Bid evaluation method and device, electronic equipment, storage medium and program product

Abstract

The embodiment of the invention relates to the technical field of automatic bid evaluation, and discloses a bid evaluation method and device, electronic equipment, a storage medium and a program product, and the method comprises the steps: generating a task execution graph corresponding to a bid invitation file through a main agent, and enabling the structure of the task execution graph to be a directed acyclic graph composed of a plurality of nodes and edges, each node represents a sub-review task; determining a target sub-agent matched with each sub-review task from a plurality of sub-agents, and generating a scheduling plan in combination with the task execution graph; executing the scheduling plan to enable the plurality of target sub-agents to sequentially process the sub-review tasks in the directed acyclic graph based on the bidding file to obtain a plurality of sub-review results; and generating a total review result corresponding to the bidding file based on the plurality of sub-review results. Through the automatic bid evaluation mode which is driven by the main agent and executed by the sub-agents and is based on the directed acyclic graph, the evaluation efficiency and consistency are improved, and the subjectivity of bid evaluation is reduced.

Description

Technical Field

[0001] This application relates to the field of automated bid evaluation technology, and in particular to a bid evaluation method, apparatus, electronic device, storage medium, and program product. Background Technology

[0002] Bidding and tendering review is a crucial step in ensuring the fair and efficient selection of partners for projects. Traditional bidding and tendering methods rely heavily on human experts, which has inherent drawbacks such as low efficiency, strong subjectivity, and poor consistency.

[0003] Improving the efficiency and consistency of bid evaluation and reducing subjectivity in the bid evaluation process are urgent technical problems that need to be solved. Summary of the Invention

[0004] The purpose of this application is to provide at least one evaluation method, device, electronic device, storage medium and program product, which can at least solve the problems of low evaluation efficiency, strong subjectivity and poor consistency, and at least achieve the effect of improving evaluation efficiency and consistency and reducing the subjectivity of evaluation.

[0005] To address the aforementioned technical problems, at least one embodiment of this application provides a bid evaluation method applied to a bid evaluation system. The bid evaluation system includes a main intelligent agent and multiple sub-intelligent agents, comprising: generating a task execution graph corresponding to a tender document using the main intelligent agent; the task execution graph being a directed acyclic graph (DAG) consisting of multiple nodes and edges, where each node represents a sub-review task, each edge represents a logical dependency between sub-review tasks, and each sub-review task corresponds to a review element in the tender document; determining a target sub-intelligent agent matching each sub-review task from the multiple sub-intelligent agents, and generating a scheduling plan based on the task execution graph; executing the scheduling plan so that multiple target sub-intelligent agents sequentially process the sub-review tasks in the DAG based on the tender document, obtaining multiple sub-review results; and generating a total review result corresponding to the tender document based on the multiple sub-review results.

[0006] The traditional, unstructured review process that relies on human experience is transformed into an automated process based on a directed acyclic graph, driven by a main intelligent agent and executed by sub-intelligent agents, which improves review efficiency and consistency and reduces the subjectivity of the evaluation.

[0007] Furthermore, the task execution diagram clearly defines all sub-review tasks and their logical order, making the entire review process no longer a black box, enhancing the objectivity and auditability of the review, and ensuring that any conclusion can be traced back to specific task nodes and execution paths, thus guaranteeing the systematicness and traceability of the review process.

[0008] Furthermore, the "master-sub agent" architecture design enables the system to adapt to new review requirements by adding or modifying "sub agents" without refactoring the entire system, significantly improving the system's scalability and adaptability.

[0009] Furthermore, the structure of directed acyclic graphs naturally supports parallel processing of subtasks that have no dependencies, thereby significantly shortening the review cycle (e.g., reviewing the same qualification of multiple bidders at the same time).

[0010] In some optional embodiments, generating the task execution graph corresponding to the tender document using the main intelligent agent includes: extracting review elements from the tender document using the main intelligent agent to obtain multiple review elements; determining the review stage to which each review element belongs; and generating the task execution graph based on the preset logical dependencies corresponding to the review stages and the multiple review elements.

[0011] By categorizing review elements into "review stages" (such as preliminary review, technical review, business review, and comprehensive review) and building dependencies based on the inherent and recognized business logic between these review stages (such as "if qualifications are not met, it is a veto and no technical review is required"), the generated review process is ensured to comply with industry standards and regulations, avoiding logical confusion and improving the accuracy and compliance of task decomposition.

[0012] In some optional embodiments, determining the target sub-agent matching each sub-review task from the plurality of sub-agents includes: obtaining a capability description vector for each sub-agent to obtain multiple capability description vectors; obtaining a task description vector for each sub-review task to obtain multiple task description vectors; calculating the similarity between the task description vector and each capability description vector for each task description vector to obtain multiple similarities corresponding to each task description vector; and determining the target sub-agent matching each sub-review task based on the multiple similarities corresponding to each task description vector.

[0013] By simultaneously mapping task requirements (review elements) and executor capabilities (sub-agents) to a high-dimensional semantic space for matching, this method surpasses simple keyword-based matching. It more accurately identifies the sub-agents best suited for specific review tasks, thereby improving the quality and accuracy of individual reviews and achieving precise task-sub-agent matching. Furthermore, even if the description of review elements changes (e.g., "qualification review" becomes "certification examination"), as long as the semantics remain similar, the correct sub-agent can still be matched. When introducing sub-agents with new capabilities, only their capability description vectors need to be updated to integrate them into the scheduling system. This gives the system strong adaptive and evolutionary capabilities, enhancing its flexibility and robustness.

[0014] In some optional embodiments, the method further includes: obtaining tender documents and bid documents; decomposing the tender documents and bid documents according to semantic units to obtain multiple text blocks; storing the multiple text blocks in a pre-built knowledge base; wherein, the process of the target sub-agent processing the sub-review task includes: retrieving text blocks corresponding to the review elements corresponding to the sub-review task from the knowledge base to obtain multiple target text blocks, wherein the multiple target text blocks belong to the multiple text blocks; processing the sub-review task based on the multiple target text blocks to obtain the sub-review result.

[0015] By pre-slicing and storing unstructured documents, the sub-agent does not need to "read through" the entire text during execution. It can directly retrieve the most relevant clauses or content fragments of the bidding document through semantic retrieval, solving the attention distraction and illusion problems that large models encounter when processing long texts. This makes the review basis more accurate and the reasoning more reliable, greatly improving the accuracy and efficiency of information processing. It avoids repeatedly inputting the entire massive bidding document into the large model, significantly reducing token consumption, lowering the model's computational burden and cost, and increasing the review processing speed. Furthermore, the retrieved "text blocks" can serve as citation evidence for the sub-review results, clearly demonstrating the specific original text on which the sub-agent made its judgments, enhancing the interpretability and credibility of the review results, and ensuring the transparency of the review basis.

[0016] In some optional embodiments, generating the overall review result corresponding to the tender document based on the multiple sub-review results includes: performing deviation detection on each sub-review result to obtain a detection result; when the detection result indicates that at least one sub-review result has a deviation, determining the sub-review tasks affected by the at least one sub-review result to obtain at least one target sub-review task; using the target sub-agent corresponding to each target sub-review task, reprocessing the target sub-review tasks according to the order in the directed acyclic graph to obtain at least one new sub-review result, completing one round of deviation detection and correction; re-performing deviation detection and correction on the new at least one sub-review result until the detection result obtained within a preset number of rounds indicates that the current sub-review results are all without deviation, generating the overall review result based on the current sub-review results; after the deviation detection and correction of the preset number of rounds, when the detection result obtained by re-performing deviation detection indicates that at least one current sub-review result has a deviation, correcting the current at least one sub-review result using expert correction, and generating the overall review result based on the current multiple sub-review results.

[0017] Through multiple rounds of testing and correction, the impact of errors in a single stage on the final result can be significantly reduced, greatly improving the reliability and robustness of the review results. This enables self-verification and error correction of sub-review results, and the system no longer outputs results only once, but possesses the ability to "self-question" and "iteratively optimize." Furthermore, this embodiment constructs a complete error correction process from "machine-to-machine" review to "human-to-machine" collaboration. First, it attempts to correct deviations through the system itself (recalculation). When the system cannot resolve the issue on its own, human experts are smoothly introduced. This design ensures automation while providing a final safeguard for key decisions, achieving a balance in human-machine collaboration. By locating the "affected sub-review tasks," precise local rollback and recalculation are achieved, rather than starting from scratch, ensuring the efficiency of the correction process.

[0018] In some optional embodiments, the method further includes: constructing sample data based on at least one expert correction result obtained after expert correction, obtaining at least one sample data, wherein each expert correction result is a sub-review result; and optimizing the corresponding target sub-agent using each sample data.

[0019] The corrective actions of experts are transformed into optimization samples for the system. This allows the system to learn from errors, target its weaknesses for improvement, and reduce similar errors in future reviews, thus achieving continuous learning and performance improvement for the sub-agents. As the sub-agents continue to learn and optimize, their outputs will match the experts' expectations more and more, gradually reducing the need for expert intervention and lowering long-term maintenance costs and reliance on experts.

[0020] At least one embodiment of this application also provides a bid evaluation device applied to a bid evaluation system. The bid evaluation system includes a main intelligent agent and multiple different sub-intelligent agents, comprising: a first generation module, used to generate a task execution graph corresponding to a tender document using the main intelligent agent. The task execution graph is a directed acyclic graph composed of multiple nodes and edges. Each node represents a sub-review task, and each edge represents a logical dependency between sub-review tasks. Each sub-review task corresponds to a review element in the tender document; a matching module, used to determine the target sub-intelligent agent matching each sub-review task from the multiple sub-intelligent agents and generate a scheduling plan in conjunction with the task execution graph; an execution module, used to execute the scheduling plan so that multiple target sub-intelligent agents sequentially process the sub-review tasks in the directed acyclic graph based on the tender document to obtain multiple sub-review results; and a second generation module, used to generate a total review result corresponding to the tender document based on the multiple sub-review results.

[0021] In some optional embodiments, the first generation module is used to extract review elements from the tender document using the main intelligent agent to obtain multiple review elements; determine the review stage to which each review element belongs; and generate the task execution graph based on the preset logical dependency relationship corresponding to the review stage and the multiple review elements.

[0022] In some optional embodiments, the matching module is used to obtain the capability description vector of each sub-agent, resulting in multiple capability description vectors; obtain the task description vector of each sub-review task, resulting in multiple task description vectors; calculate the similarity between the task description vector and each capability description vector for each task description vector, resulting in multiple similarities corresponding to each task description vector; and determine the target sub-agent matched for each sub-review task based on the multiple similarities corresponding to each task description vector.

[0023] In some optional embodiments, the system further includes: a text processing module for acquiring tender documents and bid documents; decomposing the tender documents and bid documents according to semantic units to obtain multiple text blocks; storing the multiple text blocks in a pre-built knowledge base; the process of the target sub-agent processing the sub-review task includes: retrieving text blocks corresponding to the review elements corresponding to the sub-review task from the knowledge base to obtain multiple target text blocks, wherein the multiple target text blocks belong to the multiple blocks; processing the sub-review task based on the multiple target text blocks to obtain the sub-review result.

[0024] In some optional embodiments, the second generation module is used to perform deviation detection on each sub-review result to obtain a detection result; when the detection result indicates that at least one sub-review result has a deviation, determine the sub-review tasks affected by the at least one sub-review result to obtain at least one target sub-review task; use the target sub-agent corresponding to each target sub-review task to reprocess the target sub-review tasks according to the order in the directed acyclic graph to obtain at least one new sub-review result, completing one round of deviation detection and correction; re-perform deviation detection and correction on the at least one new sub-review result until the detection result obtained within a preset number of rounds indicates that the current sub-review results are all without deviation, and generate the total review result based on the current sub-review results; after the deviation detection and correction of the preset number of rounds, when the detection result obtained by re-perform deviation detection indicates that at least one current sub-review result has a deviation, correct the at least one current sub-review result using expert correction, and generate the total review result based on the current multiple sub-review results.

[0025] In some optional embodiments, it further includes: an optimization module, configured to construct sample data based on at least one expert correction result obtained after expert correction, to obtain at least one sample data, wherein each expert correction result is a sub-review result; and to optimize the corresponding target sub-agent using each sample data.

[0026] At least one embodiment of this application also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the above-described evaluation method.

[0027] At least one embodiment of this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described evaluation method.

[0028] At least one embodiment of this application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described evaluation method. Attached Figure Description

[0029] One or more embodiments are illustrated by way of example with reference to the accompanying drawings, and these illustrative descriptions do not constitute a limitation on the embodiments.

[0030] Figure 1 This is a flowchart of a bid evaluation method provided in one embodiment of this application; Figure 2 This is a schematic diagram of a task execution graph provided in one embodiment of this application; Figure 3 This is a schematic diagram of a bid evaluation device provided in another embodiment of this application. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the various embodiments of this application will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been provided in the various embodiments of this application to help readers better understand this application. However, the technical solutions claimed in this application can be implemented even without these technical details and various changes and modifications based on the following embodiments. The division of the various embodiments below is for the convenience of description and should not constitute any limitation on the specific implementation of this application. The various embodiments can be combined with and referenced by each other without contradiction.

[0032] Example 1: The evaluation method of this embodiment can be applied to electronic devices with communication, computing, and data storage capabilities. This electronic device can be used in an evaluation system, which further includes a main intelligent agent and multiple different sub-intelligent agents. The specific process can be as follows: Figure 1 As shown, it includes S101-S104.

[0033] S101, use the main intelligent agent to generate the task execution diagram corresponding to the tender document.

[0034] The task execution graph is a directed acyclic graph consisting of multiple nodes and edges. Each node represents a sub-review task, each edge represents the logical dependency between sub-review tasks, and each sub-review task corresponds to a review element in the tender document.

[0035] Evaluation criteria are indicators set in the bidding documents based on the characteristics and requirements of the bidding project, used to measure the quality of bids. These indicators usually cover multiple dimensions such as technology, commerce, and price, and are the main basis for judging whether bids meet the bidding requirements and whether they are qualified to win the bid.

[0036] Logical dependency refers to the interrelationship and mutual constraint between review tasks. This relationship determines that during the bid evaluation process, the evaluation results of some review tasks may directly affect the evaluation of other review tasks, or the evaluation of some review tasks may need to be based on the evaluation results of other review tasks.

[0037] The embodiments of this application do not limit the specific types of intelligent agents, such as the main intelligent agent and the sub-intelligent agents. For example, the main intelligent agent can be any one of the GPT series; the sub-intelligent agent can also be any one of the GPT series, or DeepSeek, or Wenxin Yiyan, etc.

[0038] In some examples, electronic devices maintain an agent capability registry, which contains metadata for multiple sub-agents. This metadata includes the sub-agent's identifier, capability description vector, and so on. The capability description vector indicates the sub-agent's task processing capabilities. When a new sub-agent is added to the evaluation system, it can first be registered in the agent capability registry.

[0039] In some examples, generating a task execution graph corresponding to a tender document using a master intelligent agent may include: extracting review elements from the tender document using the master intelligent agent to obtain multiple review elements; determining the review stage to which each review element belongs; and generating a task execution graph based on the preset logical dependencies corresponding to the review stages and the multiple review elements.

[0040] The process of extracting review elements from the tender document using the main intelligent agent to obtain multiple review elements can be as follows: based on the preset first prompt template and the tender document, generate a review element extraction prompt; input the review element extraction prompt into the main intelligent agent, and use the main intelligent agent to extract the review elements from the tender document.

[0041] The embodiments of this application do not limit the specific text content of the preset first prompt template, and can be set based on experience.

[0042] In some examples, when using the main AI agent to extract review elements from the tender document, the corresponding descriptive content can also be extracted simultaneously. For instance, based on a preset second prompt template and the tender document, a prompt for extracting review elements and descriptions can be generated; this prompt can then be input into the main AI agent, which can then extract the descriptive content corresponding to the review elements in the tender document.

[0043] The embodiments of this application do not limit the specific text content of the preset second prompt template, and can be set based on experience.

[0044] In some examples, before the main agent extracts review elements using the tender documents, the process includes: obtaining the tender documents and bid documents; decomposing the tender documents and bid documents into semantic units to obtain multiple text blocks; and storing these text blocks in a pre-built knowledge base. After storing the text blocks in the knowledge base, an index is assigned to each text block. The index of the text block in the knowledge base includes the document's identifier (e.g., tender document identifier (review project identifier), bid document identifier). It should be noted that the tender document identifier and each bid document identifier within the same review project are interconnected; that is, the text block corresponding to the bid document can be retrieved through the tender document identifier.

[0045] When the main agent extracts review elements using the tender document, the constructed prompts (review element extraction prompts, review element and description extraction prompts) can use the identifier of the tender document. The main agent can retrieve the corresponding text block of the tender document from the knowledge base based on the identifier, thereby facilitating the extraction of review elements from the retrieved text block, or the extraction of review elements and corresponding description content.

[0046] In some examples, determining the review stage to which each review element belongs may include: generating a review element classification prompt based on a preset third prompt template; inputting the review element classification prompt into the main intelligence agent, which then classifies the review element into the review stage. The embodiments of this application do not limit the types of review stages included. For example, review stages may include qualification review (preliminary review), skills review, business review, and comprehensive review.

[0047] Regarding the specific text content of the preset third prompt template, the embodiments of this application do not impose limitations and can be set based on experience. For example, the third prompt template is: categorize the review factor "[Factor Name]" into one of the following review stages: [Qualification Review, Technical Review, Business Review, Comprehensive Review], and only output the review stage name.

[0048] In some examples, determining the review stage to which each review element belongs may include: first classifying each review element based on predefined rules; and then using an agent to classify review elements into review stages for those that cannot be classified based on the predefined rules.

[0049] By first using pre-defined rules and then using intelligent agents to categorize and review elements, different evaluation items can be automatically adapted without relying on fixed categorization templates.

[0050] In some examples, each review element is categorized based on predefined rules, which may include categorizing review elements according to predefined keywords (such as "qualification" -> "preliminary review", "technology" -> "technical review").

[0051] In some examples, determining the review stage to which each review element belongs may include: obtaining the description content corresponding to the review element; and determining the review stage to which each review element belongs based on the review element and the description content.

[0052] The specific implementation method for determining the review stage to which each review element belongs based on the review elements and their descriptions can be found in the above-mentioned implementation method for determining the review stage to which a review element belongs based directly on the review elements themselves, which will not be repeated here.

[0053] Regarding the specific implementation of obtaining the description content corresponding to the review elements, it can be obtained together when obtaining the review elements, or it can be obtained separately using the main intelligent agent. The embodiments of this application do not limit this.

[0054] The embodiments of this application do not limit the specific logical dependencies that correspond to the review stages. For example, the general logical dependencies corresponding to the review stages can be applied here. For example, when the review stages include qualification review, technical review, commercial review, and comprehensive review, the default logical dependencies corresponding to the review stages can be qualification review -> technical review, commercial review -> comprehensive review, that is, qualification review is performed first, followed by technical and commercial reviews, and finally comprehensive review.

[0055] Based on the pre-defined logical dependencies and multiple review elements corresponding to the review phase, a task execution graph is generated. The task execution graph is a directed acyclic graph (DAG). For example, a task execution graph can be like this: Figure 2 As shown in the diagram, this task execution diagram includes Qualification Review Task 1 and Qualification Review Task 2, which belong to the qualification review stage; Technical Review Task 1 and Technical Review Task 2, which belong to the technical review stage; Business Review Task 1 and Business Review Task 2, which belong to the business review stage; and Comprehensive Review Task 1, which belongs to the comprehensive review stage. The execution of Technical Review Task 1, Technical Review Task 2, Business Review Task 1, and Business Review Task 2 all depend on the completion and passing of Qualification Review Task 1 and Qualification Review Task 2. The execution of Comprehensive Review Task 1 depends on the completion of Technical Review Task 1, Technical Review Task 2, Business Review Task 1, and Business Review Task 2.

[0056] By categorizing review elements into "review stages" (such as preliminary review, technical review, business review, and comprehensive review) and building dependencies based on the inherent and recognized business logic between these review stages (such as "if qualifications are not met, it is a veto and no technical review is required"), the generated review process is ensured to comply with industry standards and regulations, avoiding logical confusion and improving the accuracy and compliance of task decomposition.

[0057] In some examples, the generated task planning diagram includes information for each sub-review task, including the tender document identifier corresponding to the review element, in addition to the review element itself. In this task planning diagram, the tender document identifiers included in each sub-review task are the same.

[0058] S102, determine the target sub-agent matching each sub-review task from multiple sub-agents, and generate a scheduling plan by combining the task execution graph.

[0059] In some examples, determining the target sub-agent matching each sub-review task from multiple sub-agents may include: obtaining the capability description vector of each sub-agent, resulting in multiple capability description vectors; obtaining the task description vector of each sub-review task, resulting in multiple task description vectors; calculating the similarity between each task description vector and each capability description vector, resulting in multiple similarities corresponding to each task description vector; and determining the target sub-agent matching each sub-review task based on the multiple similarities corresponding to each task description vector.

[0060] By simultaneously mapping task requirements (review elements) and executor capabilities (sub-agents) to a high-dimensional semantic space for matching, this method surpasses simple keyword-based matching. It more accurately identifies the sub-agents best suited for specific review tasks, thereby improving the quality and accuracy of individual reviews and achieving precise task-sub-agent matching. Furthermore, even if the description of review elements changes (e.g., "qualification review" becomes "certification examination"), as long as the semantics remain similar, the correct sub-agent can still be matched. When introducing sub-agents with new capabilities, only their capability description vectors need to be updated to integrate them into the scheduling system. This gives the system strong adaptive and evolutionary capabilities, enhancing its flexibility and robustness.

[0061] Electronic devices maintain an agent capability registry, which contains capability description vectors for each agent. In some examples, obtaining the capability description vector of each sub-agent may include retrieving the capability description vector of each sub-agent from the agent capability registry.

[0062] In some examples, obtaining the task description vector for each sub-review task may include encoding the review elements corresponding to the sub-review task to obtain the task description vector for the sub-review task.

[0063] In other examples, obtaining the task description vector for each sub-review task may include: obtaining the review elements and description content corresponding to the sub-review task; encoding the review elements and description content to obtain the task description vector for the sub-review task.

[0064] The embodiments of this application do not limit how the similarity between the task description vector and the capability description vector is calculated. For example, cosine similarity, L1 similarity, Euclidean similarity, etc., can be used.

[0065] The embodiments of this application do not limit how to determine the target sub-agent for each sub-review task based on multiple similarities corresponding to each task description vector. In some examples, the sub-agent with the highest similarity among the multiple similarities corresponding to the task description vector can be used as the target sub-agent for matching the sub-review task. In other examples, similarities exceeding a similarity threshold can be selected from the multiple similarities corresponding to the task description vector. If multiple similarities exceed the similarity threshold, the sub-agent corresponding to any one of these similarities (similarity exceeding the similarity threshold) can be used as the target sub-agent for matching the sub-review task; or, the task processing queue of the multiple similarities exceeding the similarity threshold can be obtained, and the sub-agent with the fewest tasks in the task processing queue can be selected as the target sub-agent for matching the sub-review task. In other embodiments, the sub-agent corresponding to any one of the similarities with the largest preset value among the multiple similarities corresponding to the task description vector can be used as the target sub-agent for matching the sub-review task.

[0066] After identifying the sub-agent corresponding to each sub-review task, a scheduling plan can be generated based on the sub-agent corresponding to each sub-review task and the task execution graph. The scheduling plan is used to represent the execution order of the sub-review tasks and the sub-agents that need to be scheduled to process each sub-review task.

[0067] S103, execute the scheduling plan so that multiple target sub-agents process the sub-review tasks in the directed acyclic graph in sequence based on the tender documents, and obtain multiple sub-review results.

[0068] In some cases, electronic devices execute scheduling plans, sending sub-review tasks whose current dependencies have been met and are pending to the task processing queue of the corresponding target sub-agent for processing.

[0069] In some examples, the process of the target sub-agent handling the sub-review task includes: retrieving the text blocks corresponding to the review elements in the knowledge base based on the review elements corresponding to the sub-review task, obtaining multiple target text blocks, and multiple target text blocks belonging to multiple text blocks; and processing the sub-review task based on the multiple target text blocks to obtain the sub-review result.

[0070] In some examples, each sub-review task in the execution schedule includes information including, in addition to the review elements, the tender document identifier corresponding to the review elements. In the knowledge base, the text block corresponding to the tender document is associated with the tender document identifier. The text block obtained by decomposing the tender document can be retrieved from the knowledge base through the tender document identifier. The tender document is associated with both the tender document identifier and the tender document identifier. That is, the text block corresponding to the tender document can be retrieved from the knowledge base through both the tender document identifier and the tender document identifier.

[0071] Based on the review elements corresponding to the sub-review tasks, the text blocks corresponding to the review elements are retrieved from the knowledge base. This may include: retrieving text blocks related to the review elements in the tender documents from the knowledge base based on the tender document identifier, and retrieving text blocks related to the review elements from the bid documents (if there are multiple bid documents, the text blocks of the multiple bid documents are partitioned based on the bid document identifier).

[0072] In some examples, sub-review tasks are processed based on multiple target text blocks to obtain sub-review results. This may include: constructing review prompts based on the retrieved multiple target text blocks, review elements, and a preset fourth prompt template; inputting the review prompts into the target sub-agent, and using the sub-agent to output the sub-review results. If there is only one bid document, the sub-review results include the review conclusions for that bid document based on the review element; if there are multiple bid documents, the sub-review results include separate review conclusions for each of the multiple bid documents based on the review element.

[0073] The embodiments of this application do not limit the specific text content of the fourth prompt template, and it can be set based on experience.

[0074] By pre-slicing and storing unstructured documents, the sub-agent does not need to "read through" the entire text during execution. It can directly retrieve the most relevant clauses or content fragments of the bidding document through semantic retrieval, solving the attention distraction and illusion problems that large models encounter when processing long texts. This makes the review basis more accurate and the reasoning more reliable, greatly improving the accuracy and efficiency of information processing. It avoids repeatedly inputting the entire massive bidding document into the large model, significantly reducing token consumption, lowering the model's computational burden and cost, and increasing the review processing speed. Furthermore, the retrieved "text blocks" can serve as citation evidence for the sub-review results, clearly demonstrating the specific original text on which the sub-agent made its judgments, enhancing the interpretability and credibility of the review results, and ensuring the transparency of the review basis.

[0075] S104 generates the overall review result corresponding to the tender document based on the results of multiple sub-reviews.

[0076] In some cases, generating the overall review result for the tender document based on multiple sub-review results may include: directly summarizing the results of multiple sub-reviews to obtain the overall review result for the tender document.

[0077] In some examples, the overall review result for the tender document is generated based on multiple sub-review results. This process includes: performing deviation detection on each sub-review result to obtain a detection result; when the detection result indicates that at least one sub-review result is biased, determining the sub-review tasks affected by at least one sub-review result to obtain at least one target sub-review task; using the target sub-agent corresponding to each target sub-review task, reprocessing the target sub-review tasks according to the order in the directed acyclic graph to obtain at least one new sub-review result, completing one round of deviation detection and correction; re-performing deviation detection and correction on the new at least one sub-review result until the detection results obtained within a preset number of rounds indicate that the current sub-review results are all unbiased, and generating the overall review result based on the current sub-review results; after the deviation detection and correction of the preset number of rounds, when the detection result obtained by re-performing deviation detection indicates that at least one current sub-review result is biased, correcting the current at least one sub-review result using expert correction, and generating the overall review result based on the current multiple sub-review results.

[0078] The embodiments of this application do not limit the specific number of rounds. For example, the preset number of rounds may be three or two.

[0079] In some examples, bias detection of sub-review results may include: when the sub-review results include scores, performing statistical anomaly detection on the scores in the sub-review results; when the sub-review results include text descriptions, performing human feedback detection on the scores in the sub-review results.

[0080] In some examples, statistical anomaly detection of the scores in the sub-review results may include: determining the probability of occurrence of each score based on the scores corresponding to each tender document in the sub-review results using the Z-score (standard score) algorithm (or using other algorithms to determine the abnormal distribution of data); determining the statistical detection result based on the probability of occurrence of each score; if there is a probability of occurrence below the probability threshold, then the sub-review result is determined to have a statistical anomaly; if there is no probability of occurrence below the probability threshold, then the sub-review result is determined not to have a statistical anomaly.

[0081] In some cases, when a sub-review result includes a text description, performing manual feedback checks on the scores in that sub-review result may include: displaying the text description included in the sub-review result and receiving feedback from experts on that text description. If the expert feedback indicates that the text description is without bias, the feedback result indicates that the manual feedback check has passed; if the expert feedback indicates that the text description is biased, the feedback result indicates that the manual feedback check has failed.

[0082] If the deviation detection result indicates that the sub-review result has statistical anomalies or the feedback result indicates that the manual feedback detection failed, then the deviation detection result indicates that the sub-review result has deviations. If the deviation detection result indicates that the sub-review result does not have statistical anomalies and the feedback result indicates that the manual feedback detection passed, then the deviation detection result indicates that the sub-review result has no deviations.

[0083] In some examples, identifying the sub-review task affected by at least one sub-review result includes: identifying at least one sub-review task corresponding to at least one sub-review result, and sub-review tasks depending on the at least one sub-review task, as the sub-review task affected by at least one sub-review result.

[0084] Through multiple rounds of testing and correction, the impact of errors in a single stage on the final result can be significantly reduced, greatly improving the reliability and robustness of the review results. This enables self-verification and error correction of sub-review results, and the system no longer outputs results only once, but possesses the ability to "self-question" and "iteratively optimize." Furthermore, this embodiment constructs a complete error correction process from "machine-to-machine" review to "human-to-machine" collaboration. First, it attempts to correct deviations through the system itself (recalculation). When the system cannot resolve the issue on its own, human experts are smoothly introduced. This design ensures automation while providing a final safeguard for key decisions, achieving a balance in human-machine collaboration. By locating the "affected sub-review tasks," precise local rollback and recalculation are achieved, rather than starting from scratch, ensuring the efficiency of the correction process.

[0085] In some examples, the process also includes: constructing sample data based on at least one expert correction result obtained after expert correction, obtaining at least one sample data, with each expert correction result being a sub-review result; and optimizing the corresponding target sub-agent using each sample data.

[0086] The corrective actions of experts are transformed into optimization samples for the system. This allows the system to learn from errors, target its weaknesses for improvement, and reduce similar errors in future reviews, thus achieving continuous learning and performance improvement for the sub-agents. As the sub-agents continue to learn and optimize, their outputs will match the experts' expectations more and more, gradually reducing the need for expert intervention and lowering long-term maintenance costs and reliance on experts.

[0087] Example 2: Another embodiment of this application relates to a bid evaluation device applied to a bid evaluation system. The bid evaluation system includes a main intelligent agent and multiple different sub-intelligent agents. The implementation details of the bid evaluation device in this embodiment are described below. The following content is only for ease of understanding and is not essential for implementing this solution. A schematic diagram of the bid evaluation device in this embodiment can be seen as follows: Figure 3As shown, it includes: a first generation module 31, used to generate a task execution graph corresponding to the tender document using the main intelligent agent. The task execution graph is a directed acyclic graph composed of multiple nodes and edges. Each node represents a sub-review task, and each edge represents the logical dependency relationship between sub-review tasks. Each sub-review task corresponds to a review element in the tender document; a matching module 32, used to determine the target sub-intelligent agent matching each sub-review task from multiple sub-intelligent agents and generate a scheduling plan in combination with the task execution graph; an execution module 33, used to execute the scheduling plan so that multiple target sub-intelligent agents process the sub-review tasks in the directed acyclic graph sequentially based on the tender document to obtain multiple sub-review results; and a second generation module 34, used to generate the overall review result corresponding to the tender document based on the multiple sub-review results.

[0088] In some optional embodiments, the first generation module 31 is used to extract review elements from the bidding documents using the main intelligent agent to obtain multiple review elements; determine the review stage to which each review element belongs; and generate a task execution graph based on the preset logical dependency relationship corresponding to the review stage and the multiple review elements.

[0089] In some optional embodiments, the matching module 32 is used to obtain the capability description vector of each sub-agent, resulting in multiple capability description vectors; obtain the task description vector of each sub-review task, resulting in multiple task description vectors; calculate the similarity between the task description vector and each capability description vector for each task description vector, resulting in multiple similarities corresponding to each task description vector; and determine the target sub-agent matched for each sub-review task based on the multiple similarities corresponding to each task description vector.

[0090] In some optional embodiments, the system further includes: a text processing module 35, used to obtain tender documents and bid documents; decompose the tender documents and bid documents according to semantic units to obtain multiple text blocks; store the multiple text blocks in a pre-built knowledge base; the process of the target sub-agent processing the sub-review task includes: based on the review elements corresponding to the sub-review task, retrieving the text blocks corresponding to the review elements from the knowledge base to obtain multiple target text blocks, wherein the multiple target text blocks belong to multiple blocks; processing the sub-review task based on the multiple target text blocks to obtain the sub-review result.

[0091] In some optional embodiments, the second generation module 34 is used to perform deviation detection on each sub-review result to obtain a detection result; when the detection result indicates that at least one sub-review result has a deviation, determine the sub-review task affected by at least one sub-review result to obtain at least one target sub-review task; use the target sub-agent corresponding to each target sub-review task to reprocess the target sub-review task in the order of the directed acyclic graph to obtain at least one new sub-review result, completing one round of deviation detection and correction; re-perform deviation detection and correction on the new at least one sub-review result until the detection result obtained within a preset number of rounds indicates that the current sub-review results are all without deviation, and generate a total review result based on the current sub-review results; after the deviation detection and correction of the preset number of rounds, when the detection result obtained by re-perform deviation detection indicates that at least one current sub-review result has a deviation, correct the current at least one sub-review result using expert correction, and generate a total review result based on the current multiple sub-review results.

[0092] In some optional embodiments, it further includes: an optimization module 36, used to construct sample data based on at least one expert correction result obtained after expert correction, to obtain at least one sample data, each expert correction result being a sub-review result; and to optimize the corresponding target sub-agent using each sample data.

[0093] It is worth mentioning that all modules involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. Furthermore, to highlight the innovative aspects of this application, this embodiment does not introduce units that are not closely related to solving the technical problems proposed in this application; however, this does not mean that other units are absent in this embodiment.

[0094] Example 3: Another embodiment of this application relates to an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the evaluation methods in the above embodiments.

[0095] The memory and processor are connected via a bus, which can include any number of interconnecting buses and bridges, connecting various circuits of one or more processors and memories. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and will not be described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over the wireless medium via an antenna, which further receives data and transmits it to the processor.

[0096] The processor manages the bus and general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory is used to store data used by the processor during operation.

[0097] Example 4: Another embodiment of this application relates to a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the method embodiments described above.

[0098] That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0099] Example 5: Another embodiment of this application relates to a computer program product, including a computer program. When executed by a processor, the computer program implements the method embodiments described above.

[0100] Those skilled in the art will understand that the above embodiments are specific embodiments for implementing this application, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of this application.

Claims

1. A bid evaluation method, characterized in that, This is applied to a bidding evaluation system, which includes a main intelligent agent and multiple different sub-intelligent agents, including: The main intelligent agent is used to generate a task execution graph corresponding to the tender document. The structure of the task execution graph is a directed acyclic graph consisting of multiple nodes and edges. Each node represents a sub-review task, each edge represents the logical dependency between sub-review tasks, and each sub-review task corresponds to a review element in the tender document. The target sub-agent matching each sub-review task is determined from the plurality of sub-agents, and a scheduling plan is generated in combination with the task execution graph; The scheduling plan is executed so that multiple target sub-agents sequentially process the sub-review tasks in the directed acyclic graph based on the tender documents, and obtain multiple sub-review results; Based on the results of the multiple sub-reviews, the overall review result corresponding to the tender document is generated.

2. The method according to claim 1, characterized in that, The step of generating the task execution graph corresponding to the tender document using the main intelligent agent includes: The main intelligent agent is used to extract the evaluation elements from the bidding documents to obtain multiple evaluation elements; Determine the review stage to which each review element belongs; Based on the preset logical dependencies corresponding to the review phase and the multiple review elements, the task execution graph is generated.

3. The method according to claim 1, characterized in that, The step of determining the target sub-agent matching each sub-review task from the plurality of sub-agents includes: Obtain the capability description vector for each sub-agent to obtain multiple capability description vectors; Obtain the task description vector for each sub-review task to get multiple task description vectors; For each task description vector, the similarity between the task description vector and each capability description vector is calculated to obtain multiple similarities corresponding to each task description vector; Based on multiple similarities corresponding to each task description vector, the target sub-agent matched for each sub-review task is determined.

4. The method according to claim 1, characterized in that, Also includes: Obtain the tender documents and bid documents; The tender documents and bid documents are decomposed into multiple text blocks according to semantic units; Store the multiple text blocks into a pre-built knowledge base; The process by which the target sub-agent handles the sub-review task includes: Based on the review elements corresponding to the sub-review tasks, text blocks corresponding to the review elements are retrieved from the knowledge base to obtain multiple target text blocks, and the multiple target text blocks belong to the multiple text blocks; The sub-review task is processed based on the multiple target text blocks to obtain the sub-review results.

5. The method according to claim 1, characterized in that, The process of generating the overall review result corresponding to the tender document based on the results of the multiple sub-reviews includes: Deviation detection is performed on each sub-review result to obtain the detection result; When the detection result indicates that at least one sub-review result is biased, the sub-review task affected by the at least one sub-review result is determined, and at least one target sub-review task is obtained. Using the target sub-agent corresponding to each target sub-review task, the target sub-review tasks are reprocessed according to the order in the directed acyclic graph to obtain at least one new sub-review result, thus completing one round of deviation detection and correction. The deviation detection and correction are performed again on the at least one new sub-review result until the detection results obtained within the preset number of rounds indicate that the current sub-review results are all without deviation. Then, the total review result is generated based on the current sub-review results. After a preset number of deviation detections and corrections, if the detection result obtained from a re-detection indicates that at least one sub-review result has a deviation, the at least one sub-review result is corrected using expert correction, and the overall review result is generated based on the current multiple sub-review results.

6. The method according to claim 5, characterized in that, Also includes: Based on at least one expert correction result obtained after expert correction, sample data is constructed to obtain at least one sample data, and each expert correction result is a sub-review result. The corresponding target sub-agent is optimized using each sample data.

7. A bid evaluation device, characterized in that, This is applied to a bidding evaluation system, which includes a main intelligent agent and multiple different sub-intelligent agents, including: The first generation module is used to generate a task execution graph corresponding to the tender document using the main intelligent agent. The task execution graph is a directed acyclic graph consisting of multiple nodes and edges. Each node represents a sub-review task, each edge represents the logical dependency between sub-review tasks, and each sub-review task corresponds to a review element in the tender document. The matching module is used to determine the target sub-agent that matches each sub-review task from the plurality of sub-agents, and generate a scheduling plan in combination with the task execution graph; The execution module is used to execute the scheduling plan so that multiple target sub-agents sequentially process the sub-review tasks in the directed acyclic graph based on the tender documents, and obtain multiple sub-review results; The second generation module is used to generate the overall review result corresponding to the tender document based on the results of the multiple sub-reviews.

8. An electronic device, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the evaluation method as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the evaluation method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the evaluation method according to any one of claims 1 to 6.

Images