Rag-llm-based construction engineering bid risk identification method and system

Abstract

The application discloses a building engineering bid risk identification method and system based on RAG-LLM, relates to the field of building engineering bidding analysis, and comprises the following steps: collecting bidding multi-source data of a target bidder to be analyzed from a multi-source data platform, wherein the multi-source data comprises bidding price sequences of the target bidder in each bid section of a target project and bidding submission timestamp sequences composed of timestamps of each submitted price, and the multi-source data further comprises cross-project participation records of the target bidder; extracting bidding comprehensive features from the bidding multi-source data, wherein the bidding comprehensive features comprise differential price abnormality, out-of-time sequence abnormality and cross-project activity level; and inputting the bidding comprehensive features into a RAG-LLM system, so as to recall a knowledge paragraph set from a bid case knowledge base through RAG and drive the LLM to output a corresponding building engineering bid risk grade. Thus, the fusion of structured behavior modeling and semantic analogy large model reasoning is realized, and the accuracy and interpretability of bid identification are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of construction engineering bidding analysis, and in particular to a construction engineering bid rig risk identification method and system based on RAG-LLM (Retrieval-Augmented Generation Large Language Model). BACKGROUND

[0002] Before the implementation of a project, a contracting unit is usually selected through a public bidding process. In a multi-section bidding process organized by a large real estate company, participants often use strategies such as "sectional bidding" and "slight difference bidding" to conceal bid rigging intentions: several associated companies under the same control subject submit bids for different sub-sections, and by controlling the bid difference to a very small interval (such as 0.1%-0.5%) or submitting bids at different time intervals (such as several minutes apart), traditional detection methods based on price clustering or legal relationship graphs cannot detect the abnormality. SUMMARY

[0003] The present application provides a construction engineering bid rig risk identification method and system based on RAG-LLM, a storage medium, a computer program product, and an electronic device to at least solve the problem that the current related technology cannot accurately identify and warn bid rigging behavior in a complex multi-section bidding environment.

[0004] In a first aspect, the present application provides a construction engineering bid rig risk identification method based on RAG-LLM, which comprises: collecting bidding multi-source data of a target bidder to be analyzed from a multi-source data platform; the multi-source data includes a bidding price sequence of the target bidder in each section of a target project and a bidding timestamp sequence composed of the timestamp of each bid submission, and the multi-source data also includes the cross-project participation record of the target bidder; extracting bidding comprehensive features from the bidding multi-source data; the bidding comprehensive features include: a slight difference bid abnormality degree corresponding to the bidding price sequence, a time sequence abnormality degree corresponding to the bidding timestamp sequence, and a cross-project activity degree corresponding to the cross-project participation record; inputting the bidding comprehensive features into a RAG-LLM system to recall a set of bid rig knowledge paragraphs matching the bidding comprehensive features from a bid rig case knowledge base through RAG, and using a preset prompt word to drive the LLM to output a corresponding construction engineering bid rig risk level.

[0005] In a second aspect, the embodiments of the present application provide a construction engineering bid-rigging risk identification system based on RAG-LLM, which comprises: a multi-source data acquisition unit configured to acquire bid multi-source data of a target bidder to be analyzed from a multi-source data platform; the multi-source data comprises bid price sequences of the target bidder in each bid section of a target project and bid submission timestamp sequences composed of timestamps of each submitted bid price, and the multi-source data further comprises cross-project participation records of the target bidder; a comprehensive feature extraction unit configured to extract bid comprehensive features from the bid multi-source data; the bid comprehensive features comprise: differential bid price abnormality corresponding to the bid price sequences, out-of-sequence sequence abnormality corresponding to the bid submission timestamp sequences, and cross-project activity corresponding to the cross-project participation records; and a RAG-LLM risk analysis unit configured to input the bid comprehensive features into a RAG-LLM system, to recall, by RAG, a set of bid-rigging knowledge paragraphs matched with the bid comprehensive features from a bid-rigging case knowledge base, and to output a corresponding construction engineering bid-rigging risk level by LLM driven by a preset prompt word.

[0006] In a third aspect, an electronic device is provided, which comprises: at least one processor, and a memory connected to the at least one processor in communication, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the steps of the construction engineering bid-rigging risk identification method based on RAG-LLM of any embodiment of the present application.

[0007] In a fourth aspect, the embodiments of the present application provide a storage medium having a computer program stored thereon, characterized in that the program, when executed by a processor, implements the steps of the construction engineering bid-rigging risk identification method based on RAG-LLM of any embodiment of the present application.

[0008] In a fifth aspect, the embodiments of the present application provide a computer program product comprising computer programs / instructions, which, when executed by a processor, implement the steps of the construction engineering bid-rigging risk identification method based on RAG-LLM of any embodiment of the present application.

[0009] The construction engineering bid-rigging risk identification method and system based on RAG-LLM provided by the present application can at least produce the following technical effects:

[0010] (1) By unified collection and multi-dimensional fusion of the bid offer sequence, the bid submission timestamp sequence, and the cross-project participation record, key features under the collusion mode of "micro-differential bidding", "peak-shifting bidding", and "associated company rotation participation" can be accurately described. Based on these rich and complementary feature dimensions, abnormal signals in the aspects of price, time, and participation structure can be captured simultaneously during risk identification, and the detection rate and fine-grained positioning ability of hidden bid-rigging behaviors are significantly improved through multi-source and multi-dimensional bid feature modeling.

[0011] (2) The bid comprehensive features are introduced into the RAG retrieval module to recall the bid-rigging case knowledge passages with the highest feature matching degree in real time, and the LLM is driven for risk level assessment based on the context, which not only makes the model judgment closer to the real case, but also provides a traceable knowledge basis for each risk output. Therefore, through the fusion of semantic-level case matching and reasoning, the robustness and interpretability of the bid-rigging identification system in the face of new or variant bid-rigging methods are greatly improved.

[0012] Through the technical solution, the whole process automation from data collection, feature extraction, case retrieval to risk assessment is realized. Once new bid data is generated, a pre-warning report can be output in real time, greatly shortening the response time delay of bid evaluation. At the same time, through the extensible knowledge base maintenance and prompt word adjustment mechanism, the system can quickly adapt to new industry rules and bid-rigging cases without model retraining, maintaining long-term effectiveness and low maintenance cost. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, a brief introduction to the drawings needed to be used in the embodiments or prior art description will be given below. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0014] Figure 1 A flowchart of an example of a construction engineering bid-rigging risk identification method based on RAG-LLM according to an embodiment of the present application is shown;

[0015] Figure 2 An operation flowchart of an example of extraction of cross-project activity level according to an embodiment of the present application is shown;

[0016] Figure 3 An operation flowchart of an example of tensor decomposition based on a non-negative block tensor decomposition model to calculate cross-project activity level according to an embodiment of the present application is shown;

[0017] Figure 4An operation flow chart of an example of predicting an output construction project bid encirclement risk level by driving a RAG-LLM system according to an embodiment of the present application is shown;

[0018] Figure 5 An operation flow chart of an example of fine-tuning for a RAG-LLM system is shown;

[0019] Figure 6 An experimental comparison effect schematic diagram of an example of Accuracy and F1 Score is shown;

[0020] Figure 7 A structural block diagram of an example of a RAG-LLM-based construction project bid encirclement risk identification system according to an embodiment of the present application is shown;

[0021] Figure 8 A system framework schematic diagram of an example of a RAG-LLM-based construction project bid encirclement risk identification system according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in detail with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0023] It should be noted that in the prior art, three types of means are mainly used for bid encirclement risk identification: price characteristic analysis method, legal person and correlation relationship mining method, and static rule screening method.

[0024] In the price characteristic analysis method, statistical threshold or distance / density-based clustering algorithm is mainly used to detect the concentration of bid prices. For example, bids with a price difference less than a certain preset proportion (such as 1%) are considered as suspicious bid rigging; or all prices are clustered in the numerical space, and extreme clusters are determined as abnormal. However, when the bid rigging party controls the bid price within an extremely narrow range (such as 0.1%-0.3%), and the range is highly consistent with the real bid price, simple threshold detection or clustering boundary will misjudge it as normal fluctuation, resulting in that the bid rigging case with high concealment is ignored. In addition, in the case where the bid scale or the project bid section price difference itself is extremely small, the "widening" or "tightening" of the threshold setting may cause a large number of false positives or false negatives.

[0025] In methods for mining legal entities and related relationships, corporate relationship graphs constructed based on data such as corporate equity, director relationships, and historical cooperation records can reveal long-term and relatively stable control chains. However, parties involved in bid-rigging often rapidly reconstruct corporate relationships through short-term equity transfers, the establishment of temporary shell companies, or the use of nominee accounts, resulting in blind spots in the timeliness of conventional graphs. In particular, when multiple subsidiaries appear alternately in different projects and complete equity transfers before and after the bidding deadline, such "dynamic shell" relationships are difficult to capture in real time, causing a lag in the relationship mining module and preventing it from being included in risk assessment.

[0026] In static rule-based screening methods, common screening rules include "bidding from the same address," "registering with the same email address," and "bidding price difference being too small." These rules are easy to implement and can quickly filter out some low-level collusion behaviors, but once the rule items are fixed, they are difficult to adapt to new collusion methods. Faced with the diverse operations of bidders, rule maintenance requires frequent manual updates, and the addition of new rules also requires expert judgment and historical data verification, resulting in high overall costs and slow response times.

[0027] It should be understood that the above description of the relevant technologies is intended only to help the public better understand the inventive spirit and motivation of this application, and is not intended to limit this application. Furthermore, the technical solutions described in the above-mentioned relevant technologies are not prior art, and may also be undisclosed technical solutions, such as those under research or in the laboratory stage.

[0028] Figure 1 A flowchart illustrating an example of a RAG-LLM-based method for identifying bid-rigging risks in construction projects according to an embodiment of this application is shown.

[0029] Regarding the execution subject of the method in the embodiments of this application, it can be any controller or processor with computing or processing capabilities. By adopting a fusion mechanism of structured modeling and generative language reasoning, structured data provides a quantitative basis for behavior analysis, while the large language model improves the accuracy of recognition by performing case-based understanding and inductive judgment of behavior patterns through retrieval enhancement and generation capabilities.

[0030] In some examples, it can be integrated into an electronic device or terminal through software, hardware, or a combination of both, and the type of terminal or electronic device can be diverse, such as mobile phones, tablets, or desktop computers, etc.

[0031] like Figure 1 As shown, in step S110, the multi-source bidding data of the target bidder to be analyzed is collected from the multi-source data platform.

[0032] In the construction project bidding supervision scene, the information involved in the bidding behavior is relatively scattered, and there are data barriers or inconsistent data granularity between different systems. Therefore, by breaking through the data interface of heterogeneous platforms, the behavior information of the target bidder in multiple dimensions can be uniformly gathered. Specifically, the multi-source data platform includes but is not limited to government public resource transaction platform, large real estate enterprise procurement system, industry supervision platform and third-party evaluation agency database, etc., and the data interface is standardized designed to support automatic retrieval and format unification of bidding information. In addition, the target bidder can refer to any bidder and can be specified by user input, such as bidding entry or entities with bidding intention for specific projects, etc.

[0033] Here, the multi-source data includes the bidding submission timestamp sequence composed of the bidding price sequence of the target bidder in each bid section of the target project and the timestamp of each submission price, and the multi-source data also includes the cross-project participation record of the target bidder.

[0034] Specifically, the bidding price sequence refers to the valid bidding price data submitted by the target bidder in multiple bid sections of a target project, for example, the data is indexed by bid section number and organized as an ordered sequence according to the submission order of the price, which helps to identify price coordination behavior (such as 0.1%-0.5% micro-difference control). The bidding submission timestamp sequence refers to the submission time corresponding to each price operation, which is used to analyze the concentration or misplacement of bidding operations in the time dimension, and provides a data basis for identifying mis-time bidding behavior and other coordination behavior patterns. The cross-project participation record refers to the participation of the target bidder in other projects in the same period or near time window, including bidding time, project attributes, cooperation subjects, bid section quantity and other information, which facilitates to describe the breadth and frequency of its bidding behavior, and helps to identify whether it participates frequently, whether it participates with other subjects, whether it concentrates bidding on a certain bidding organization, etc. The above multi-source data is cleaned, desensitized and stored in an internal database structure, so as to construct a structured and timely bidding behavior portrait by uniformly collecting and standardizing the multi-dimensional data such as bidding price, timestamp and participation record.

[0035] In step S120, the bidding comprehensive features are extracted from the bidding multi-source data. Thus, the original bidding multi-source data collected is converted into structured numerical features that can be used for pattern recognition.

[0036] Here, the bidding comprehensive features include: the micro-difference price abnormality corresponding to the bidding price sequence, the mis-time sequence abnormality corresponding to the bidding submission timestamp sequence, and the cross-project activity corresponding to the cross-project participation record.

[0037] Regarding the abnormality of differential bidding, for the bidding sequence of a target bidder in multiple bid sections, if the bidding difference always remains in a small range (such as 0.1%-0.5%), it indicates the possibility of coordination by controlling the bidding interval. Exemplarily, the local fluctuation can be calculated using the sliding window method, and by calculating the bidding standard deviation, the mean of the difference between adjacent bids, the range (maximum-minimum), and the relative volatility (relative to the bidding control price), etc., a comprehensive abnormal score index is finally obtained. The higher the score, the more abnormal the bidding distribution is, and there may be signs of human manipulation.

[0038] Regarding the abnormality of staggered sequence, if a control subject operates multiple company bidders in multiple bid sections, it will often precisely control the bidding submission time points of each bidding unit to avoid system detection of operation traces. However, such processing may deviate from the normal bidding time distribution (for example, the "last sprint" mode), and thus obtain a corresponding quantitative staggered sequence abnormality index.

[0039] Regarding cross-project activity, it is used to assess the target bidder's participation intensity and correlation in similar projects within the same period. If a company participates in a large number of projects in a short period of time and frequently appears in the same bid section or project of the same bidding unit with a few companies, it is suspected of being involved in a bid-rigging network. The calculation method of activity can be diverse, exemplarily, by comprehensively calculating the bidding frequency, participation project distribution density, and repeated cooperation subject proportion, etc., and a time sequence weighting mechanism can also be introduced for weighted processing of short-term concentrated bidding behavior, etc.

[0040] Thus, by establishing a "behavior-abnormality" correspondence relationship based on bidding comprehensive features, the mapping and conversion from raw data to structured knowledge are realized, enabling the system to describe complex and hidden coordination behavior characteristics in a data-driven manner.

[0041] In step S130, the bidding comprehensive features are input into the RAG-LLM system to recall a set of bid-rigging knowledge paragraphs matching the bidding comprehensive features from the bid-rigging case knowledge base through RAG, and output the corresponding construction engineering bid-rigging risk level using the pre-set prompt word driven LLM.

[0042] Here, the case knowledge is matched through the RAG (Retrieval-Augmented Generation) mechanism and combined with the quantitative features at the semantic level to realize analogical reasoning and level judgment of bid-rigging risks. Exemplarily, the comprehensive feature vector of the target bidder is input into the vector retrieval engine to retrieve the most similar bid-rigging case paragraph from the bid-rigging case knowledge base.

[0043] It should be noted that multiple knowledge segments of bid-rigging cases are pre-stored in the bid-rigging case knowledge base. Each knowledge segment is indexed through knowledge processing metadata. For example, these case knowledge segments can be pre-processed with features to extract the price distribution, time-based collaboration patterns, and corporate network relationships of previously identified bid-rigging behaviors. Combined with the specific bid-rigging behavior patterns of the cases (e.g., price aggregation collaboration behavior, bid-rigging network behavior, etc.), one or more features are selected from micro-difference price anomaly, staggered time sequence anomaly, and cross-project activity to construct the indexing metadata of the corresponding case knowledge segments, so as to achieve feature matching with the input comprehensive bidding features.

[0044] After obtaining the set of knowledge segments that match the comprehensive characteristics of the bid, these segments are combined with the original comprehensive characteristics to construct prompt words, which are then input into a large language model (such as Deepseek or Qwen series). Here, the prompt words have a clearly defined structure, including project background, behavioral patterns, knowledge references, output requirements, etc., guiding the model to reason based on case comparison, determine whether the behavior has the risk of bid rigging, and output the risk level (such as low, medium, high). At the same time, the inference basis is generated, such as "the bids are extremely aggregated and the time intervals are low, which is suspected to be human collusion."

[0045] More specifically, the input format in the prompt word template is designed as follows:

[0046] Role definition: such as "You are now a bid evaluation analyst, and you need to analyze whether the target bidder has any risk of bid rigging";

[0047] Input feature summary: including price volatility, bidding time distribution, activity value, etc.;

[0048] Search paragraph summary: Displays the content of the knowledge paragraphs related to bid-rigging from the 3-5 most relevant cases recalled;

[0049] Question instructions: such as "Based on the above content, please determine whether there is a risk of bid rigging, and provide the risk level and reasons."

[0050] This enables intelligent reasoning and transformation between structured abnormal behavior and semantic knowledge, overcoming the limitation of traditional models that can only make "yes / no" judgments but cannot explain "why". By combining case semantics, behavioral features, and generative understanding, the system can not only judge the risk of bid rigging, but also provide readable, traceable, and regulatory-relevant judgment criteria, greatly improving the interpretability and credibility of the identification results.

[0051] Regarding the extraction of the differential bid abnormality in step S120, in some examples of the embodiments of the present application, the bid price sequence is divided by a sliding window, the standardized value of the bid price data in each window is calculated, the distribution pattern of the window standardized value sequence is analyzed by a lightweight neural network model, and the corresponding differential bid abnormality is output.

[0052] Here, in the risk identification scene of the construction project bid surrounding, the bid rigging behavior of the bidder often shows as a small amplitude adjustment of the bid in a short period and a specific bid section to evade global threshold detection, and it is difficult to be distinguished in the whole bid section or the whole time series statistics. The sliding window segmentation combined with the window standardization can amplify the "differential" features in the local context. Then, the distribution pattern of the standardized sequence is learned by a lightweight neural network, and the model can automatically capture the subtle differences between "normal fluctuations" and "abnormal aggregation" patterns.

[0053] More specifically, in the sliding window division, the bid price sequence of the target bidder in a bid section is

[0054] , formula (1)

[0055] In the formula, is the th bid price, is the total number of bid prices. The window length and the step size are taken, is divided into overlapping subsequences:

[0056] , formula (2)

[0057] Wherein

[0058] , formula (3)

[0059] In the formula, denotes the window length, denotes the step size, denotes the number of windows, is the bid subsequence of the th window.

[0060] It should be noted that the bid rigging bidder often hides the bid rigging intention through "small amplitude aggregation" in a continuous bid, and this pattern is easily covered by normal bid fluctuations in the global perspective. The sliding window segmentation can decompose the long sequence into several local paragraphs, and focus on mining the subtle bid distribution changes in each paragraph. Decomposing the long sequence into context-related local signals significantly improves the sensitivity to the "short-term concentrated bid" pattern; the window overlap strategy ensures that the cutting boundary can also capture the abnormality, reducing the missed detection.

[0061] In window-wise normalization, for each window Calculate local mean and standard deviation:

[0062] , equation (4)

[0063] Then get the normalized sequence

[0064] , equation (5)

[0065] Form the vector .

[0066] In the formula, , are the mean and standard deviation of the first window, respectively, is a small constant to prevent division by zero, is the normalized window feature vector.

[0067] It should be understood that the absolute value range of different projects and bidders is quite different. If the original values are directly compared, it will be misjudged due to different dimensions. Standardization within each window can smooth out the scale difference and highlight local abnormal fluctuations.

[0068] Further, by designing a lightweight network, the model structure can include a one-dimensional convolutional neural network (1D-CNN): extract sequence microstructure features between windows and within windows, suitable for sliding window sequential data, such as 2 layers of Conv1D+ReLU+Pooling+Flatten+Dense, and finally connected to Sigmoid output. The training data set can be formed by splicing historical normal bidding samples and already determined surrounding samples, and the label is "normal" or "micro-difference anomaly". In addition, the loss function can use binary cross-entropy, and the optimization goal is window micro-difference anomaly binary classification.

[0069] Calculate the abnormality degree for each window:

[0070] , equation (6)

[0071] In the formula, represents the micro-difference bid abnormality degree of the first window, represents the neural network mapping with parameters . Specifically, , and the larger the value, the more abnormal the window "micro-difference" is.

[0072] In this way, the sequence obtained after window normalization carries local fluctuation pattern information. The lightweight neural network can learn the subtle differences in shape and spectrum between normal and abnormal fluctuations, thereby giving each window an abnormality degree score.

[0073] Finally, aggregate the local anomaly scores of all windows into the differential bid anomaly score of the whole sequence . Take the maximum value to capture the strongest abnormal signal:

[0074] , formula (7)

[0075] In the formula, represents the differential bid anomaly score output of the whole sequence.

[0076] Therefore, by sliding window + standardization, small bid aggregation in the same context is highlighted, and subtle string intention that traditional global statistics ignore can be captured. In addition, the step and window length can be adjusted to support automatic adaptation to projects with different bid densities and bid periods, and multi-scale compatibility is realized.

[0077] Regarding the extraction of cross-project activity in step S130, in some examples of the embodiments of the present application, the interval value of adjacent timestamps in the bid submission timestamp sequence is calculated to obtain a corresponding interval value sequence; the interval value sequence is reconstructed based on an autoencoder model to obtain a corresponding reconstruction error, and the autoencoder model is pre-trained according to a sample set of normal patterns; and the reconstruction error is used to determine the time error sequence anomaly degree.

[0078] It should be noted that the string bidder often intentionally splits the submission time within the bid deadline, such as "evenly scattered" at intervals of a few minutes or "jumping submission" at low peak times. By calculating the time interval sequence of adjacent submissions, this "time error" behavior can be converted into a quantifiable time sequence signal.

[0079] Specifically, the bid submission timestamp sequence of the target bidder is extracted from the procurement platform log , wherein is the exact time of the th submission.

[0080] The adjacent time difference is calculated.

[0081] If the system has multiple bid sections submitted in parallel, it needs to be grouped according to "project ID + bidder ID" first, and each group is calculated separately. In this way, discrete timestamp data is converted into an interval sequence of the same length, forming a unified time sequence analysis basis.

[0082] It should be noted that the bid submission interval is often extremely unevenly distributed. Specifically, normal bids may be concentrated in the last few minutes, resulting in a large number of small values; while the time error strategy submits throughout the bid period, resulting in long-tailed intervals. Direct modeling of the original is difficult, and normalization and transformation are needed to highlight abnormal patterns.

[0083] Specifically, by logarithmic transformation The long-tail distribution is compressed, and the model's recognition ability for short, medium, and long intervals is enhanced, thus effectively alleviating the long-tail problem through logarithmic transformation, making the autoencoder more easily convergent. In addition, missing or abnormal extreme values (such as or exceeding the upper limit) can be truncated to the trainable range by upper and lower limits.

[0084] Then, time series reconstruction based on the autoencoder is performed. It should be noted that the normal bid submission interval sequence has a clear "one submission final sprint" pattern. After pre-training with only normal samples, the model has a low reconstruction error for the normal pattern. The dispersion or interval jump of the wrong time submission is difficult to reconstruct well, resulting in a higher error. In addition, the encoder and decoder of the autoencoder can use a symmetric multi-layer LSTM structure.

[0085] In the pre-training of the autoencoder, the training set can use a large number of real "unserial bid" bidder submission time series to ensure representativeness and time period diversity, and can use mean square reconstruction error to monitor the reconstruction error distribution on independent normal samples to determine the reconstruction level baseline.

[0086] After deployment is completed, the model only does forward inference, inputs the preprocessed sequence, and outputs the reconstructed sequence . The reconstruction error at each time point is calculated. The LSTM autoencoder is trained on normal samples so that it can capture sequence dependence and periodic trends.

[0087] Further, the wrong time abnormality degree can reflect the deviation of the entire bid time series from the normal pattern. Through the statistical distribution of the reconstruction error, the local peak or overall dispersion can be quantified as a single indicator.

[0088] Specifically, the error sequence is obtained:

[0089] , equation (8)

[0090] By using the quantile strategy for error aggregation:

[0091] , equation (9)

[0092] By aggregating the errors, both local and global deviations are included in the measurement, taking into account both sudden wrong time and long-term dispersion risk patterns. In addition, when exceeds the threshold , a high-risk alert can be triggered, and the threshold can be calculated using a dynamic statistical method based on the median absolute deviation (MAD) to achieve dynamic adaptive adjustment.

[0093] In the course of practicing this invention, it was found that the bidding time distribution in the field of construction engineering bidding often follows a "one-stage submission and final sprint" pattern. This refers to the behavior of most bidders submitting all their bid documents in a short period of time before the bid deadline. Specifically, in the long period after the bidding begins, the submission frequency is low and the intervals are large; however, once the deadline approaches (e.g., the last 5-10 minutes), most bidding operations are suddenly concentrated, the time intervals are drastically shortened, and a "sprint-style" high-density submission is presented.

[0094] The main reasons for this model are as follows: Bidders often observe and prepare documents in the early stages, only submitting them all at once in the final stage after confirming everything is correct and the pricing strategy is finalized, in order to reduce the risk of being imitated or copied by other bidders. Therefore, using adjacent submission intervals... Representing time series characteristics, when When far from the deadline, Larger and more volatile; when As the deadline approaches, A significant drop forms a "sharp decline" range. By modeling this pattern as a "normal" baseline, the autoencoder can learn the typical "low-frequency observation + high-frequency sprint" time series structure. Any behavior that significantly deviates from this structure (such as uniform submission throughout the day or deliberate staggered distribution) can be highlighted in the reconstruction error, helping to identify and avoid risks caused by timing errors.

[0095] Figure 2 A flowchart illustrating an example of extracting cross-project activity according to an embodiment of this application is shown.

[0096] like Figure 2 As shown, in step S210, the three-dimensional project participation tensor of the target bidder is constructed.

[0097] It should be noted that in the construction industry, there is a significant phenomenon of bidders repeatedly participating across multiple projects and time periods. Some unusual bidders often concentrate their participation in multiple projects or tenders within the same timeframe (such as the same quarter or the same month). This high-frequency, cross-project participation pattern is relatively rare in normal business, but it is a common strategy used by bid-rigging groups to evade supervision. Therefore, by constructing a three-dimensional tensor of "bidder-project-timeframe," it is possible to quantify the behavioral trajectory of bidders participating in the engineering market from a multi-dimensional interactive perspective, and to capture unconventional, group-based active patterns.

[0098] Specifically, the three dimensions of the 3D project participation tensor are bidder, bidding participation project, and bidding participation time slice. Specifically, in the first dimension ( Assign a unique number to all bidders in the second dimension ( Assign a unique number to all engineering projects / sections during the observation period, in the third dimension ( ) Divide the analysis period into uniform time slices (e.g. in units of month, week or quarter), each time slice is uniquely numbered.

[0099] Construct a three-dimensional tensor by tensor filling , element represents whether the bidder participated in the project in the time slice . Specifically, represents a participation record, represents no participation record. The multi-dimensional tensor structure breaks the expression bottleneck of multi-project, multi-period, group behavior of two-dimensional table.

[0100] In step S220, the three-dimensional participation tensor is decomposed by tensor to obtain a bidder factor matrix, so as to calculate the cross-project activity degree.

[0101] It should be noted that in the three-dimensional tensor, the normal bidder shows a scattered distribution, and the project and time slice are weakly related, while the abnormal bidder (such as a bid-rigging gang) often shows an implicit pattern of high project concentration and time slice synchronous outbreak. Therefore, various non-limiting high-order tensor decomposition methods can be used, such as CP decomposition, Tucker decomposition, dynamic tensor decomposition or block tensor decomposition. Through high-order decomposition, the "participation mode factor" hidden in the multi-dimensional data is automatically extracted, and the high-dimensional behavior is compressed into a low-dimensional feature matrix. The decomposition realizes the automatic "clustering" of the participation mode, thereby revealing the group activity rule of the bidder and distinguishing the behavior difference between the active bidder and the ordinary bidder.

[0102] Specifically, after tensor decomposition, each row of the bidder factor matrix reflects the "participation intensity" of the bidder under the high-order activity mode. By integrating the intensity scores under each mode, the degree of cross-project, multi-period activity of the bidder within the entire observation period can be quantified as a sensitive index of bid-rigging gang behavior.

[0103] Exemplarily, for the bidder , the cross-project activity degree of the bidder can be defined as:

[0104] , formula (10)

[0105] In the formula, is the score of the bidder under the mode .

[0106] By constructing the tenderer-project-time slice as a three-dimensional tensor and performing high-order decomposition on the tensor, complex group collaborative behaviors can be captured, and hidden group active patterns can be automatically discovered. Normal tenderers usually show sparse participation in terms of project dispersion and time dispersion, while abnormal groups show high-dimensional interaction in terms of project concentration and time synchronization explosion in a specific mode. The tenderer factor matrix obtained by decomposition can quantify the score of each person in multiple potential active patterns, and then refine it into a single cross-project activity index, providing objective and interpretable multi-dimensional behavior characteristics for subsequent risk identification.

[0107] In some examples of the embodiments of the present application, the cross-project activity can be obtained by a non-negative block tensor decomposition (NBTD) method.

[0108] Figure 3 An operation flowchart of an example of tensor decomposition based on a non-negative block tensor decomposition model to calculate cross-project activity according to an embodiment of the present application is shown.

[0109] As shown in Figure 3 , in step S310, the three-dimensional participation tensor is divided according to the number of tenderers and the project category to generate a plurality of non-overlapping sub-tensor blocks.

[0110] Here, in order to solve the problem of large-scale tenderer-project-time three-dimensional participation data with large calculation amount and obvious mode difference, the original three-dimensional tensor is first divided into blocks according to the tenderer attributes and the project category. By dividing the tenderer set into several groups and dividing the project set into several categories, a plurality of non-overlapping sub-tensor blocks are obtained, each block containing only the participation of the corresponding tenderer and the corresponding project in each time slice.

[0111] Specifically, the three-dimensional participation tensor is divided into groups according to the number of tenderers and groups according to the project category, generating a total of non-overlapping sub-tensor blocks .

[0112] , formula (11)

[0113] In the formula, denotes the original participation tensor, denotes the total number of tenderers, denotes the total number of projects, denotes the total number of time slices, and are the number of tenderer groups and the number of project categories, respectively, denotes the i th tenderer group, denotes the j th project category, and denotes the k th time slice. group bidders and the first sub-tensors corresponding to the first group of bidders, the number of bidders in the first group, the number of project categories.

[0114] In this way, each sub-tensor block can be independently decomposed and analyzed, and the overall computational complexity can be significantly reduced by using distributed or multi-thread parallel computing. In addition, the grouping of bidders can be divided according to the size of the enterprise, the historical participation frequency or the geographical region; the project category can be defined based on the type of the project, the scale of the bidding and procurement or the function of the bid section. In this way, the participation patterns between different groups and categories often have differences, and by blocking, the interference between cross-group and cross-category patterns can be avoided, and the sensitivity of tensor decomposition to each pattern can be improved.

[0115] In step S320, two-stage non-negative decomposition is performed on each sub-tensor block.

[0116] On each sub-tensor block, the non-negative CP decomposition is used to extract the “bidder-project-time” three-dimensional pattern factor. The non-negative constraint ensures that the factor matrix extracted by decomposition has an interpretable intensity meaning, which can intuitively reflect the participation degree of each bidder in a specific project and time dimension. The factor values are all non-negative, representing the participation frequency or activity level; therefore, the original data can be approximated with a lower rank, and the dominant behavior pattern can be efficiently refined.

[0117] Specifically, in the time dimension decomposition of the first stage, for each sub-tensor block CP decomposition is directly performed.

[0118] , formula (12)

[0119] In the formula, CP decomposition rank, indicating how many groups of factors are extracted; indicates the weight of the first group of factors, indicates the bidder factor vector, indicates the project factor vector, indicates the time factor vector and , indicates the vector outer product operation.

[0120] In the spatial dimension decomposition of the second stage, the factor set after CP decomposition is regarded as a tensor component, and non-negative spatial decomposition is directly performed.

[0121] Specifically, in the spatial dimension decomposition of the second stage, the bidder factor matrix and the project factor matrix After treating it as a composite matrix, directly multiply it on the left by the non-negative bidder subspace basis. And right multiplication of nonnegative item subspace basis Perform nonnegative space decomposition:

[0122] Equation (13)

[0123] In the formula, , Denotes the basis matrix of the bidder's subspace. Represents the basis matrix of the project subspace. Represents the core interaction matrix. Represents the rank of the spatial decomposition.

[0124] Here, for the bidder and project factor matrices, a non-negative Tucker decomposition is further applied to compress the CP decomposition factors into subspaces, reducing redundant factors and extracting higher-order coupling features. This maps similar bidders or projects to the same subspace, achieving aggregation of group behavior patterns. Thus, the coupling relationship between bidders and projects is effectively aggregated into a few high-meaning features, and group association identification assists in identifying cross-project collaborative groups.

[0125] In step S330, the spatiotemporal activity matrix of each sub-block is constructed.

[0126] Equation (14)

[0127] In the formula, Indicates the spatiotemporal activity of a sub-block. This represents the matrix transpose operation. , This indicates the spatiotemporal activity of a sub-block.

[0128] In step S340, the spatiotemporal activity matrix is ​​solved by introducing a time decay factor to obtain the bidder's cross-project activity in different time slices.

[0129] After combining the subspace decomposition results with a time factor, an exponential time decay is introduced to weight the activity of each bidder in different time slices. Specifically, the target bidder... In time slice activity level Expressed by the following formula:

[0130] Equation (15)

[0131] In the formula, For matrix The Middle Line number Column elements, The time decay coefficient, for the current analysis time point, denotes a time decay factor.

[0132] Here, the decay coefficient can be determined according to the project bidding cycle length and the regulatory window time, and can also be fitted by historical data. Thus, the latest behavior data at the regulatory moment has a higher weight, which can timely capture the sudden concentrated bidding behavior of the surrounding bidder, and gradually fade the long-term participation by using the decay weight, and reduce the influence of the out-of-date behavior. The analysis focus is placed on the spatio-temporal activity close to the current review time point, which is more in line with the dynamic of the surrounding bidding site.

[0133] In step S350, the cross-project activity of each time slice is aggregated to obtain the cross-project activity of the target bidder .

[0134] , formula (16)

[0135] In the formula, denotes a time weighting function, is a slope parameter of the weighting function, is a center offset time of the weighting function, denotes the cross-project activity of the target bidder in the time slice . In some embodiments, the slope and the center offset can be optimized by maximum likelihood estimation to match the typical project submission rhythm.

[0136] Here, the spatio-temporal activity results of the bidder in all time slices are fused into a single score according to the logistic regression time weighting, to reflect the overall cross-project participation intensity and concentration of the bidder. Thus, through the comprehensive index , the cross-project play of the bidder in the whole life cycle can be reflected.

[0137] Through the embodiments of the present application, the original “bidder-project-time” three-dimensional participation data is divided into sub-blocks which are more homogeneous in structure by using the non-negative block tensor decomposition model, and a multi-level, non-negative constrained decomposition operation is performed on each sub-block. This not only helps to reveal the potential collaborative mode between the local bidder group and the specific project category, but also facilitates the separation and modeling of the dynamic changes in the time dimension. Thus, in the large-scale and complex engineering bidding data environment, the detection capability of cross-project and cross-period serial bidding behavior can be significantly improved, and an abnormal activity signal with clear spatio-temporal attribution and structural explanation can be output.

[0138] It should be noted that in the behavior of stringing in the construction network, the stringing is often distributed in different projects and multiple time periods in a group and block form for coordinated manipulation. Compared with the traditional global CP or Tucker decomposition, the NBTD adopted in the embodiments of the present application can first divide the sub-tensor blocks according to the project categories and the bidder groups, and then independently apply non-negative constraints to each block for decomposition, perfectly fitting the strategy of “local coordination + global heterogeneity”. Through the combination of blocking and non-negative decomposition, on the one hand, the coordinated behavior mode in each block is highlighted, and on the other hand, the interference between the modes in different blocks is avoided, providing clearer and more meaningful high-order features for risk quantification and large model reasoning.

[0139] Figure 4 An operation flowchart of an example of predicting an output construction project bid-rigging risk level by driving a RAG-LLM system according to an embodiment of the present application is shown.

[0140] As shown in Figure 4 , in step S410, the top K case knowledge passages with the highest matching degrees are retrieved from the bid-rigging case knowledge base respectively for the micro-price abnormality degree, the out-of-time sequence abnormality degree and the cross-project activity degree in the bid submission comprehensive features.

[0141] In bid-rigging risk discrimination, the “micro-price abnormality degree”, the “out-of-time sequence abnormality degree” and the “cross-project activity degree” respectively reflect the typical risk signals in the price dimension, the time sequence dimension and the network dimension. Retrieving the most relevant historical cases for each signal separately can ensure that the model obtains the most representative evidence support when reasoning on this signal, rather than relying on the generalized global case set. Multi-dimensional feature recall also avoids the blind area of a single retrieval method for certain risk patterns.

[0142] In some embodiments, the three feature retrievals can be initiated in parallel, and the results are returned asynchronously to shorten the overall recall delay.

[0143] In step S420, the retrieved case knowledge passages are de-duplicated and combined to obtain a matched bid-rigging knowledge passage set; the total number of passages in the knowledge passage set is less than or equal to 3K.

[0144] It should be noted that a bid-rigging case knowledge passage may simultaneously have multiple risk features, so that the branch recalls from the three features may have the same or highly similar case passages. If not de-duplicated, redundant cases will cause the subsequent reasoning module to repeatedly consume the same evidence, waste computing resources, and weaken the diversity of the reasoning chain. Combining and limiting the total number of passages to 3K (K is the number of single-feature retrievals) can control the input size while ensuring coverage of all signals.

[0145] After the RAG retrieves the content, the content is combined with the preset prompt words as references and input into the LLM.

[0146] To ensure that RAG-LLM performs reasoning in order, branches, and hierarchies under multi-dimensional risk factors, rather than one-time "black box" output, the preset prompt words use a chain-of-thought prompting structure, and the RAG-LLM system includes a main module and multiple branch reasoning modules, i.e., a "one trunk + three parallel branches" structure is constructed inside the model, so that each risk factor can independently reason and jointly participate in the final integrated decision.

[0147] For example, the chain-of-thought prompting design idea is as follows:

[0148] The unified prefix states: "Please analyze and give a conclusion based on the following information."

[0149] Step1 (price dimension): "

First branch

[0150] Step2 (time sequence dimension): "

Second branch

[0151] Step3 (cross-project dimension): "

Third branch

[0152] Step4 (main trunk integration): "

Main trunk

[0153] Specifically, during reasoning, the prompt and the recalled knowledge paragraphs are transmitted into the model, the branch modules generate their own conclusions in parallel, and then they are summarized to the main trunk. The first three steps Step1-3 are bound to the corresponding branch Adapter, and the branch input only contains the branch factor and the full set of case paragraph set. The prompt of the last Step4 is bound to the main trunk LoRA Adapter, which is used for comprehensive decision. In this way, the reasoning architecture of "separation but not separation, separation but not dispersion" is realized: the branches focus on each factor, and the main trunk unifies the attribution; through chain-of-thought prompting, the context interference of synchronous reasoning is reduced, and the accuracy and explainability of the model in multi-factor discrimination are improved.

[0154] In step S431, the differential bid abnormality degree and the surrounding bid knowledge paragraph set are reasoned based on the first branch reasoning module to determine the corresponding price string bid risk identification result.

[0155] Here, the price string bid risk identification result is used to indicate whether there is a price string bid risk and the price abnormality reference case knowledge cited in the reasoning.

[0156] Here, based on the "differential bid abnormality degree" Reflects the suspicious signal of small-scale concentrated bidding of bidders in different bid sections of the same project. Through branch reasoning combined with surrounding bid case knowledge paragraph set, the corresponding price factor identification result is generated through prompt driving.

[0157] Exemplarily, the model output format can be:

[0158] Conclusion: There is / there is no price string bid risk.

[0159] Reason:...

[0160] Reference case: Case #12, paragraphs 3-5.

[0161] Here, by setting the Adapter branch dedicated to the price factor, the model can still maintain high accuracy when facing mixed features, and the reference case output enhances the auditability of the judgment, and the supervisory personnel can directly locate the case paragraph for review.

[0162] In step S432, the out-of-time sequence abnormality degree and the surrounding bid knowledge paragraph set are reasoned based on the second branch reasoning module to determine the corresponding out-of-time avoidance behavior identification result.

[0163] Here, the out-of-time avoidance behavior identification result is used to indicate whether there is an out-of-time avoidance behavior risk and the time sequence abnormality reference case knowledge cited in the reasoning. Out-of-time avoidance focuses on the abnormal distribution of bidding submission rhythm. The reconstruction error is quantified, and the factor is analyzed through the exclusive Adapter branch, which can accurately locate various out-of-time avoidance behavior patterns. In addition, the second branch Adapter can only update the LoRA for the time sequence sensitive layer. Through the parallel design of branch independent operation, cascading errors caused by misjudgment of the previous branch are avoided.

[0164] In step S433, the cross-project activity degree and the surrounding bid knowledge paragraph set are reasoned based on the third branch reasoning module to determine the corresponding cross-project string bid identification result.

[0165] Here, the cross-project string bid identification result is used to indicate whether there is a cross-project string bid risk and the project correlation abnormality reference case knowledge cited in the reasoning. Cross-project activity Reflecting the abnormal high-frequency participation behavior of bidders at different times in multiple projects, the exclusive branch can focus on this factor and combine case illustrations to focus on the group operation mode.

[0166] In step S440, the main module integrates the price string risk identification results, the time avoidance behavior identification results, and the cross-project string identification results to determine the construction project bid risk level, and generates a bid reasoning reference knowledge chain combining price anomaly reference case knowledge, time sequence anomaly reference case knowledge, and project correlation anomaly reference case knowledge.

[0167] Here, the main module assumes the "comprehensive view" responsibility, fuses the three branch independent output results with the corresponding reference cases, gives the final bid risk level and generates a complete reasoning knowledge chain.

[0168] The exemplary input features of the main module are as follows:

[0169] 1) Three-branch label vector ;

[0170] 2) Three-branch reference case paragraph list;

[0171] 3) Prompt: "Integrate the conclusions of the first three steps, output the comprehensive bid risk level of the target bidder (high / medium / low), and list the key influencing factors and corresponding case paragraphs to form a complete knowledge chain."

[0172] Further, the main Adapter integrates each branch vector, and its exemplary output format is as follows:

[0173]

Comprehensive risk level

[0174]

Core factor

[0175]

Knowledge chain

[0176] 1. Price: Case #12, paragraphs 3-5;

[0177] 2. Time sequence: Case #7, paragraph 2;

[0178] 3. Cross-project: Case #3, paragraphs 1-2;

[0179] Through structured analysis output, a report is generated and stored in the audit log, and the knowledge chain nodes and case paragraphs can also be linked to support one-key jump of the regulatory system.

[0180] Through the embodiments of the present application, in the risk identification of building engineering bid encirclement, different dimensions of abnormal signals (price difference, wrong time submission, cross-project activity) each reflect a specific bid rigging strategy, and a single model cannot be considered. The chain reasoning of the large language model of "multi-branch parallel + trunk integration" is adopted, the most relevant cases are recalled for multi-dimensional features combined with RAG, LoRA is injected into the light adapter in each branch, each branch module focuses on learning and explaining a risk factor, and then the trunk module summarizes the branch conclusions and references the evidence to form a hierarchical and traceable comprehensive judgment process. Therefore, the independence of each factor reasoning is maintained, and on the basis of multi-dimensional parallel analysis, the accuracy of single factor and comprehensive risk discrimination is significantly improved, and accurate case references are provided for each step of reasoning to ensure the logical coherence and end-to-end explainability of the final decision.

[0181] It should be noted that the bid encirclement behavior in the field of building engineering bidding usually presents a highly concealed cooperative mode under the interweaving of "multiple bidders, multiple projects, and multiple time periods". Traditional static rules, single-point detection, and shallow machine learning methods cannot effectively identify these complex risk chains that are highly dependent on knowledge.

[0182] The intelligent risk identification system based on RAG-LLM provided by the embodiments of the present application has multi-modal, multi-source heterogeneous data fusion and knowledge enhancement reasoning capability, and can more accurately realize the automatic identification and explanation of bid encirclement risks.

[0183] However, only by designing a scientific data structuring method, a chain branch reasoning structure, and an efficient fine-tuning mechanism suitable for the scene based on real samples and business characteristics of the construction industry, can the industry intelligence potential of RAG-LLM be released.

[0184] Figure 5 An operation flowchart of an example of fine-tuning for the RAG-LLM system is shown, which adopts LoRA parameter efficient adaptation fine-tuning and introduces novel aspects such as multi-task joint loss and evidence correlation optimization.

[0185] As shown in Figure 5 In step S510, a structured training sample is constructed based on historical bid encirclement samples.

[0186] Here, the structured training sample includes bid multi-source sample features of a bid encirclement bidder, historical case description contents matched with the bid multi-source sample features, and a bid encirclement risk level main label and a plurality of branch node risk sub-labels for the bid encirclement bidder. The plurality of branch node risk sub-labels are used to indicate the risk states of price bid rigging, wrong time avoidance behavior, and cross-project bid rigging.

[0187] An efficient fine-tuning process first requires the large model to "understand" the chain structure of risk patterns and knowledge traceability logic in the business scenario. To this end, the original bid samples need to be structured, with the "digital features" of bidding behavior, historical case text content and multi-layer labels corresponding one by one, and the data-knowledge-label three-way channel is opened.

[0188] Specifically, the bid price sequence, bid timestamp sequence and cross-project participation record of each bidder in the project and its sections are collected, and after normalization, sliding window, time series difference, tensor decomposition and other processing, three categories of digital features (micro-difference bid abnormality, time sequence abnormality, cross-project activity) are extracted.

[0189] In terms of label alignment of historical case content, due to the differentiated performance of RAG mechanism, it is recommended to use expert-defined method to specify the knowledge paragraph ID or quoted text of the case paragraph most relevant to the sample features in the bid case knowledge base through user input information, including past bid punishment announcements, industry risk reports and typical bid chain link descriptions, which are highly matched in semantics with the current bidding features, to realize evidence chain binding.

[0190] Regarding the labeling of multi-layer labels, each structured sample is assigned a main label (comprehensive bid risk level) and multiple branch sub-labels (price bid chain risk, time avoidance risk, cross-project bid chain risk), and the branch labels directly correspond to the output of each branch node in the subsequent RAG-LLM inference chain.

[0191] In step S520, the RAG recalls a set of knowledge paragraphs matching the bidding multi-source sample features from the bid case knowledge base, and performs semantic encoding processing to obtain the corresponding bidding sample RAG knowledge features.

[0192] It should be noted that although the evidence chain of historical case content is specified above, these contents are only used as labels, and the contents recalled using the corresponding RAG should be used in the inference link. The reason is that the RAG recall capability is also an optimization direction of the system, and if these historical case contents are directly used as background knowledge during large language model training, it will seriously limit the generalization ability of the system, and also reduce its performance in the case of incomplete background knowledge. For specific RAG operation details, refer to the description in other parts of the above, which will not be repeated here.

[0193] In step S530, the bidding multi-source sample features and the bidding sample knowledge features are spliced to generate a sample input vector.

[0194] In step S540, LoRA parameter adapters are respectively inserted into the backbone module and each branch reasoning module of the RAG-LLM system, so that the backbone network is used to output a sample input vector surrounding bid risk level main label prediction result, and each branch reasoning module is used to output a prediction result of a sample input vector corresponding to each branch node risk sub-label.

[0195] It should be noted that full parameter fine-tuning based on a large model is extremely computationally intensive, has a large implementation cost, and is prone to forgetting existing knowledge. Here, a LoRA low-rank adaptation mechanism is used, and LoRA adapters are respectively inserted into the backbone and each branch reasoning module, for example, inserted into the attention layer or FFN layer of a multi-layer Transformer, to realize efficient local parameter updating while ensuring the learned ability and business sensitivity of the main branch and branch nodes.

[0196] During training, a structured sample input vector (features + knowledge) is simultaneously input into the main branch and all branch modules. Each branch LoRA parameter is only sensitive to its branch loss, and the main task loss is optimized by the main branch LoRA adapter, and all branch information is finally fed back to the main branch module to form a global decision.

[0197] Thus, by using the LoRA low-rank adaptation mechanism, each reasoning chain node and the main decision branch can have an independent fine-tuning channel, without the need for large model full parameter backpropagation, significantly reducing the fine-tuning parameter amount and engineering implementation difficulty, and supporting rapid updating and iterative modeling under limited computing power. In addition, each branch module is independently adapted, improving the discriminant sensitivity to local risk factors; the main module is globally attributed, improving the consistency and robustness of the comprehensive decision. In addition, the system can also construct new parallel branch modules based on new surrounding bid behaviors, and has good scalability.

[0198] In step S550, a joint loss function is constructed using the main task loss, the branch node loss, and the evidence correlation loss between the bid sample RAG knowledge features and the historical case description content, and the LoRA parameter adapters are jointly trained, so that the RAG-LLM system can simultaneously improve the accuracy of the surrounding bid risk level main label, each branch node sub-label, and the case knowledge reference.

[0199] The essence of building surrounding bid risk discrimination is a chain multi-objective optimization problem, which requires both global discrimination ability and the accuracy of risk judgment of each branch node and reference evidence. Through the joint optimization of the main task loss, the branch node loss, and the evidence correlation loss, the completeness and resolution of the chain reasoning can be ensured, and the reference and explanation ability of the model to business knowledge can be strengthened.

[0200] Specifically, the joint loss function is expressed by the following formula:

[0201] , formula (17)

[0202] In the formula, represents the main task loss, which is the cross-entropy loss of the main label of the bid risk level; represents the branch node loss, which is the average of the cross-entropy loss of all branch node sub-labels; is the evidence relevance loss, which is the semantic relevance loss between the bid sample RAG knowledge features and the historical case description content; is the loss weighting coefficient.

[0203] More specifically, the main task loss Multi-class cross-entropy is used to measure the error between the model's final output risk level and the real main label:

[0204] , formula (17)

[0205] In the formula, represents the model output main risk prediction, such as high, medium and low risk levels; represents the real main label.

[0206] Each branch loss uses binary or multi-class cross-entropy respectively:

[0207] , formula (18)

[0208] In the formula, is the number of branches (such as price, time sequence, project), for example, 3. represents the branch prediction, represents the real label of the branch.

[0209] Evidence relevance loss BERTScore or sentence vector cosine similarity can be used to encourage the model to actively quote relevant historical case content in the chain reasoning output:

[0210] , formula (18)

[0211] In the formula, represents the model output reasoning chain text segment, represents the relevant historical case paragraph.

[0212] Through the fine-tuning method based on LoRA, combined with the main task loss, branch node loss and evidence correlation loss, the RAG-LLM system can improve the comprehensive risk level discrimination, branch risk type discrimination and reference ability to historical case knowledge through end-to-end joint optimization. Thus, the multi-factor discrimination, chain reasoning and evidence-based attribution intelligent risk control are realized, so that the RAG-LLM system can realize more fine and interpretable risk identification output in the actual engineering bidding business, and the intelligent supervision level of engineering bidding is improved.

[0213] In order to verify the superiority of the "micro-error-time-cross-project" three-dimensional chain risk identification scheme proposed in the embodiments of the present application over the traditional method and other single-dimensional methods, the following comparative experiments are designed.

[0214] Regarding the selection of the data set, 15 large infrastructure project bidding records of a provincial public resource trading platform during 2019-2022 can be selected, a total of 100,000 bidding streams, including 500 real supervision confirmed cases of bid rigging / collusion and a number of normal cases.

[0215] Regarding the selection of evaluation indexes, the following is selected:

[0216] Accuracy: overall risk discrimination accuracy;

[0217] F1 score: taking into account the recall of bid rigging cases and the false alarm of normal cases;

[0218] Explanation coverage: the overlap rate of model output associated case paragraphs and expert annotated paragraphs, used to measure auditability;

[0219] Average reasoning delay: time required from input to output, used to evaluate system real-time performance.

[0220] Table 1: Experimental methods and comparison schemes

[0221]

[0222] Baseline A: Common global threshold method, comparing each of the full bid segment price, full period time interval and full project frequency with a fixed threshold.

[0223] Baseline B / C / D: Verify the effect gain of single-dimensional method and RAG-LLM respectively.

[0224] This scheme E: Parallel input of three-dimensional features into corresponding branches, and then comprehensive determination by the trunk module under the driving of chain prompt words, and output of knowledge chain.

[0225] Table 2: Experimental results

[0226]

[0227] From the experimental results of Table 2 above, it can be seen that the global threshold (A) has low precision and recall, and cannot capture local or cross-project hidden patterns. The single-dimensional method (B / C) has slightly improved in its own dimension, but still lacks multi-factor fusion capability. In the price dimension + RAG-LLM (D), the accuracy is greatly improved by introducing knowledge enhancement, but the explanation coverage is insufficient. The present scheme (E) improves the accuracy by about 7% and the F1 score by about 8% through three-dimensional parallel branches and chain prompts, and the explanation coverage reaches 89.3%, which is significantly better than each baseline; although the average reasoning time delay increases, it still meets the actual risk warning needs.

[0228] Figure 6 An example of an experimental comparison effect diagram of Accuracy (accuracy) and F1 Score (F1 score) is shown.

[0229] As shown in Figure 6 , by comparing the experimental results of the five methods (A~E) on the same test set, the horizontal axis represents the method number, and the vertical axis represents the detection accuracy (Accuracy) and F1 score, respectively. It can be seen that the present scheme (E) achieves the highest value in the two indicators, which is significantly better than other comparison schemes.

[0230] It should be noted that, for the foregoing method embodiments, in order to simply describe, they are all described as a series of actions, but those skilled in the art should know that the present application is not limited to the order of the actions described, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification all belong to preferred embodiments, and the actions and modules involved are not necessarily required by the present application. In the above embodiments, the description of each embodiment is focused on, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0231] Figure 7 A structure block diagram of an example of a construction engineering bid risk identification system based on RAG-LLM according to an embodiment of the present application is shown.

[0232] As shown in Figure 7 , the construction engineering bid risk identification system based on RAG-LLM 700 includes a multi-source data acquisition unit 710, a comprehensive feature extraction unit 720, and a RAG-LLM risk analysis unit 730.

[0233] The multi-source data collection unit 710 is configured to collect bidding multi-source data of a target bidder to be analyzed from a multi-source data platform; the multi-source data includes bidding price sequences of the target bidder in each bid section of a target project and a bidding submission timestamp sequence composed of timestamps of each submission of a price, and the multi-source data further includes cross-project participation records of the target bidder.

[0234] The comprehensive feature extraction unit 720 is configured to extract bidding comprehensive features from the bidding multi-source data; the bidding comprehensive features include: a differential price abnormality corresponding to the bidding price sequence, a wrong time sequence abnormality corresponding to the bidding submission timestamp sequence, and a cross-project activity degree corresponding to the cross-project participation records.

[0235] The RAG-LLM risk analysis unit 730 is configured to input the bidding comprehensive features into a RAG-LLM system, to recall, by RAG, a set of bid-rigging knowledge paragraphs matched with the bidding comprehensive features from a bid-rigging case knowledge base, and to output, by a preset prompt word driven LLM, a corresponding construction engineering bid-rigging risk level.

[0236] Figure 8 A system framework schematic diagram of an example of a construction engineering bid-rigging risk identification system based on RAG-LLM according to an embodiment of the application is shown.

[0237] As shown in the system architecture, the four layers from bottom to top are: a multi-source data collection layer, a knowledge base and localized large model layer, a risk business process layer, and a function access and result display layer. Figure 8

[0238] In the multi-source data collection layer, the system automatically collects various formats of bidding related data from enterprise internal relational databases (through SQL interface), third party public interfaces (such as RESTful API) and Internet public pages (through crawler grabbing), including: bidding price, submission log, project meta information, contract text, supervision announcement, public opinion report, etc. For structured table data, missing value imputation, field standardization and One-Hot encoding are performed; for unstructured documents, PDF, pictures, etc., OCR+NLP technology is used for text extraction and semantic analysis. The processed raw data is temporarily stored in a data lake for unified scheduling and cleaning by the lower layer.

[0239] In the knowledge base and localized large model layer, this layer is divided into two parts:

[0240] ​1) Knowledge base construction: Multi-source cleaned data is classified and indexed according to themes such as "enterprise dimension", "contract dimension", "bid dimension", and "policy public opinion dimension". Text slices are converted to vectors using embedding models (such as industry-adapted BERT or specialized vectorization tools), and stored in the enterprise vector knowledge base. This knowledge base provides efficient context completion support for subsequent RAG retrieval.

[0241] 2) Localized RAG-LLM fine-tuning: Based on the previously designed chain prompt framework and LoRA adapter mechanism, fine-tune the large model to the enterprise business context. After inputting the bid features (micro-differential bidding, timing anomalies, cross-project activity) and retrieved case vectors, the model performs branch reasoning and generates risk conclusions and reference evidence.

[0242] In the risk business process layer, it is used to embed AI capabilities into key steps of the bidding process:

[0243] Requirement analysis: According to the tenderer's demand text, the model automatically identifies potential risk points such as unreasonable clauses or suspicious scoring rules.

[0244] Tender document preparation: When preparing the announcement and bid evaluation method, the model gives risk prompts and recommends compliance wording.

[0245] Bid pre-audit and auxiliary evaluation: Intelligent verification of bid eligibility, bidding strategy, and bid scoring, the model combines historical cases to identify risks.

[0246] Contract risk assessment: Before contract negotiation and signing, the model analyzes clause vulnerabilities and credit risks, and outputs review comments.

[0247] Each step calls the localized RAG-LLM, pulls the knowledge base cases and tool functions (such as document generation, clause comparison) in real time, and writes the analysis results back to the unified risk management platform.

[0248] In the functional access and result display layer, it faces business users and provides two sets of unified interfaces for PC and mobile. At any node in the process, the intelligent assistant can be awakened, and the user can ask questions in natural language, and the large model returns:

[0249] Risk report: structured table / card form to display risk level, core factors, and reference case paragraphs;

[0250] Decision suggestion: provide revision schemes or compliance measures for each risk point;

[0251] Visual audit: clicking on the case reference can jump to the original clause or regulatory document, supporting full-link traceability.

[0252] Thus, through the synergy of the above four-layer architecture, the data correlation, knowledge retrieval and model reasoning of each link of the system are closely linked, realizing the full-link, interpretable and auditable intelligent bid risk identification and management from multi-source data ingestion to intelligent decision support.

[0253] In some embodiments, the embodiments of the present application provide a non-volatile computer readable storage medium, wherein one or more programs including execution instructions are stored in the storage medium, the execution instructions can be read and executed by an electronic device (including but not limited to a computer, a server, or a network device, etc.) to execute the steps of any one of the above RAG-LLM-based construction engineering bid risk identification methods.

[0254] In some embodiments, the embodiments of the present application also provide a computer program product, which includes a computer program stored on a non-volatile computer readable storage medium, the computer program includes program instructions, when the program instructions are executed by a computer, the computer executes the steps of any one of the above RAG-LLM-based construction engineering bid risk identification methods.

[0255] In some embodiments, the embodiments of the present application also provide an electronic device, which includes at least one processor, and a memory connected to the at least one processor in communication, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the steps of the RAG-LLM-based construction engineering bid risk identification method.

[0256] The above products can execute the methods provided by the embodiments of the present application, and have the corresponding function modules and beneficial effects of executing the methods. Technical details not described in detail in the embodiments can be referred to the methods provided by the embodiments of the present application.

[0257] The electronic device of the embodiments of the present application exists in various forms, including but not limited to: a mobile communication device, an ultra-mobile personal computer device, a portable entertainment device, or other onboard electronic devices with data interaction function.

[0258] The device embodiments described above are only schematic, wherein the units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, that is, they can be located in one place, or distributed on multiple network units. According to actual needs, part or all of the modules can be selected to achieve the purpose of the embodiments of the present application.

[0259] Those skilled in the art can clearly understand the implementation of the various embodiments by means of software plus a general hardware platform from the above description of the embodiments, and of course, the various embodiments can also be implemented by hardware. Based on such an understanding, the above technical solutions, essentially or in other words, the part that contributes to the related art, can be embodied in the form of a software product. The computer software product can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, and the like, and includes a plurality of instructions to cause a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments or some parts of the methods.

[0260] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present application.

Claims

1. A RAG-LLM-based construction project bid rigging risk identification method, characterized in that, The method comprises: Collecting bidding multi-source data of a target bidder to be analyzed from a multi-source data platform; the multi-source data comprises bidding bid sequences of the target bidder in each bid section of a target project and bidding submission timestamp sequences composed of timestamps of each bid submission, and the multi-source data further comprises cross-project participation records of the target bidder; Extracting bidding comprehensive features from the bidding multi-source data; the bidding comprehensive features comprise: differential bid abnormality corresponding to the bidding bid sequences, out-of-time sequence abnormality corresponding to the bidding submission timestamp sequences, and cross-project activity corresponding to the cross-project participation records; Wherein, the extraction of the cross-project activity comprises: Constructing a three-dimensional project participation tensor of the target bidder, wherein the three dimensions of the three-dimensional project participation tensor are bidder, project participation, and time slice of project participation; Tensor decomposition is performed on the three-dimensional project participation tensor to obtain a bidder factor matrix to calculate the cross-project activity; Wherein, the tensor decomposition of the three-dimensional project participation tensor to obtain the bidder factor matrix to calculate the cross-project activity comprises: Based on the non-negative block tensor decomposition model, the three-dimensional project participation tensor is decomposed to obtain the bidder factor matrix to calculate the cross-project activity, and the operation specifically comprises: A three-dimensional project participation tensor is divided into groups according to the number of bidders is divided into non-overlapping sub-tensor blocks : ， In the formula, denotes the original participation tensor, denotes the total number of bidders, denotes the total number of projects, denotes the total number of time slices, and are the number of bidder groups and the number of project categories, respectively, denotes the group of bidders and the category of projects corresponding to the sub-tensor, is the number of bidders in the group, is the number of projects in the category. Performing two-stage non-negative decomposition on each sub-tensor block to extract the bidder-project-time three-dimensional mode factor: In the temporal dimension decomposition of the first stage, for each sub-tensor block Direct CP decomposition: ， wherein is the CP decomposition rank, indicating how many groups of factors are extracted; is the weight of the group of factors, is the bidder factor vector, is the project factor vector, is the time factor vector and , is the vector outer product operation; In the spatial dimension decomposition of the second stage, the factor set after CP decomposition is regarded as a tensor component, and non-negative spatial decomposition is directly performed as follows: ， In the formula, , denotes the bid-subspace basis matrix, denotes the project-subspace basis matrix, denotes the core interaction matrix, denotes the spatial decomposition rank; thus, for the bid and project factor matrices, further non-negative Tucker decomposition is imposed, the CP decomposition factors are compressed in the subspace, similar bidders or projects are mapped to the same subspace, and the group behavior pattern aggregation is realized; Constructing a spatiotemporal activity matrix of each sub-block: ， In the formula, denotes the spatio-temporal activity of a sub-block, denotes a matrix transposition operation, , denotes the spatio-temporal activity of a sub-block; By introducing a time decay factor to solve the spatio-temporal activity matrix, the target bidder is obtained In the time slice Cross-project activity : ， wherein is a matrix in the row of column element, is a time decay coefficient, is the current analysis time point, denotes a time decay factor; thereby, using decay weights to fade out far- past participation step by step, focusing the analysis on the spatio-temporal activities close to the current review time point; aggregating the cross-project activity of each time slice to obtain a cross-project activity of the target bidder :​ ， wherein denotes a time weighting function, is a weighting function slope parameter, is a weighting function center offset time, denotes a target bidder across project activity in a time slice across project activity; The bidding comprehensive features are input into the RAG-LLM system to recall a set of surrounding bid knowledge paragraphs matching the bidding comprehensive features from a surrounding bid case knowledge base through RAG, and a pre-set prompt word is used to drive the LLM to output a corresponding construction engineering surrounding bid risk level; Wherein, the recalling of the set of surrounding bid case paragraphs matching the bidding comprehensive features from the surrounding bid case knowledge base through RAG comprises: Respectively for the differential bid abnormality, the out-of-time sequence abnormality, and the cross-project activity in the bidding comprehensive features, the top K case knowledge paragraphs with the highest matching degree are retrieved from the surrounding bid case knowledge base; the surrounding bid case knowledge base pre-stores a plurality of surrounding bid case knowledge paragraphs, each surrounding bid case knowledge paragraph is indexed by knowledge processing metadata, and the indexing metadata of the case knowledge paragraph is constructed according to one or more of the differential bid abnormality, the out-of-time sequence abnormality, and the cross-project activity; The retrieved case knowledge paragraphs are de-duplicated and combined to obtain a set of matching surrounding bid knowledge paragraphs; the total number of paragraphs in the knowledge paragraph set is less than or equal to 3K.

2. The method of claim 1, wherein, For the extraction of the differential bid abnormality, comprising: Performing sliding window segmentation on the bidding bid sequences; Calculating the standardized value of the bidding bid data in each window; The distribution pattern of the window standardized value sequence is analyzed by a lightweight neural network model to output the corresponding differential bid abnormality.

3. The method of claim 1, wherein, For the extraction of the out-of-time sequence abnormality, comprising: compute interval values of adjacent timestamps in the sequence of the bid submission timestamps to obtain a corresponding sequence of interval values; reconstruct the sequence of interval values based on a self-encoder model to obtain a corresponding reconstruction error; the self-encoder model is pre-trained according to a sample set of normal patterns; determine a time sequence anomaly degree according to the reconstruction error.

4. The method of claim 1, wherein, The pre-set prompt word adopts a chain reasoning instruction structure, and the RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; 5. The method of claim 4, wherein, The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG-LLM system comprises a trunk module and a plurality of branch reasoning modules; The RAG- A joint loss function is constructed by using a main task loss, a branch node loss, and an evidence correlation loss between a bid sample RAG knowledge feature and the historical case description content, the LoRA parameter adapter is jointly trained, and the RAG-LLM system can improve the accuracy of the main label of the bid risk level, the sub-label of each branch node, and the case knowledge reference at the same time; The joint loss function is expressed by the following formula: ， In the formula, represents the main task loss, which is the cross-entropy loss of the main label of the surrounding risk level; represents the branch node loss, which is the average of the cross-entropy losses of all branch node sub-labels; is the evidence correlation loss, which is the semantic correlation loss between the bid sample RAG knowledge features and the historical case description content; is the loss weighting coefficient.

6. A RAG-LLM based construction project bid rigging risk identification system, characterized in that, The system is used for implementing the method according to any one of claims 1-5; the system comprises: A multi-source data acquisition unit is configured to acquire bid multi-source data of a target bidder to be analyzed from a multi-source data platform; the multi-source data includes bid price sequences of the target bidder in each bid section of a target project and a bid submission timestamp sequence composed of a timestamp of each submitted bid price, and the multi-source data further includes cross-project participation records of the target bidder; An integrated feature extraction unit is configured to extract bid integrated features from the bid multi-source data; the bid integrated features include: a differential bid price abnormality corresponding to the bid price sequence, a mistimed sequence abnormality corresponding to the bid submission timestamp sequence, and a cross-project activity level corresponding to the cross-project participation records; A RAG-LLM risk analysis unit is configured to input the bid integrated features into a RAG-LLM system, to recall, by the RAG, a set of bid knowledge paragraphs matching the bid integrated features from a bid case knowledge base, and to output, by the LLM driven by a preset prompt word, a corresponding construction project bid risk level.

Images