Big model-based building engineering business opportunity information retrieval agent system and method

Abstract

The application relates to the technical field of building engineering information retrieval, and discloses a building engineering business opportunity information retrieval intelligent agent system and method based on a large model. The system comprises information collection, intention mapping, mapping relationship construction, business opportunity matching, dynamic updating and retrieval management modules. The information collection module obtains multi-modal business opportunity information from multiple heterogeneous data sources, generates a business opportunity feature warehouse through integration and feature extraction; the intention mapping module receives a user retrieval request, analyzes semantic intention by using a large language model to generate an intention feature vector, and preliminarily matches in the business opportunity feature warehouse; the mapping relationship construction module dynamically maps the intention and the business opportunity feature space according to the correlation strength; the business opportunity matching module calculates the semantic correlation degree by combining the relationship and the time weight, and generates a business opportunity matching sequence; the dynamic updating module monitors data changes, updates the warehouse and adjusts the mapping relationship; and the retrieval management module generates a hierarchical retrieval strategy, constructs a business opportunity retrieval knowledge base, and meets the retrieval requirements of building enterprises.

Description

Technical Field

[0001] This invention relates to the field of construction engineering information retrieval technology, specifically to a large-model-based intelligent system and method for retrieving business opportunities in construction engineering. Background Technology

[0002] In the construction engineering sector, the acquisition and accurate matching of business opportunities are crucial prerequisites for enterprises to conduct project planning, resource allocation, and market competition. Currently, the distribution of business opportunities in the construction engineering sector exhibits significant heterogeneity. This information encompasses structured data such as bidding announcements and government procurement information from government public resource trading platforms, as well as semi-structured text such as project updates and enterprise cooperation intentions published on industry portals, and unstructured data such as engineering drawings and project site images. The significant differences in information formats, update frequencies, and core fields across different data sources make information integration extremely difficult. Traditional manual collection and processing methods not only consume substantial manpower and time but are also prone to data omissions or delays, impacting the efficiency of opportunity capture.

[0003] Users' needs for business opportunity information are becoming increasingly diverse and personalized. Project decision-makers in construction companies may need to search for business opportunities in specific regions or for specific types of projects (such as municipal engineering and building construction), while technical personnel are more concerned with project information related to specific construction techniques or materials. Traditional search systems rely heavily on keyword matching mechanisms, which struggle to accurately interpret the deep semantic intent in users' search requests, often resulting in problems such as "redundant search results" or "mismatch between needs and results."

[0004] Business opportunities in the construction engineering sector are highly time-sensitive. The speed at which key information such as project bidding announcements, prequalification dates, and bid deadlines are updated directly impacts a company's chances of participating. Traditional retrieval systems often rely on periodic batch updates, making it difficult to monitor dynamic changes in data sources in real time. This frequently leads to information lag, causing users to miss important business opportunities. Furthermore, as time and market conditions change, user search preferences and the weighting of business opportunity features also evolve. Traditional systems lack the ability to dynamically adjust the mapping between intent space and business opportunity feature space, failing to continuously adapt to user needs and market changes, further reducing the accuracy and effectiveness of business opportunity retrieval. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent agent system and method for retrieving business opportunities in construction projects based on a large model, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides an intelligent agent system for retrieving business opportunities in construction projects based on a large model, the system comprising: The information acquisition module is used to acquire multimodal business opportunity information in the field of construction engineering from multiple heterogeneous data sources, integrate the multimodal business opportunity information and extract its features to generate a business opportunity feature warehouse; The intent mapping module is used to receive the user's input search request, use a large language model to parse the semantic intent of the search request, generate an intent feature vector, and perform preliminary matching in the business opportunity feature warehouse based on the intent feature vector; The mapping relationship construction module is used to construct a dynamic mapping relationship between the intent space and the opportunity feature space based on the correlation strength between the intent feature vector and the opportunity features in the opportunity feature warehouse. The business opportunity matching module is used to calculate the semantic relevance between each business opportunity information in the business opportunity feature warehouse and the intent feature vector according to the dynamic mapping relationship, and generate a business opportunity matching sequence by combining the time weight of the business opportunity information; The dynamic update module is used to monitor data changes in business opportunity information, update the business opportunity feature warehouse according to the newly acquired business opportunity information, and adaptively adjust the dynamic mapping relationship between the intent space and the business opportunity feature space. The retrieval management module is used to generate hierarchical retrieval strategies and build a business opportunity retrieval knowledge base based on the updated business opportunity feature repository.

[0007] Preferably, the information acquisition module is implemented in the following ways: The multimodal business opportunity information is preprocessed to extract text features, image features, and numerical features from the multimodal business opportunity information; The text features, image features, and numerical features are fused in a multimodal manner to generate a unified representation of the business opportunity feature vector; Based on the source and type of the business opportunity feature vectors, the business opportunity feature vectors are classified and stored to form the business opportunity feature warehouse.

[0008] Preferably, the intent mapping module is implemented in the following ways: The intent of the retrieval request is identified using a large language model, and the intent category and confidence level of the retrieval request are output. Based on the intent category and confidence level, select the corresponding intent descriptor from the predefined intent dictionary; The intent descriptor is encoded into a high-dimensional vector to form the intent feature vector.

[0009] Preferably, the mapping relationship construction module is implemented in the following ways: Calculate the cosine similarity between the intent feature vector and each opportunity feature vector in the opportunity feature warehouse; A subset of candidate business opportunity features is obtained based on the cosine similarity filtering. Based on the distribution relationship between the candidate business opportunity feature subset and the intent feature vector, a dynamic mapping relationship between the intent space and the business opportunity feature space is constructed.

[0010] Preferably, the business opportunity matching module is implemented in the following ways: Based on the dynamic mapping relationship, obtain business opportunity information related to the intent feature vector from the business opportunity feature warehouse; Extract the time attribute of each business opportunity and calculate its time decay coefficient relative to the current time; The business opportunity matching sequence is generated by combining the semantic relevance and the time decay coefficient.

[0011] Preferably, the dynamic update module is implemented in the following ways: New business opportunity information is collected in real time, its business opportunity feature vector is extracted and added to the business opportunity feature warehouse; Detect changes in the status of business opportunity information in the business opportunity feature warehouse, and mark and remove expired business opportunity information; The dynamic mapping relationship between the intent space and the opportunity feature space is recalculated based on the updated opportunity feature repository.

[0012] Preferably, the system further includes: The feature importance filtering module is used to analyze the correlation strength between each business opportunity feature in the business opportunity feature warehouse and the historical retrieval success rate, and calculate the feature importance entropy; based on the feature importance entropy, key business opportunity features are filtered to form a core feature sequence; After generating the intent feature vector, the intent mapping module prioritizes the initial matching with the core feature sequence.

[0013] Preferably, the system further includes: The feedback learning module is used to collect user interaction behavior with search results and generate feedback signals; The matching weights between the intent recognition strategy in the intent mapping module and the business opportunity matching module are adjusted based on the feedback signal.

[0014] Preferably, the multimodal business opportunity information in the information collection module includes, but is not limited to, tender announcements, winning bid announcements, equipment and service procurement lists, engineering information, project planning documents, engineering drawings, supplier qualification documents, and market analysis reports.

[0015] Preferably, the present invention also includes a method for constructing an intelligent agent for retrieving business opportunities in construction projects based on a large model, applied to the aforementioned intelligent agent system for retrieving business opportunities in construction projects based on a large model, the method comprising the following steps: Step 1: Obtain multimodal business opportunity information in the construction engineering field from multiple heterogeneous data sources, integrate the multimodal business opportunity information and extract its features to generate a business opportunity feature warehouse; Step 2: Receive the user's search request, use a large language model to parse the semantic intent of the search request, generate an intent feature vector, and perform preliminary matching in the business opportunity feature warehouse based on the intent feature vector; Step 3: Based on the correlation strength between the intent feature vector and the opportunity features in the opportunity feature warehouse, construct a dynamic mapping relationship between the intent space and the opportunity feature space; Step 4: Based on the dynamic mapping relationship, calculate the semantic relevance between each business opportunity information in the business opportunity feature warehouse and the intent feature vector, and generate a business opportunity matching sequence by combining the time weight of the business opportunity information; Step 5: Monitor data changes in business opportunity information, update the business opportunity feature warehouse based on newly acquired business opportunity information, and adaptively adjust the dynamic mapping relationship between the intent space and the business opportunity feature space; Step 6: Based on the updated business opportunity feature repository, generate a hierarchical retrieval strategy and build a business opportunity retrieval knowledge base.

[0016] Compared with the prior art, the beneficial effects of the present invention are: By establishing a multi-module collaborative working mechanism, this solution effectively addresses numerous issues in the current process of retrieving business opportunities in the construction industry, providing construction companies with a more efficient and accurate solution for acquiring business opportunities. In the information collection phase, the module can acquire multimodal business opportunity information from multiple heterogeneous data sources and generate a business opportunity feature warehouse through data integration and feature extraction. This breaks down the fragmentation of information from different data sources under traditional manual collection methods, avoids integration difficulties caused by data format differences, and reduces the investment of manpower and time costs. It ensures that companies can obtain more comprehensive and complete business opportunity information, eliminating the risk of missing potential opportunities due to data omissions or format incompatibility.

[0017] At the user intent parsing and initial matching level, the intent mapping module leverages the semantic understanding capabilities of a large language model to deeply analyze the underlying intent in user search requests, generating accurate intent feature vectors. Based on these vectors, initial matching is then performed in a business opportunity feature repository. This process overcomes the limitations of traditional keyword matching mechanisms, moving beyond reliance on surface-level text for matching. Instead, it accurately identifies key elements of user needs, such as project type, cooperation mode, and geographical scope, thereby reducing interference from irrelevant information and allowing users to quickly access business opportunities highly relevant to their needs, improving search efficiency and user experience.

[0018] The mapping relationship construction module builds a dynamic mapping relationship between the intent space and the opportunity feature space based on the correlation strength between intent feature vectors and opportunity features. This design enables the system to continuously adapt to changes in user needs and opportunity information features. As the market environment evolves, the importance of different opportunity features and users' preferences for specific needs may change. The dynamic mapping relationship can reflect these changes in a timely manner, avoiding the "mismatch between needs and results" problem caused by fixed mapping relationships in traditional systems, and ensuring that the system maintains high matching accuracy regardless of market changes.

[0019] Based on dynamic mapping relationships, the opportunity matching module generates an opportunity matching sequence by incorporating time weights into the opportunity information. This considers both the semantic relevance between user intent and opportunity information, as well as the timeliness of the opportunity information. In the construction engineering field, time factors such as bid deadlines and prequalification periods directly affect a company's participation opportunities. The introduction of time weights allows for the priority presentation of more timely opportunity information to users, preventing users from missing key deadlines due to unreasonable information sorting, and further ensuring that companies can promptly seize important business opportunities.

[0020] The dynamic update module monitors changes in business opportunity information in real time, promptly updating the business opportunity feature repository with newly acquired information. Simultaneously, it adaptively adjusts the dynamic mapping between the intent space and the business opportunity feature space. This mechanism solves the problem of lagging information updates in traditional systems, ensuring that the information in the business opportunity feature repository is always up-to-date. It also allows the mapping relationship to be optimized with the addition of new information, avoiding a decrease in search accuracy due to untimely information updates or fixed mapping relationships. This enables the system to continuously meet user needs and adapt to market changes.

[0021] The search management module generates hierarchical search strategies and builds a business opportunity search knowledge base based on the updated business opportunity feature repository, providing users with more targeted search solutions. The hierarchical search strategies can provide personalized search paths based on the different needs of different user groups (such as the regions and project scales that decision-makers focus on, and the construction techniques that technical personnel focus on). The construction of the business opportunity search knowledge base further integrates historical search data and high-quality business opportunity information, providing richer reference for subsequent searches, helping users find target information faster, and improving overall search efficiency and experience. Attached Figure Description

[0022] Figure 1 This is a sequence diagram of the intelligent agent system for retrieving business opportunities in construction projects based on a large model, as described in this invention. Figure 2 A flowchart illustrating the implementation of the information acquisition module; Figure 3 A flowchart for the implementation of the intent mapping module; Figure 4 Flowchart for the business opportunity matching module; Figure 5 This is a flowchart illustrating the implementation of the feature importance filtering module. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Please see Figure 1 This invention provides an intelligent agent system for retrieving business opportunities in construction projects based on a large model. The system includes: The information acquisition module acquires multimodal business opportunity information in the construction engineering field from multiple heterogeneous data sources, including bidding announcements, project planning documents, engineering drawings, supplier qualification documents, and market analysis reports. This information is then integrated and its features extracted to generate a business opportunity feature warehouse. The intent mapping module receives user-input search requests, uses a large language model to parse the semantic intent, generates intent feature vectors, and performs preliminary matching within the business opportunity feature warehouse based on these vectors. The mapping relationship construction module constructs a dynamic mapping relationship between the intent space and the business opportunity feature space based on the correlation strength between the intent feature vectors and the business opportunity features in the warehouse. The business opportunity matching module calculates the semantic relevance between each business opportunity in the warehouse and the intent feature vector based on the dynamic mapping relationship, and generates a business opportunity matching sequence by incorporating the time weight of the business opportunity information. The dynamic update module monitors changes in business opportunity information, updates the business opportunity feature warehouse based on newly acquired information, and adaptively adjusts the dynamic mapping relationship between the intent space and the business opportunity feature space. The retrieval management module generates a hierarchical retrieval strategy based on the updated business opportunity feature warehouse and constructs a business opportunity retrieval knowledge base. The entire system achieves efficient and accurate business opportunity information retrieval through modular collaboration, adapting to the dynamic needs of the construction engineering field.

[0025] Example 1: See Figure 2The information collection module acquires multimodal business opportunity information in the construction engineering field from multiple heterogeneous data sources, including bidding announcements, project planning documents, engineering drawings, supplier qualification documents, and market analysis reports. These data sources include government bidding platforms, industry databases, publicly available enterprise information databases, and materials provided by professional service organizations. Bidding announcements typically contain structured text information such as project name, bidding entity, project overview, budget amount, and bid deadline, and may also include drawings or technical specifications as attachments. Project planning documents involve more detailed project descriptions, including project background, construction goals, technical indicators, implementation plans, and expected results, often presented in lengthy text and numerical tables. Engineering drawings, as visual information, include floor plans, elevations, sections, and various detailed drawings, reflecting the project's spatial layout, structural form, and construction requirements. Supplier qualification documents include business licenses, qualification certificates, performance records, and financial reports, typically in scanned or structured data format. Market analysis reports provide macro-level information such as industry trends, competitive landscape, and material price fluctuations, combining textual descriptions and statistical data.

[0026] These multimodal business opportunity information are preprocessed to extract textual, image, and numerical features. Textual feature extraction begins with cleaning and standardizing the original text, removing irrelevant characters, unifying the encoding format, and performing word segmentation. For the text content in tender announcements and project planning documents, natural language processing techniques are used for word embedding and semantic parsing. A pre-trained language model converts the text into a high-dimensional vector representation, capturing key information such as project type, scale requirements, and technical characteristics. Image feature processing targets scanned copies of engineering drawings and qualification documents. Convolutional neural networks are used to extract visual elements, identifying graphic symbols, annotations, and features such as seals and signatures in the documents, converting them into feature vectors. Numerical feature extraction primarily targets structured data such as budget amounts, project durations, and registered capital. Statistical methods are used for normalization to eliminate the influence of dimensions, converting the data into numerical feature vectors.

[0027] This project fuses textual, image, and numerical features using a multimodal approach, employing an attention mechanism for weighted integration to generate a unified business opportunity feature vector. The attention mechanism dynamically adjusts the weights based on the importance of different features; for example, textual features may have higher weights for tender notices, while image features may have higher weights for engineering drawings. The fusion process ensures consistency across modal information within the vector space, forming a comprehensive feature representation that reflects the business opportunity content. Based on the source and type of the business opportunity feature vectors, they are categorized and stored to form a business opportunity feature repository. This categorization is based on the attributes of the data source and the type of business opportunity; for example, tender notices are categorized by project size and region, engineering drawings by project type and design stage, and supplier qualification documents by qualification level and industry certification. The business opportunity feature repository uses a distributed database architecture, supporting efficient read / write and expansion, and ensuring data consistency and integrity. Preprocessing also includes data cleaning and deduplication to remove redundant and invalid information, improving the accuracy of feature extraction. During multimodal fusion, feature alignment technology is used to ensure consistency across different modalities in the vector space, enhancing the representational power of the business opportunity feature vectors. The business opportunity feature repository is updated through scheduled tasks and real-time monitoring to ensure the timely integration of new data. The implementation of the entire information collection module guarantees the comprehensiveness and accuracy of business opportunity information, providing a high-quality data foundation for subsequent retrieval.

[0028] Taking a specific business opportunity information processing process as an example, when the system receives a new bidding announcement for a highway construction project, the information acquisition module first parses the announcement text, extracting textual and numerical information such as the project name, bidding unit, budget amount, and construction period requirements. Simultaneously, the accompanying engineering drawings are parsed, extracting image features such as route design and bridge structure. These features are then fused using multimodal methods to generate a unified business opportunity feature vector, which comprehensively reflects the project's key attributes. Based on project type and region, this feature vector is categorized and stored in the "Transportation Infrastructure - Under Bidding" category of the business opportunity feature repository. The entire processing is automated, requiring no manual intervention, demonstrating the system's efficiency and accuracy.

[0029] Example 2: See Figure 3The intent mapping module receives user-input search requests and uses a large language model to parse their semantic intent, generating intent feature vectors. For example, if a user inputs "find recent bidding information for large bridge construction projects in the Yangtze River Delta region," this search request first undergoes text preprocessing, including word segmentation, stop word removal, and grammatical correction. The large language model, based on a deep learning architecture, analyzes the semantic content of the request, identifying core intent elements including project type (bridge construction), geographical scope (Yangtze River Delta region), project scale (large), and time requirement (recent). The model outputs the intent category as "project bidding query" and provides a confidence score. This confidence score reflects the model's assessment of the accuracy of intent recognition, based on the clarity and standardization of the request expression.

[0030] Based on the identified intent category and confidence level, the system selects the corresponding intent descriptor from a predefined intent dictionary. The intent dictionary contains standardized expressions of professional terms in the construction engineering field; for example, "bridge construction" is mapped to "bridge_construction," "Yangtze River Delta region" to "yangtze_river_delta," and "large-scale project" to "large_scale." These descriptors have been reviewed by domain experts to ensure semantic accuracy and consistency. The selection process also considers the impact of confidence level; when the confidence level is high, the matching descriptor is used directly, while when the confidence level is low, fuzzy matching or a clarification request mechanism may be triggered.

[0031] The selected intent descriptors are encoded into high-dimensional vectors, forming intent feature vectors. The encoding process uses a neural network encoder trained on a large amount of text in the construction engineering field, which can convert text descriptors into numerical representations with rich semantic information. The generated intent feature vectors not only contain literal meanings but also capture contextual relationships and domain knowledge, such as the potential connections between "bridge construction" and concepts like "steel structure" and "prestressed concrete."

[0032] The mapping relationship construction module calculates the similarity metric between the intent feature vector and the feature vectors of each opportunity in the opportunity feature repository. This calculation is based on a vector space model, measuring the relative position and direction of two vectors in the semantic space. For a user's request to find bridge construction projects, the system calculates the similarity score between the intent feature vector and all opportunity feature vectors in the repository. The similarity score reflects the semantic relevance of each opportunity to the user's request; for example, tender notices related to bridge construction receive higher scores, while projects related to building construction or municipal roads receive lower scores.

[0033] A subset of candidate business opportunity features is obtained based on similarity scores. The selection process employs a dynamic threshold mechanism, with threshold settings taking into account both the confidence level of intent recognition and the explicitness of the user's request. For explicit requests with high confidence, a higher threshold is used to ensure accuracy; for vague or broad requests, the threshold is appropriately lowered to provide more relevant results. The candidate subset contains several business opportunity feature vectors with the highest scores, representing business opportunities that are most likely to meet the user's needs.

[0034] Based on the distribution relationship between candidate opportunity feature subsets and intent feature vectors, a dynamic mapping relationship between the intent space and the opportunity feature space is constructed. This process analyzes the clustering features and distribution patterns of opportunity feature vectors in the candidate subsets, establishing a mapping model from intent vectors to opportunity vectors. For example, for bridge construction requests, the system may discover that related opportunity feature vectors have similar values ​​in specific dimensions, thus establishing a dedicated mapping channel for this type of request. The dynamic mapping relationship is represented by an adjustable weight matrix, which records the response strength of different feature dimensions to various intents, enabling adaptive optimization based on real-time query results.

[0035] Throughout the implementation process, the system maintains the timeliness and adaptability of the mapping relationship. When new business opportunity information is added to the repository or user request patterns change, the mapping relationship is adjusted accordingly. This dynamic adjustment mechanism enables the system to continuously optimize search results and adapt to changes in user needs and industry dynamics. By establishing a precise intent-opportunity mapping relationship, the system can quickly locate relevant business opportunity information, laying the foundation for subsequent refined matching and ranking.

[0036] Taking a real-world application scenario as an example, when a user queries "smart building renovation projects in the Beijing-Tianjin-Hebei region," the intent mapping module first parses out the core elements: region (Beijing-Tianjin-Hebei) and project type (smart building renovation). After selecting standardized descriptors from the intent dictionary, an intent feature vector is generated. The system calculates the similarity between this vector and feature vectors in the business opportunity feature repository, filtering out renovation project information related to smart buildings. Simultaneously, the mapping relationship construction module analyzes the feature distribution of these candidate projects, discovering that these projects exhibit significant clustering characteristics in the technical feature dimension, thereby optimizing the mapping weight for this type of request. The entire process is completed within milliseconds, providing users with accurate initial matching results.

[0037] Example 3: See Figure 4The opportunity matching module retrieves opportunity information related to intent feature vectors from the opportunity feature repository based on dynamic mapping relationships. The dynamic mapping relationship provides an association matrix between intent vectors and opportunity vectors. This matrix is ​​calculated and maintained in real-time, recording the correspondence between different intent types and various opportunity features. When a user submits a search request, the system first generates intent feature vectors through the intent mapping module, then queries this association matrix to quickly locate all potentially relevant opportunity information in the opportunity feature repository. This relevant information constitutes an initial candidate set, containing a varying number of opportunity entries, each with a complete feature vector representation.

[0038] Extracting the time attributes of each business opportunity is a crucial step in the matching process. Time attributes include multiple dimensions such as publication time, last update time, and expiration date. The system standardizes these timestamps, converting them into a unified timestamp format, and calculates their time decay coefficient relative to the current time. The time decay coefficient is calculated using a time-series weighting model based on exponential decay, which reflects the characteristic that information value decreases over time. Newer business opportunities have higher initial weights, which gradually decay over time, but the decay rate varies depending on the type of business opportunity. For example, bidding information usually has a defined validity period and decays faster; while supplier qualification information has a longer validity period and decays relatively more slowly.

[0039] A business opportunity matching sequence is generated by combining semantic relevance and time decay coefficient. Semantic relevance is calculated using a vector space model to quantify the semantic similarity between the intent feature vector and each business opportunity feature vector. The time decay coefficient reflects the timeliness value of the business opportunity information. The two are combined using a weighted fusion method, with semantic relevance holding the main weight and the time decay coefficient serving as an adjustment factor. The final comprehensive score calculation formula is as follows: ; in: This indicates the final matching score for the business opportunity. Indicates semantic relevance score, This represents the difference between the current time and the time the business opportunity was posted. and These are the weighting coefficients for semantic relevance and time factor, respectively. This is the decay rate parameter. The weighting coefficient and decay parameter are set based on domain knowledge and historical data analysis; different parameter values ​​can be used for different business opportunity types. The system sorts all candidate business opportunities in descending order according to the comprehensive score, forming the final business opportunity matching sequence.

[0040] The dynamic update module collects new business opportunity information in real time, sourced from multiple heterogeneous data sources under continuous monitoring. The frequency of new information acquisition depends on the characteristics of the data source; for frequently updated bidding platforms, a collection interval of minutes is used; for less frequently changing qualification information databases, a collection cycle of days is used. Upon acquiring each new business opportunity, the system immediately initiates a feature extraction process, converting it into a standardized business opportunity feature vector and adding it to the business opportunity feature repository. The addition of new features triggers updates to the repository's index structure, ensuring that query efficiency is not affected.

[0041] Detecting changes in the status of business opportunity information in the business opportunity feature repository is an ongoing process. The system periodically scans all business opportunity information in the repository, checking their status identifiers and expiration information. For expired bidding information, supplier information with invalid qualifications, and project announcements marked as withdrawn, the system automatically updates their status to expired and excludes this information from subsequent searches. Status detection is based not only on explicit timestamp information but also on a comprehensive judgment by analyzing auxiliary information such as the update status of the information source website and related news reports. For expired business opportunity information, the system does not immediately delete it but first marks it for removal, retaining historical records for analysis and reference.

[0042] The dynamic mapping between the intent space and the opportunity feature space is recalculated based on the updated opportunity feature repository. This process involves updating the similarity matrix and adjusting the mapping weights. When new opportunity information is added, the system analyzes the distribution characteristics of the new feature vectors and assesses their impact on the existing mapping. For a large number of similar new pieces of information, the mapping may be reconstructed; for sporadic new information, an incremental update method is used to adjust the existing mapping weights. The recalculation process uses a distributed computing framework, decomposing the computational task into multiple nodes for parallel execution, ensuring computational efficiency under large data volumes. The updated mapping takes effect immediately, ensuring that subsequent retrieval requests can be processed based on the latest data status.

[0043] The entire business opportunity matching and dynamic update process forms a closed-loop system, continuously adapting to data changes and user needs. The generation of matching sequences considers not only static semantic relevance but also dynamic temporal characteristics, ensuring that the result sorting aligns with user intent while reflecting the freshness of the information. The dynamic update mechanism guarantees that the system always operates based on the latest and most accurate business opportunity information, avoiding the provision of outdated or invalid information. This design enables the system to maintain long-term accuracy and usability in real-world application environments, meeting the real-time requirements of business opportunity retrieval in the construction engineering field.

[0044] In practice, when a user queries "recent green building certification projects," the system first parses the core need through intent mapping, and then retrieves all green building-related project information from the business opportunity feature repository. For each matching project, its semantic relevance score and time decay coefficient are calculated. Recently released LEED certification projects receive a higher ranking due to their higher time decay coefficient, while similar projects released six months ago, although semantically similar, rank lower due to time decay. Simultaneously, the dynamic update module continuously monitors new green building project information. Once new relevant business opportunities are discovered, the feature repository is immediately updated and the mapping relationship adjusted to ensure that the latest information is included in subsequent queries. This entire process achieves a balance between accuracy and timeliness, providing users with the most valuable business opportunity matching results.

[0045] Example 4: See Figure 5 The feature importance filtering module analyzes the correlation strength between each opportunity feature in the opportunity feature repository and the historical search success rate. The system records user interaction data for each search operation, including search requests, returned results, user clicks, and subsequent business follow-ups. This historical data is anonymized and stored in a dedicated log database, forming a historical record of search behavior. By analyzing these records, the system can identify which opportunity features are more likely to lead to successful search matches. For example, analysis shows that project information containing a specific budget range receives more user attention than information with only an estimated amount, and tender documents with detailed technical specifications are more likely to match user needs than simply described documents.

[0046] The process of calculating feature importance entropy involves the application of information theory methods. The system calculates the distribution differences of each business opportunity feature across different search results, evaluating the feature's contribution to the discriminative power of the search results. A lower importance entropy value indicates a stronger predictive ability of the feature for the search results; a higher entropy value indicates lower discriminative power. During the calculation, the system considers the distribution of feature values, the strength of their association with user behavior, and the interrelationships between features. For example, the entropy value of the project's location feature is usually low because user searches often have a clear regional bias; while the entropy value of the project size feature may be high because different users have significant differences in their preferences for project size.

[0047] Key business opportunity features are selected based on feature importance entropy to form a core feature sequence. The system sets a dynamic threshold mechanism to automatically adjust the selection criteria according to the current volume and type of business opportunity data. For key features with entropy values ​​below the threshold, the system includes them in the core feature sequence. These features typically include highly structured information such as project type, budget amount, bidding status, and geographical location. The core feature sequence employs an optimized storage structure and a dedicated indexing mechanism to support fast query and matching operations. Features in the sequence are sorted by importance entropy value, with the most important features at the beginning of the sequence for priority processing.

[0048] Table 1: Importance Analysis of Business Opportunity Characteristics:

[0049] After generating the intent feature vector, the intent mapping module prioritizes preliminary matching with the core feature sequence. The matching process is first executed within a subset of core features, quickly narrowing the search scope by calculating the similarity between the intent feature vector and the core feature vector. This priority matching mechanism significantly reduces the amount of data that needs to be processed and improves search efficiency. For example, when a user queries "Beijing residential construction projects," the system first matches business opportunities with the geographical location "Beijing" and the project type "residential construction" in the core feature sequence, forming a preliminary candidate set.

[0050] The initial matching results serve as an initial candidate set for subsequent modules to refine. This initial candidate set contains business opportunities that are most likely to meet user needs; these opportunities have already been filtered at the core feature level and possess high relevance. Subsequent business opportunity matching modules, based on this initial candidate set, further consider other feature dimensions and time factors to generate the final matching sequence. This hierarchical processing mechanism ensures both retrieval accuracy and improves system response speed.

[0051] The feature importance filtering module periodically recalculates the feature importance entropy and updates the core feature sequence. The update frequency is dynamically adjusted based on data changes; a recalculation process is triggered when a certain number of new business opportunities are added, or when user search patterns change significantly. During the update process, the system evaluates the importance of new features. Newly emerging feature types are included in the evaluation scope after short-term observation and data accumulation. The update of the core feature sequence adopts a smooth transition method to avoid sudden changes affecting search stability.

[0052] In practical applications, the feature importance filtering module continuously monitors the performance of each feature. For example, if the click-through rate and conversion rate of the feature "green building certification" continue to rise, and its importance entropy value decreases accordingly, the system will automatically include it in the core feature sequence. Conversely, if the importance of a traditional feature gradually decreases, the system will remove it from the core sequence, but it will still remain in the complete feature set for subsequent matching. This dynamic adjustment mechanism allows the system to adapt to changes in industry trends and user preferences, always maintaining optimal search results.

[0053] The entire feature selection process is fully automated, requiring no manual intervention. The system automatically discovers patterns between features and retrieval success rates using machine learning algorithms, continuously optimizing the feature importance assessment model. The maintenance of core feature sequences utilizes a distributed storage architecture, supporting fast access and high-concurrency processing. As the data volume in the opportunity feature repository grows, the system automatically expands its processing capacity to ensure that feature selection efficiency remains unaffected. This design enables the system to handle large-scale opportunity data, providing users with fast and accurate retrieval services.

[0054] Example 5: The feedback learning module collects user interactions with search results, which constitute a crucial data source for system optimization. Interactions include both explicit and implicit feedback. Explicit feedback directly expresses user satisfaction, such as rating, liking, or disliking search results, and actions explicitly marked as "useful" or "irrelevant." Implicit feedback indirectly reflects preferences through user behavior patterns, including click order in the results list, dwell time, scope of detailed content viewed, download or favorite actions, and the behavior of completely ignoring certain results. The system records these interaction details using event tracking technology, including timestamps, action types, and metadata such as identifiers of specific business opportunities involved. All feedback data undergoes anonymization, removing personally identifiable information before being stored in a feedback log, forming a continuously growing feedback dataset.

[0055] The process of generating feedback signals involves parsing and quantifying the raw interaction data. The system converts different types of interaction behaviors into uniform feedback intensity values, which reflect the user's level of acceptance of a specific search result. For example, directly giving a high rating or adding a item to favorites generates a strong positive feedback signal, while quickly skipping it after a brief browsing generates a weaker negative feedback signal. Feedback signals also consider contextual factors, such as comparisons between different results within the same search session and the business opportunity information ultimately selected by the user. These signals are encoded as structured data, containing fields such as search session identifier, business opportunity information identifier, feedback type, feedback intensity, and time information.

[0056] The intent recognition strategy in the intent mapping module is adjusted based on feedback signals. The system analyzes intent recognition biases reflected in the feedback data, such as finding that certain specific expressions in search requests frequently lead to irrelevant results. Based on these findings, the intent recognition logic of the large language model is fine-tuned. The adjustment process employs incremental learning, using positive and negative samples extracted from the feedback data to optimize model parameters. For example, when multiple feedback responses indicate that a user query for "prefabricated building project" actually requires information on the entire design and construction process, not just the construction phase, the system adjusts the intent parsing weight for such requests, strengthening the recognition importance of the semantic element "entire process." Model updates are performed in a rolling manner, periodically integrating new feedback data to avoid the instability caused by frequent updates.

[0057] The matching weights in the opportunity matching module are adjusted based on user preference patterns revealed by feedback signals. The system analyzes the impact of different feature dimensions in the feedback data on user decisions and recalibrates the weighting ratios of semantic relevance and time decay coefficients. For example, if feedback shows that users prefer recently published information, the weight of the time decay coefficient will be appropriately increased even if the semantic relevance is slightly lower. The adjustment of matching weights also considers differences in opportunity types, using different weight configurations for different categories of opportunities such as bidding information, supplier information, and project information. The adjustment process is implemented through algorithm optimization, with the goal of making the ranking of matching results more consistent with users' actual selection patterns.

[0058] The feedback learning module establishes a long-term learning mechanism to continuously track and optimize the effects of adjustments. The system monitors the search performance after adjustments and evaluates the optimization effect by comparing user satisfaction indicators before and after the adjustments. Adjustments with significant effects are retained and strengthened, while adjustments with insignificant effects are rolled back or corrected. This continuous optimization enables the system to adapt to changes in user needs and industry trends, maintaining continuous improvement in search performance. The learning mechanism also includes handling of abnormal feedback, which can identify and filter accidental erroneous feedback or malicious operations, ensuring the reliability of the learning process.

[0059] The application of feedback data employs a differentiated strategy, treating new and experienced users differently. Initial feedback from new users is given higher weight to quickly establish a personalized preference model; feedback from experienced users is used for fine-tuning and optimization, gradually refining the user profile. The system also identifies group feedback patterns; when multiple users exhibit similar responses to the same search request, a global optimization process is initiated, adjusting system-level parameter settings. This learning strategy, combining personalization and generalization, allows the system to both accommodate individual differences and maintain overall performance improvement.

[0060] In practical applications, the feedback learning module continuously accumulates user behavior data, forming rich training samples. These samples are used to continuously improve the accuracy of intent recognition and opportunity matching. For example, when it is found that multiple users mark the qualification information of a specific supplier as useful, the system will increase the weight of such information in the matching process; when users frequently ignore certain types of project announcements, the recommendation priority of such announcements will be reduced. Through this continuous learning based on actual usage feedback, the system gradually adapts to users' real needs and preferences, providing more accurate opportunity retrieval services. The entire feedback learning process forms a closed-loop optimization system, enabling the opportunity retrieval agent to continuously evolve and improve, better serving the needs of opportunity discovery and decision support in the construction engineering field.

[0061] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0062] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A large-scale model-based intelligent system for retrieving business opportunities in construction projects, characterized in that, include: The information acquisition module is used to acquire multimodal business opportunity information in the field of construction engineering from multiple heterogeneous data sources, integrate the multimodal business opportunity information and extract its features to generate a business opportunity feature warehouse; The intent mapping module is used to receive the user's input search request, use a large language model to parse the semantic intent of the search request, generate an intent feature vector, and perform preliminary matching in the business opportunity feature warehouse based on the intent feature vector; The mapping relationship construction module is used to construct a dynamic mapping relationship between the intent space and the opportunity feature space based on the correlation strength between the intent feature vector and the opportunity features in the opportunity feature warehouse. The implementation of the mapping relationship construction module includes: calculating the cosine similarity between the intent feature vector and each opportunity feature vector in the opportunity feature warehouse; filtering candidate opportunity feature subsets based on the cosine similarity; constructing a dynamic mapping relationship between the intent space and the opportunity feature space based on the distribution relationship between the candidate opportunity feature subsets and the intent feature vectors; specifically, by analyzing the clustering features and distribution patterns of the opportunity feature vectors in the candidate subsets, establishing a mapping model from intent vectors to opportunity vectors, and using an adjustable weight matrix from intent vectors to opportunity vectors to establish a representation of the response intensity of each feature dimension to different intents, thereby constructing a mapping relationship between the intent space and the opportunity feature space. The business opportunity matching module is used to calculate the semantic relevance between each business opportunity information in the business opportunity feature warehouse and the intent feature vector according to the dynamic mapping relationship, and generate a business opportunity matching sequence by combining the time weight of the business opportunity information; The dynamic update module is used to monitor data changes in business opportunity information, update the business opportunity feature warehouse according to the newly acquired business opportunity information, and adaptively adjust the dynamic mapping relationship between the intent space and the business opportunity feature space. The retrieval management module is used to generate hierarchical retrieval strategies and build a business opportunity retrieval knowledge base based on the updated business opportunity feature repository.

2. The intelligent agent system for retrieving business opportunities in construction projects based on a large model as described in claim 1, characterized in that, The information collection module is implemented in the following ways: The multimodal business opportunity information is preprocessed to extract text features, image features, and numerical features from the multimodal business opportunity information; The text features, image features, and numerical features are fused in a multimodal manner to generate a unified representation of the business opportunity feature vector; Based on the source and type of the business opportunity feature vectors, the business opportunity feature vectors are classified and stored to form the business opportunity feature warehouse.

3. The intelligent agent system for retrieving business opportunities in construction projects based on a large model as described in claim 1, characterized in that, The implementation of the intent mapping module includes: The intent of the retrieval request is identified using a large language model, and the intent category and confidence level of the retrieval request are output. Based on the intent category and confidence level, select the corresponding intent descriptor from the predefined intent dictionary; The intent descriptor is encoded into a high-dimensional vector to form the intent feature vector.

4. The intelligent agent system for retrieving business opportunities in construction projects based on a large model as described in claim 1, characterized in that, The business opportunity matching module is implemented in the following ways: Based on the dynamic mapping relationship, obtain business opportunity information related to the intent feature vector from the business opportunity feature warehouse; Extract the time attribute of each business opportunity and calculate its time decay coefficient relative to the current time; The business opportunity matching sequence is generated by combining the semantic relevance and the time decay coefficient.

5. The intelligent agent system for retrieving business opportunities in construction projects based on a large model according to claim 1, characterized in that, The implementation methods of the dynamic update module include: New business opportunity information is collected in real time, its business opportunity feature vector is extracted and added to the business opportunity feature warehouse; Detect changes in the status of business opportunity information in the business opportunity feature warehouse, and mark and remove expired business opportunity information; The dynamic mapping relationship between the intent space and the opportunity feature space is recalculated based on the updated opportunity feature repository.

6. The intelligent agent system for retrieving business opportunities in construction projects based on a large model according to claim 1, characterized in that, Also includes: The feature importance filtering module is used to analyze the correlation strength between each business opportunity feature in the business opportunity feature warehouse and the historical retrieval success rate, and calculate the feature importance entropy; based on the feature importance entropy, key business opportunity features are filtered to form a core feature sequence; After generating the intent feature vector, the intent mapping module prioritizes the initial matching with the core feature sequence.

7. The intelligent agent system for retrieving business opportunities in construction projects based on a large model according to claim 1, characterized in that, Also includes: The feedback learning module is used to collect user interaction behavior with search results and generate feedback signals; The matching weights between the intent recognition strategy in the intent mapping module and the business opportunity matching module are adjusted based on the feedback signal.

8. The intelligent agent system for retrieving business opportunities in construction projects based on a large model according to claim 1, characterized in that, The multimodal business opportunity information in the information collection module includes, but is not limited to, tender announcements, winning bid announcements, equipment and service procurement lists, engineering information, project planning documents, engineering drawings, supplier qualification documents, and market analysis reports.

9. A method for constructing an intelligent agent for retrieving business opportunities in construction projects based on a large model, applied to the intelligent agent system for retrieving business opportunities in construction projects based on a large model as described in any one of claims 1-8, characterized in that, Includes the following steps: Step 1: Obtain multimodal business opportunity information in the construction engineering field from multiple heterogeneous data sources, integrate the multimodal business opportunity information and extract its features to generate a business opportunity feature warehouse; Step 2: Receive the user's search request, use a large language model to parse the semantic intent of the search request, generate an intent feature vector, and perform preliminary matching in the business opportunity feature warehouse based on the intent feature vector; Step 3: Based on the correlation strength between the intent feature vector and the opportunity features in the opportunity feature warehouse, construct a dynamic mapping relationship between the intent space and the opportunity feature space; Step 4: Based on the dynamic mapping relationship, calculate the semantic relevance between each business opportunity information in the business opportunity feature warehouse and the intent feature vector, and generate a business opportunity matching sequence by combining the time weight of the business opportunity information; Step 5: Monitor data changes in business opportunity information, update the business opportunity feature warehouse based on newly acquired business opportunity information, and adaptively adjust the dynamic mapping relationship between the intent space and the business opportunity feature space; Step 6: Based on the updated business opportunity feature repository, generate a hierarchical retrieval strategy and build a business opportunity retrieval knowledge base.

Images