehensive confidence score is extracted from the fused probability assignment. The comprehensive confidence score output by the dual-path confidence-weighted fusion subunit is passed to the business rule engine, which then combines the comprehensive confidence score with the final... The system performs joint judgment on key fields in structured business data. When a key field matches a preset hard filtering rule, the overall confidence score is forcibly corrected to the corresponding hard value. When a key field matches a preset weighted adjustment rule, the overall confidence score is multiplied by the corresponding adjustment weight coefficient. The threshold adaptive adjustment subunit reads manual feedback data records within a preset statistical period from the relational database in the data storage layer and calculates the confirmation accuracy of the recommendation category and the filtering category. Based on the trend of the confirmation accuracy, it dynamically raises or lowers the upper and lower thresholds used for three-class classification in the business rule engine. The final confidence score corrected by the business rule engine is output to the output and application layers through the inter-layer data transfer interface.
[0067] The connection between the output and the application layer is as follows: The final confidence level input received from the decision fusion layer is fed into the automatic bidding information classification subunit. The automatic bidding information classification subunit compares the final confidence level with the upper and lower thresholds and outputs a classification label. When the final confidence level is greater than or equal to the upper threshold, a recommendation label is output; when the final confidence level is less than the upper threshold but greater than the lower threshold, a pending review label is output; when the final confidence level is less than or equal to the lower threshold, a filter label is output. The classification labels and corresponding final confidence level values output by the automatic bidding information classification subunit are passed to the recommendation scoring and ranking subunit. The recommendation scoring and ranking subunit reads the commercial priority weight coefficient and the announcement publication timestamp of each bidding announcement and calculates the time interval between the current time and the publication timestamp. It then combines the final confidence level, commercial priority weight coefficient, and time decay factor... After weighted summation, a ranking score is generated and sorted in descending order of score. The ranking results output by the recommendation scoring subunit and the risk characteristics output by the risk warning subunit are displayed together on the visualization dashboard. The risk warning subunit extracts the bid deadline field from the structured fields of the bidding announcement and calculates the difference between it and the current time. When the difference is less than the preset emergency threshold, an emergency warning mark is generated. At the same time, the budget amount field is extracted from the structured fields, and a high-amount warning mark is generated when the budget amount exceeds the preset high-amount threshold. The visualization dashboard displays the classification results, ranking position, and warning marks of each bidding announcement in a graphical interface. It also receives correction operation data for the classification results submitted by the operations personnel through the interface interaction and outputs the correction operation data as human feedback data to the continuous learning layer through the inter-layer data transfer interface.
[0068] The internal connections of the continuous learning layer are as follows: Human feedback data received from the output and application layers is input to the misclassified sample recovery subunit. This subunit extracts the unique identifier of the corrected label from the human feedback data and, based on this identifier, reads the original fused feature vector, original first confidence level, original second confidence level, original comprehensive confidence level, and original classification label from the data storage layer. The read original data is then combined with the manually corrected true label to form a complete misclassified sample record, which is then stored in the misclassified sample pool. The misclassified sample records in the misclassified sample recovery subunit are output to the active learning sampling subunit, where active learning... The sampling subunit calculates the uncertainty score and diversity score for each sample in the misclassified sample pool. The uncertainty score is calculated by taking the absolute value of the difference between the original comprehensive confidence score of the sample and 0.5, and then taking its negative value so that samples with confidence scores close to 0.5 receive higher uncertainty scores. The diversity score is calculated by averaging the cosine distance between the fused feature vector of the sample and the feature vectors of the selected sample set. The uncertainty score and diversity score are weighted according to preset weights to obtain a comprehensive sampling score, and a preset number of samples are selected as high-value training samples from high to low scores. The active learning sampling subunit... The selected high-value training samples trigger the model periodic retraining subunit to perform retraining. This subunit merges the high-value training samples with historical training data to form an expanded training set, and then performs incremental iterative training on the three base models: XGBoost classifier, Random Forest classifier, and LightGBM classifier, updating the splitting thresholds of decision tree nodes and the weights of leaf nodes in each model. After retraining, the model periodic retraining subunit outputs model weight update signals to the online learning subunit and model fusion subunit of the machine learning engine layer via a backfeedback link to synchronize the retrained new model parameters to the online inference process. The system outputs vector database update signals to the RAG retrieval enhancement subunit of the large model semantic understanding layer to write the text semantic vectors and screening conclusions of the target cases that have been manually confirmed as positive matches into the vector database. It outputs threshold adjustment signals to the threshold adaptive adjustment subunit of the decision fusion layer to recalibrate the upper and lower thresholds according to the confidence distribution of the retrained model on the validation set. It outputs feature weight optimization signals to the feature cross selection subunit of the feature engineering layer to update the priority ranking of feature selection according to the feature importance score output by the retrained model. This forms a complete closed-loop evolution mechanism that drives each inference layer and decision layer from the bottom continuous learning layer upwards.
[0069] In this embodiment, by defining in detail the data input relationships, processing and transmission order, and output convergence methods among the sub-units within each layer of the nine-layer intelligent agent architecture, as well as the forward data transmission paths and reverse feedback driving paths between layers, the data processing flow within the intelligent agent is formed from the collection, cleaning, and feature construction of the original bidding announcement data, to dual-path parallel reasoning, evidence fusion decision-making, result classification and display, and then to the closed-loop feedback of manual feedback sampling and model retraining. This forms a complete and rigorous end-to-end intelligent screening link, with each sub-unit operating collaboratively according to the preset connection relationship and the data flow being clear and controllable. This provides a structured architectural foundation for achieving high accuracy and continuous self-evolution in bidding information screening.
[0070] In one embodiment, such as Figure 4 As shown, in step S10, the bidding announcement data is obtained, and the text content of the bidding announcement data is extracted and the multi-dimensional structured fields are parsed to obtain structured business data and text semantic data. This process is executed sequentially by the data acquisition layer, data preprocessing layer and feature engineering layer of the intelligent agent, and specifically includes the following steps: S11-S13.
[0071] S11. The data acquisition layer uses a website bidding information crawling tool to periodically crawl the original bidding announcement page data of external bidding websites, and performs deduplication processing based on the page content fingerprint through a data deduplication engine.
[0072] The website bidding information crawling tool initiates HTTP requests to the target bidding information publishing website according to a preset crawling schedule and obtains the HTML page data of the bidding announcement list page. After parsing the Uniform Resource Locator (URL) of each bidding announcement's details page from the list page, it sequentially initiates requests to each details page and downloads the complete HTML page data of the details page. The incremental synchronization mechanism compares the unique identifier set of the already crawled announcements with the current crawling results to identify newly released bidding announcements. The data deduplication engine extracts the URL hash value and text content fingerprint of each newly crawled page data. It matches the URL hash value with the stored crawling records and compares the text content fingerprint with the stored content fingerprint database. When the URL hash value already exists or the similarity between the text content fingerprint and an existing fingerprint exceeds a preset duplication threshold, the data is marked as duplicate data and discarded. The original bidding announcement page data that has been determined to be non-duplicate after deduplication is stored in the raw database of the data storage layer and passed down to the data preprocessing layer.
[0073] S12. The data preprocessing layer receives the original bidding announcement page data after deduplication, and sequentially calls the text cleaning subunit to remove HTML tags and noisy text, the structured extraction subunit to extract key fields based on named entity recognition, the normalization subunit to perform robust scaling processing on numerical fields, and the anomaly detection subunit to remove abnormal data based on the isolated forest algorithm, to obtain standardized data.
[0074] The text cleaning subunit uses rule-based regular expression matching to systematically remove HTML tags, CSS style code, JavaScript code, and non-text content such as page navigation bars, footers, and sidebar advertisements from the original HTML page data, retaining only the main text area containing the core information of the bidding announcement. It also performs unified conversion and compression on full-width characters, traditional Chinese characters, and redundant whitespace in the main text, resulting in a well-formatted plain text bidding announcement. The structured extraction subunit loads a pre-trained named entity recognition model to perform sequence labeling on the plain text, identifying organizational entities such as the procuring entity name and agency name. Simultaneously, it uses regular expression templates to match numerical and time-related fields such as budget amount, bid deadline, bid opening time, and bid price, categorizing and storing the identified entities and fields according to preset field names. The normalization subunit removes currency unit characters from the extracted budget amount field and converts it to a floating-point value in ten thousand yuan. For the customer personnel size field, it divides it into six tiers according to preset binning boundary values and replaces the original value with the tier number. The date field is uniformly converted to a standard date format. After normalization, the values of all fields fall within a unified dimensional range. The anomaly detection subunit loads a pre-trained isolated forest model, using each normalized data record as an input sample. The isolated forest model constructs multiple isolated trees by randomly selecting features and split points, and calculates the average path length of each sample across these trees as an anomaly score. Data records with anomaly scores exceeding a preset threshold are marked as anomalous and removed from the data stream. The retained data records are passed to the feature engineering layer as standardized data.
[0075] S13. The feature engineering layer receives standardized data, calls the label system construction sub-unit to construct nine-dimensional structured labels that at least cover industry, personnel size, whether it is a specific type of enterprise, and procurement budget, calls the word embedding sub-unit to perform semantic encoding on the text content to obtain text semantic vectors, calls the feature cross selection sub-unit to perform feature cross combination of the nine-dimensional structured labels and text semantic vectors and filter based on model gain, and outputs structured business data and text semantic data.
[0076] The tagging system construction subunit reads the values of each field in the standardized data. It inputs the name of the procuring entity into the enterprise information query interface to obtain supplementary fields such as the national standard industry classification code, whether it is a specific type of enterprise, the number of years the enterprise has been established, and whether it has a history of cooperation with the company. These supplementary fields are then merged with the original fields in the standardized data to form a nine-dimensional structured tag. The national standard industry classification code is one-hot encoded, the "whether it is a specific type of enterprise" attribute is binary encoded, the number of years established is binned according to a preset range, and the historical cooperation status is sequentially encoded into five levels: never cooperated, had intentions, participated in bidding, won a bid, and is currently cooperating. The word embedding subunit loads a pre-trained BERT model. The plain text of the tender announcement is segmented and input into the Transformer encoder network of the BERT model. The vector corresponding to the text's start marker position in the hidden state vector output from the last layer of the encoder network is taken as the sentence-level semantic representation of the tender announcement text, or the weighted average of the word vectors in the text is taken as the text semantic vector. The dimension of the text semantic vector is a preset fixed dimension. The feature cross-selection subunit concatenates the encoding vector of the nine-dimensional structured label with the text semantic vector, and performs feature cross-operation on the concatenated vector to generate second-order and third-order cross-combination features containing structured and semantic features. Using the XGBoost model, the information gain value of each feature dimension is calculated on historical training data. After sorting the gain values from high to low, the top several features with the cumulative contribution rate of the gain values reaching the preset ratio are selected as the filtered feature set. The filtered feature set is output as structured business data and text semantic data.
[0077] In this embodiment, the timeliness and non-redundancy of input data are ensured by the timed capture and content fingerprint deduplication mechanism of the data acquisition layer. The unstructured raw HTML pages are gradually transformed into standardized data through text cleaning, structured extraction, normalization processing and isolated forest anomaly detection in the data preprocessing layer. The structured business attributes and deep text semantics are organically integrated through the nine-dimensional label construction, BERT semantic encoding and gain-based feature cross-selection in the feature engineering layer. This makes the feature data input to the subsequent dual-path inference stage have both accurate business field expression ability and rich semantic representation ability, providing a high-quality feature foundation for improving the accuracy of target information selection.
[0078] In one embodiment, such as Figure 5 As shown, in step S20, the step of generating the first confidence score by inputting structured business data into the ensemble tree model of the first inference path is executed by the machine learning engine layer of the agent, specifically including the following steps: S21-S22.
[0079] S21. Structured business data is input in parallel into the three integrated tree-based models of the machine learning engine layer: XGBoost classifier, random forest classifier, and LightGBM classifier.
[0080] The XGBoost classifier performs forward inference on the input structured business data with a hyperparameter configuration of 500 trees, a maximum tree depth of 8, and a learning rate of 0.05. Internally, the XGBoost classifier constructs classification and regression trees one by one using a gradient boosting framework. During training, each new tree is fitted with the residual gradient of the preceding tree set. During inference, after the input sample reaches a leaf node through the decision path of each tree, the prediction scores of each leaf node are accumulated and mapped to a probability value using the Sigmoid function. The Random Forest classifier performs inference on the input data with a hyperparameter configuration of 300 trees and a maximum tree depth of 12. Internally, the Random Forest classifier uses the Bootstrap method to sample multiple sub-training sets with replacement from the training set and trains a classification decision tree independently on each sub-training set. During inference, the input sample simultaneously passes through all decision trees for voting, and the proportion of votes is used as the probability value of the sample belonging to the target. The LightGBM classifier performs inference on input data with a hyperparameter configuration of 600 trees and 63 leaf nodes. It employs a histogram-based decision tree algorithm to discretize continuous features into multiple bins and find the optimal split point within each bin. A leaf node-first growth strategy is used to construct the tree structure. During inference, the input sample traverses the paths of each tree to reach a leaf node, outputting the corresponding predicted score, which is then transformed to obtain a probability value. The three ensemble tree-based models operate independently during inference, generating three independent predicted probability values.
[0081] S22. Using the model fusion subunit, the predicted probability values output by the three integrated tree-based models are stacked and fused to generate the first confidence level.
[0082] The model fusion subunit maintains a secondary logistic regression model. During the training phase, the secondary logistic regression model is trained using the predicted probability values of the three base models on the validation set as input features and the true labels of the validation set samples as the target variable, learning the optimal weight coefficients assigned to the predicted values of each base model. During the inference phase, the model fusion subunit concatenates the first predicted probability value output by the XGBoost classifier, the second predicted probability value output by the Random Forest classifier, and the third predicted probability value output by the LightGBM classifier into a three-dimensional feature vector. This three-dimensional feature vector is input into the trained secondary logistic regression model. The secondary logistic regression model uses the learned weight coefficients to perform a weighted linear combination of the three-dimensional feature vector and maps the combination result to a probability value between 0 and 1 using the Sigmoid function, outputting the first confidence level that the tender notice belongs to the target tender information. The online learning subunit is equipped with the FTRL algorithm. When the continuous learning layer outputs the model weight update signal to the machine learning engine layer through the back feedback link, the online learning subunit reads the manually corrected sample data and the corresponding real labels, and uses the FTRL algorithm to incrementally update the weights of the leaf nodes of the three ensemble tree-based models. During the update process, the FTRL algorithm introduces L1 regularization and L2 regularization terms to impose sparsity constraints on the weights in order to control the model complexity and prevent overfitting.
[0083] In this embodiment, structured business data is simultaneously input into three integrated tree-based models—XGBoost classifier, Random Forest classifier, and LightGBM classifier—each with its own advantages in model structure and learning strategy, for parallel inference. A Stacking fusion strategy based on a secondary logistic regression model is used to weight and fuse the prediction results of the three base models to generate a first confidence score. This allows the first inference path to comprehensively utilize three different modeling mechanisms—gradient boosting, self-help aggregation, and histogram acceleration—to capture nonlinear relationships in structured features from multiple perspectives. Furthermore, an online learning subunit configured with the FTRL algorithm provides the model with the ability to adjust inference weights in real time after receiving human feedback, effectively improving the classification accuracy of the first inference path for structured business data and its adaptability to changes in data distribution.
[0084] In one embodiment, such as Figure 6 As shown, step S20, which involves retrieving historical cases similar to the text semantic data from a preset vectorized domain knowledge base, concatenating the text semantic data with the historical cases, and inputting the concatenated data into the large language model to generate the second confidence score, is executed by the large model semantic understanding layer of the agent. Specifically, it includes the following steps: S23. The RAG retrieval enhancement subunit transforms the text semantic data into query vectors, performs a two-stage retrieval combining BM25 coarse recall and vector cosine fine ranking in the vector database of the data storage layer, and re-ranks the recall results by maximum marginal relevance to obtain similar historical cases.
[0085] The RAG retrieval enhancement subunit uses the text semantic vector output from the feature engineering layer as the query vector. This text semantic vector is a fixed-dimensional dense vector output by the BERT model encoder after semantically encoding the main text of the tender announcement. Each dimension of the query vector represents the activation intensity of the text in the corresponding semantic dimension. The first stage of retrieval is the BM25 coarse recall stage. After segmenting the main text of the tender announcement into words, the RAG retrieval enhancement subunit calculates the BM25 weight of each word on the document frequency table. A sparse bag-of-words query vector is constructed using the BM25 weights. BM25 scores are calculated for the text fields of all historical cases in the vector database. Based on the BM25 scores, the top 50 candidate historical cases with the highest scores are recalled from high to low. The second stage of retrieval is the vector cosine similarity fine-tuning stage. The RAG retrieval enhancement subunit calculates the cosine similarity between the query vector and the text semantic vectors corresponding to the 50 candidate historical cases recalled in the first stage. The cosine similarity is calculated by dividing the dot product of the two vectors by the product of their magnitudes. The candidate historical cases are sorted from high to low based on the cosine similarity, and the top 10 candidate historical cases are selected. After fine-tuning, the RAG retrieval enhancement subunit performs maximum marginal relevance re-sorting on the top 10 candidate historical cases. Maximum marginal relevance re-sorting calculates the comprehensive score by subtracting the similarity between the case and the query vector from the maximum similarity between the case and the cases in the selected case set multiplied by a preset diversity penalty coefficient when selecting the next case in each round. Cases are selected sequentially according to the comprehensive score until a preset selection number K is reached. The K historical cases after re-sorting are considered similar historical cases.
[0086] S24. The Few-shot learning subunit selects the K samples with the highest semantic similarity to the current tender announcement from similar historical cases as inference reference examples for the large language model. The Prompt engineering subunit provides a prompt template for the constraint output format. The text descriptions and corresponding screening conclusions of the K samples, the text semantic data of the current tender announcement, and the prompt template are concatenated and input into the large language model. The large language model outputs the inference results containing the conclusions and confidence scores in a structured format. The second confidence score is extracted from the inference results.
[0087] The Few-shot learning subunit reads the K historical cases obtained after reordering and organizes the tender announcement text summary, structured field summary, manually annotated screening conclusion labels, and confidence scores of each historical case according to a preset case display format. The Prompt engineering subunit provides a prompt template that constrains the output format. The prompt template includes a description of the role setting of the large language model, an inference task description, output format constraints, and a structured output framework with empty slots. The output format constraints require the large language model to output in JSON format, and the JSON must include the matching conclusion field, confidence score field, and inference reason field, without outputting other irrelevant content. The formatted text of the K historical cases, the original tender announcement text summary and structured field summary corresponding to the text semantic data of the current tender announcement, and the prompt template are concatenated in the order of case first, current input in the middle, and prompt template last. The concatenated complete text is used as the input prompt for the large language model. After receiving input prompts, the large language model performs autoregressive generation token by token in its Transformer decoder network. During the generation process, it uses the matching patterns of K historical cases provided in the input prompts as inference references to perform semantic inference on the degree of matching between the current tender announcement and the target business, and generates an inference result string that conforms to the JSON format constraints in the prompt template. The inference result string is then parsed using JSON to extract the value of the confidence score field, and the extracted value is the second confidence score.
[0088] In this embodiment, the RAG retrieval enhancement subunit uses textual semantic data as query vectors to perform a two-stage cascaded retrieval combining BM25 sparse retrieval and vector cosine dense retrieval in the vector database. The recall results are then reordered based on maximum marginal relevance to balance retrieval recall and result diversity. The Few-shot learning subunit organizes similar historical cases into structured reference examples and, in conjunction with the format constraint template of the Prompt engineering subunit, guides the large language model to perform inference in a controlled structured output format. This allows the second inference path to fully utilize the domain experience knowledge contained in the historical annotated cases to assist in the semantic understanding of the current tender announcement. This improves the accuracy of deep semantic matching evaluation while avoiding the format uncertainty and inference illusion caused by the free generation of the large language model.
[0089] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.
[0090] In one embodiment, a tender information filtering system based on intelligent agents is provided, which corresponds one-to-one with the tender information filtering method based on intelligent agents described in the above embodiments. For example... Figure 7 As shown, the agent-based bidding information screening system 700 includes: a data acquisition module 701, a dual-path inference module 702, a decision fusion module 703, and a rule output module 704. Detailed descriptions of each functional module are as follows: The data acquisition module 701 is used to acquire bidding announcement data, extract text content from the bidding announcement data and parse multi-dimensional structured fields to obtain structured business data and text semantic data. The dual-path inference module 702 includes a first inference unit and a second inference unit. The first inference unit is used to input structured business data into an integrated tree model to generate a first confidence level. The second inference unit is used to retrieve historical cases similar to the text semantic data from a preset vectorized domain knowledge base, and then input the text semantic data and historical cases into a large language model to generate a second confidence level. The decision fusion module 703 is used to calculate the conflict coefficient between the first confidence level and the second confidence level. Only when the conflict coefficient is lower than the preset conflict threshold, the first confidence level and the second confidence level are fused according to evidence theory to obtain the comprehensive confidence level. The rule output module 704 is used to forcibly correct the overall confidence level based on pre-configured business rules, and generate filtering result labels for the bidding announcement data based on the corrected confidence level. The filtering result labels include at least one of recommended, pending review, or filtered.
[0091] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.
[0092] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.
[0093] This application provides an intelligent agent for executing a tender information filtering method based on any one of the embodiments of this application. The intelligent agent is installed on... Figure 8 The computer device shown includes: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor. When the processor executes the computer program, it implements the steps in any of the above method embodiments, or when the processor executes the computer program, it implements the functions of each module / unit in the above device embodiments.
[0094] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete this application. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the computer device.
[0095] Those skilled in the art will understand that Figure 8 The computer device described is merely an example and does not constitute a limitation on the computer device. It may include more or fewer components than shown, or combine certain components, or different components. For example, the computer device may also include input / output devices, network access devices, buses, etc.
[0096] The aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.
[0097] The memory can be an internal storage unit of the computer device, such as a hard drive or RAM. The memory can also be an external storage device of the computer device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory can include both internal and external storage units of the computer device.
[0098] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.
[0099] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0100] In the embodiments provided in this application, it should be understood that the disclosed apparatus / devices and methods can be implemented in other ways. For example, the apparatus / device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.
[0101] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0102] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.
Claims
1. An agent-based method for filtering tender information, characterized in that, The method is executed by an intelligent agent and includes: Obtain bidding announcement data, extract text content and parse structured fields from the bidding announcement data to obtain structured business data and text semantic data; The structured business data is input into the ensemble tree model of the first inference path to generate a first confidence score; at the same time, historical cases similar to the text semantic data are retrieved from the preset vectorized domain knowledge base, and the text semantic data and the historical cases are concatenated and input into the large language model to generate a second confidence score. Calculate the conflict coefficient between the first confidence level and the second confidence level. Only when the conflict coefficient is lower than a preset conflict threshold, perform evidence theory fusion on the first confidence level and the second confidence level to obtain a comprehensive confidence level. The comprehensive confidence level is forcibly corrected based on pre-configured business rules, and a filtering result label for the bidding announcement data is generated based on the corrected confidence level. The filtering result label includes at least one of: recommended, pending review, or filtered. The intelligent agent adopts a top-down nine-layer decoupled architecture, including: The data storage layer includes a raw database, a feature library, a model library, a vector database, and a relational database, which provides vertical data reading and persistence support for other layers. The data acquisition layer is equipped with website tag crawling tools, an incremental synchronization mechanism, and a data deduplication engine; The data preprocessing layer is configured with text cleaning subunits, structure extraction subunits, normalization subunits, and anomaly detection subunits; The feature engineering layer is configured with a label system construction subunit, a word embedding subunit, and a feature cross selection subunit; The machine learning engine layer is configured with multiple integrated tree-based models, model fusion sub-units, and online learning sub-units; The large model semantic understanding layer is configured with a RAG retrieval enhancement subunit, a Prompt engineering subunit, a Few-shot learning subunit, and a semantic similarity calculation unit, and retrieves the historical cases by interacting with the vector database in the data storage layer. The decision fusion layer is configured with a dual-path confidence-weighted fusion subunit, a business rule engine, and a threshold adaptive adjustment subunit. The output and application layer is configured with an automatic classification sub-unit for bidding information, a recommendation scoring and ranking sub-unit, a risk warning sub-unit, and a visual dashboard. The continuous learning layer is configured with a misjudged sample recovery subunit, an active learning sampling subunit, and a model periodic retraining subunit.
2. The agent-based bid information screening method of claim 1, wherein, The internal and cross-layer connections of the intelligent agent are as follows: In the data acquisition layer, the outputs of the website tag crawling tool and the incremental synchronization mechanism are jointly input to the data deduplication engine, and the data deduplication engine outputs the deduplicated raw data to the data preprocessing layer. In the data preprocessing layer, the raw data is input to the text cleaning subunit, the cleaned text is passed to the structure extraction subunit, the output of the structure extraction subunit is passed to the normalization subunit and the anomaly detection subunit for parallel processing, and the outputs of the normalization subunit and the anomaly detection subunit are converged and output to the feature engineering layer. In the feature engineering layer, the structured field data in the input data is passed to the label system construction subunit, and the unstructured text data in the input data is passed to the word embedding subunit. The outputs of the label system construction subunit and the word embedding subunit are jointly input to the feature cross selection subunit. The fused feature vector output by the feature cross selection subunit is simultaneously passed to the machine learning engine layer and the large model semantic understanding layer, forming a dual-path parallel inference architecture. In the machine learning engine layer, the fused feature vector is simultaneously input to multiple parallel ensemble tree-based models, and the outputs of each ensemble tree-based model converge to the model fusion subunit. The model fusion subunit outputs the generated first confidence score to the decision fusion layer. In the large model semantic understanding layer, the RAG retrieval enhancement subunit calls the vector database in the data storage layer to retrieve the historical cases. The retrieval results, along with the current text data and the reference examples selected by the Few-shot learning subunit, are input to the Prompt engineering subunit to generate input prompts. The input prompts are inferred by the large language model to generate the second confidence score, and the second confidence score is output to the decision fusion layer. In the decision fusion layer, the dual-confidence weighted fusion subunit receives the first confidence level and the second confidence level and performs evidence fusion. The fusion result is passed to the business rule engine for forced correction. The threshold adaptive adjustment subunit dynamically adjusts the decision threshold of the business rule engine based on historical feedback data. The corrected final confidence level is output to the output and application layer. In the output and application layer, the received final confidence level is processed sequentially by the automatic classification subunit and the recommendation scoring and ranking subunit. The processing results, together with the risk features output by the risk warning subunit, are displayed on the visualization dashboard. The user manual feedback data received by the visualization dashboard is output to the continuous learning layer. In the continuous learning layer, the manual feedback data is input to the misjudged sample recovery subunit. After the active learning sampling subunit selects high-value training samples, it triggers the model periodic retraining subunit to perform retraining. After retraining, through the reverse feedback link, it outputs model weight update signals to the online learning subunit and the model fusion subunit of the machine learning engine layer, vector library update signals to the RAG retrieval enhancement subunit of the large model semantic understanding layer, threshold adjustment signals to the threshold adaptive adjustment subunit of the decision fusion layer, and feature weight optimization signals to the feature cross-selection subunit of the feature engineering layer, forming a closed-loop evolution mechanism.
3. The agent-based bidding information screening method according to claim 2, characterized in that, The process of acquiring bidding announcement data, extracting text content and parsing multi-dimensional structured fields from the bidding announcement data to obtain structured business data and text semantic data is executed sequentially by the data acquisition layer, data preprocessing layer, and feature engineering layer, specifically including: The data acquisition layer uses the website bidding information crawling tool to periodically crawl the original bidding announcement data of external bidding websites, and uses the data deduplication engine to deduplicate based on content fingerprints; The data preprocessing layer receives the deduplicated data and sequentially calls the text cleaning subunit to remove HTML tags and noisy text, the structured extraction subunit to extract key fields based on named entity recognition, the normalization subunit to perform robust scaling processing on numerical fields, and the anomaly detection subunit to remove abnormal data based on the isolated forest algorithm to obtain standardized data. The feature engineering layer receives the standardized data, calls the tag system construction subunit to construct nine-dimensional structured tags that at least cover industry, personnel size, whether it is a specific type of enterprise, and procurement budget, calls the word embedding subunit to perform semantic encoding on the text content to obtain a text semantic vector, calls the feature cross selection subunit to perform feature cross combination of the nine-dimensional structured tags and the text semantic vector and filter based on model gain, and outputs the structured business data and text semantic data.
4. The agent-based bidding information screening method according to claim 3, characterized in that, In the feature engineering layer, the nine-dimensional structured tags include national standard industry, industry related to the target enterprise, customer personnel size, whether it is a specific type of enterprise, establishment year, cooperation status, procurement budget, procurement demand number of people, and bidding qualification requirements; the word embedding subunit uses TF-IDF or BERT model to semantically encode the bidding announcement text, and the customer personnel size is binned to a preset level, and the procurement budget is processed by robust scaling.
5. The agent-based bidding information screening method according to claim 2, characterized in that, The step of inputting the structured business data into the ensemble tree model of the first inference path to generate a first confidence score, executed by the machine learning engine layer, specifically includes: The structured business data is input in parallel into the XGBoost classifier, random forest, and LightGBM ensemble tree-based models of the machine learning engine layer; The outputs of the three base models are stacked and fused using the model fusion subunit to generate the first confidence score. The online learning subunit is equipped with the FTRL algorithm, which is used to perform incremental weight updates on the three integrated tree-based models when it receives the update signal from the back feedback link.
6. The agent-based bidding information screening method according to claim 2, characterized in that, The step of concatenating the text semantic data with the historical cases and inputting the result into the large language model to generate a second confidence score, which is executed by the semantic understanding layer of the large model, specifically includes: The RAG retrieval enhancement subunit transforms the text semantic data into query vectors, performs a two-stage retrieval combining BM25 coarse recall and vector cosine fine ranking in the vector database of the data storage layer, and re-ranks the recall results by maximum marginal relevance to obtain the similar historical cases. The Few-shot learning subunit selects the K most relevant samples from the similar historical cases; The Prompt subunit provides a prompt template for the constraint output format. The K samples, the text semantic data, and the prompt template are concatenated and input into the large language model. The large language model outputs the inference result containing the conclusion and confidence in a structured format, from which the second confidence is extracted.
7. The agent-based bidding information screening method according to claim 2, characterized in that, The decision fusion layer specifically performs the following operations: The dual-path confidence weighted fusion subunit constructs basic probability allocation functions for the machine learning engine layer and the large model semantic understanding layer, and calculates the conflict coefficient K based on Dempster-Shafer evidence theory. When K ≤ 0.8, the first confidence level and the second confidence level are fused using the Dempster combination rule to obtain the comprehensive confidence level; when K > 0.8, fusion is rejected and the corresponding bidding announcement data is directly marked as pending manual review. The business rule engine has built-in hard filtering rules and weighted adjustment rules to forcibly correct the labeling results after the comprehensive confidence or rejection fusion. The threshold adaptive adjustment subunit dynamically adjusts the upper and lower thresholds used for the three categories based on historical feedback data, where the three categories correspond to recommendation, pending review, and filtering.
8. A tender information screening system based on intelligent agents, characterized in that, The agent-based bidding information screening system is equipped with an agent, including: The data acquisition module is used to acquire bidding announcement data, extract text content from the bidding announcement data and parse multi-dimensional structured fields to obtain structured business data and text semantic data. The dual-path inference module includes a first inference unit and a second inference unit. The first inference unit is used to input the structured business data into an integrated tree model to generate a first confidence level. The second inference unit is used to retrieve historical cases similar to the text semantic data from a preset vectorized domain knowledge base, and then concatenate the text semantic data with the historical cases and input them into a large language model to generate a second confidence level. The decision fusion module is used to calculate the conflict coefficient between the first confidence level and the second confidence level, and only when the conflict coefficient is lower than a preset conflict threshold, the first confidence level and the second confidence level are fused according to evidence theory to obtain a comprehensive confidence level. The rule output module is used to forcibly correct the comprehensive confidence level based on pre-configured business rules, and generate filtering result tags for the bidding announcement data based on the corrected confidence level. The filtering result tags include at least one of recommended, pending review, or filtered. The intelligent agent adopts a top-down nine-layer decoupled architecture, including: The data storage layer includes a raw database, a feature library, a model library, a vector database, and a relational database, which provides vertical data reading and persistence support for other layers. The data acquisition layer is equipped with website tag crawling tools, an incremental synchronization mechanism, and a data deduplication engine; The data preprocessing layer is configured with text cleaning subunits, structure extraction subunits, normalization subunits, and anomaly detection subunits; The feature engineering layer is configured with a label system construction subunit, a word embedding subunit, and a feature cross selection subunit; The machine learning engine layer is configured with multiple integrated tree-based models, model fusion sub-units, and online learning sub-units; The large model semantic understanding layer is configured with a RAG retrieval enhancement subunit, a Prompt engineering subunit, a Few-shot learning subunit, and a semantic similarity calculation unit, and retrieves the historical cases by interacting with the vector database in the data storage layer. The decision fusion layer is configured with a dual-path confidence-weighted fusion subunit, a business rule engine, and a threshold adaptive adjustment subunit. The output and application layer is configured with an automatic classification sub-unit for bidding information, a recommendation scoring and ranking sub-unit, a risk warning sub-unit, and a visual dashboard. The continuous learning layer is configured with a misjudged sample recovery subunit, an active learning sampling subunit, and a model periodic retraining subunit.
Citation Information
Patent Citations
Intelligent bid evaluation and bidding content page number indexing method and system fused with large language model
CN122364230A
Vehicle guidance by a multimodal large language model
US20260175868A1
