Intelligent bid evaluation auxiliary method and system based on large language model

Abstract

This invention provides a smart evaluation assistance method and system based on a large language model. The method initializes the received evaluation task and generates a corresponding evaluation strategy through rule mapping. Then, based on this evaluation strategy and the evaluation documents in the initialized evaluation task, it uses a pre-built large oracle model for semantic localization to extract full information. A two-layer evaluation information database is constructed based on this full information. Based on this database, a multi-role intelligent agent collaborative mechanism is used for division of labor, collaboration, and information exchange to generate scoring results and supporting explanations. Finally, a review report containing traceability evidence is generated based on the scoring results and supporting explanations and fed back to the business system. The intelligent agent experience database corresponding to the multi-role intelligent agent collaborative mechanism is updated according to the feedback results. This solves the problems of incomplete information retrieval, inaccurate review results, low evaluation efficiency, insufficient strategy adaptability, and inefficient multi-role collaboration in traditional evaluation assistance technologies.

Description

Technical Field

[0001] This invention relates to the field of computers, and specifically to a smart evaluation assistance method and system based on a large language model. Background Technology

[0002] In the bidding and evaluation process in fields such as power and engineering construction, traditional evaluation relies on manual review of massive amounts of bid documents, which faces problems such as fragmented information, complex review rules, and inefficient multi-role collaboration. Existing auxiliary evaluation technologies suffer from pain points such as unclear slicing strategies, missing recall content, incorrect scoring, and lengthy call chains. They are unable to simulate the real review logic of evaluation experts and cannot meet the requirements for accurate, efficient, and autonomous evaluation.

[0003] Existing technologies face the following key technical challenges in implementing intelligent assisted bidding evaluation: 1. Incomplete information retrieval problem: Traditional bidding evaluation technology based on RAG (Retrieval-augmented Generation) slice retrieval breaks down the bid documents into fragmented segments for retrieval, severing the contextual logic. This results in the loss of key content such as cross-page tables and chapter-related information. For example, core review information such as "led by a national-level project" in the bid documents for compensation devices is easily missed.

[0004] 2. Inaccurate review results: The external technical black box has an uncertain ability to interpret specific charts and formats, and lacks a deep integration of bidding business knowledge and expert experience, which leads to misunderstanding of the scoring criteria and omission of key information. For example, a utility model patent may be misjudged as an invention patent for scoring, or scoring may exceed the limit of prompt words.

[0005] 3. Low evaluation efficiency: Cross-system and cross-platform data interaction leads to a lengthy call chain (traditional chain up to 4 layers), making logs difficult to parse and problems difficult to locate. In addition, it relies too much on high-end GPU computing power, and the review of a single node takes too long. For example, the traditional solution has an average response time of 500 seconds to process a hundred-page bid document.

[0006] 4. Insufficient strategy adaptability: The bidding projects cover different categories (such as insulators and compensation devices) and different review rules. The traditional system strategy configuration is fixed and it is difficult to dynamically adapt to diverse scenarios. Repeated development and adjustment are required, and the update and iteration cycle is long.

[0007] 5. Inefficient multi-role collaboration: Information exchange among multiple roles such as bidding agents, evaluation experts, and suppliers relies on manual transmission. Deviations in the transmission of review rules and untimely feedback lead to process bottlenecks. For example, expert review opinions need to be manually summarized, and rule adjustments require redeploying the system.

[0008] Therefore, it is of great significance to find solutions to the problems of incomplete information retrieval, inaccurate review results, low evaluation efficiency, insufficient strategy adaptability, and inefficient multi-role collaboration in traditional bid evaluation assistance technologies. Summary of the Invention

[0009] To address the problems of incomplete information retrieval, inaccurate review results, low evaluation efficiency, insufficient strategy adaptability, and inefficient multi-role collaboration in existing bid evaluation assistance technologies, this invention proposes a smart bid evaluation assistance method and system based on a large language model.

[0010] Firstly, a smart evaluation assistance method based on a large language model is provided, including: Initialize and map rules based on the received evaluation tasks, and generate corresponding evaluation strategies; Based on the aforementioned review strategy and the bid evaluation documents in the initialized bid evaluation task, the full information is extracted using a pre-built large language model for semantic localization, and a two-layer review information database is constructed based on the full information. Based on the aforementioned two-layer review information database, a multi-role intelligent agent collaborative mechanism is used to carry out division of labor and information interaction, generating scoring results and explanations of the basis; Based on the scoring results and the explanation of the basis, a review report containing the traceability basis is generated and fed back to the business system. The intelligent agent experience base corresponding to the multi-role intelligent agent collaboration mechanism is updated according to the feedback results.

[0011] Secondly, a smart evaluation assistance system based on a large language model is provided, including: The mapping module is used to initialize and map rules based on the received evaluation tasks, and generate corresponding evaluation strategies. The extraction module is used to extract full information based on the review strategy and the evaluation documents in the initialized evaluation task using a pre-built large language model for semantic localization, and to build a two-layer review information database based on the full information. The collaboration module is used to perform division of labor and information interaction based on the two-layer review information database through a multi-role intelligent agent collaboration mechanism, and generate scoring results and explanations of the basis. The generation module is used to generate a review report containing traceability basis based on the scoring results and the explanation, and feed it back to the business system. The module also updates the agent experience base corresponding to the multi-role agent collaboration mechanism based on the feedback results.

[0012] In another aspect, this application also provides an electronic device, comprising: at least one processor and a memory; the memory and the processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, a smart evaluation assistance method based on a large language model, as described above, is implemented.

[0013] Furthermore, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed, implements the intelligent evaluation assistance method based on a large language model as described above.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention provides a smart evaluation assistance method and system based on a large language model. The method initializes the received evaluation task and generates a corresponding evaluation strategy through rule mapping. Then, based on this evaluation strategy and the evaluation documents in the initialized evaluation task, it uses a pre-built large oracle model for semantic localization to extract full information. A two-layer evaluation information database is constructed based on this full information. Based on this database, a multi-role intelligent agent collaborative mechanism is used for division of labor, collaboration, and information exchange to generate scoring results and supporting explanations. Finally, a review report containing traceability evidence is generated based on the scoring results and supporting explanations and fed back to the business system. The intelligent agent experience database corresponding to the multi-role intelligent agent collaborative mechanism is updated according to the feedback results. This solves the problems of incomplete information retrieval, inaccurate review results, low evaluation efficiency, insufficient strategy adaptability, and inefficient multi-role collaboration in traditional evaluation assistance technologies. Attached Figure Description

[0015] Figure 1 The flowchart shows the intelligent evaluation assistance method based on a large language model according to the present invention. Figure 2 This is a schematic diagram of the architecture of the intelligent bid evaluation assistance method based on a large language model according to the present invention. Figure 3 This is a schematic diagram of the hybrid retrieval content extraction process of the intelligent evaluation assistance method based on a large language model according to the present invention; Figure 4 This is a schematic diagram of the multi-role intelligent agent collaborative review process of the intelligent evaluation assistance method based on a large language model according to the present invention; Figure 5 This is a schematic diagram of the intelligent evaluation and evaluation assistance system based on a large language model according to the present invention; Figure 6 This is a schematic diagram of an electronic device structure according to the present invention. Detailed Implementation

[0016] With the accelerated digital transformation of bidding and tendering, the requirements for the efficiency and accuracy of bid evaluation in fields such as power and engineering construction are increasing. The traditional bid evaluation process relies on manual work to complete tasks such as reviewing bid documents, matching evaluation rules, and calculating scores. Faced with bid documents that are often hundreds of pages long, this is not only time-consuming and labor-intensive, but also prone to evaluation bias due to subjective judgment and information omissions.

[0017] The existing solutions described above still have the following key shortcomings when dealing with document generation in the bidding and tendering field: 1. Insufficient information integrity: Due to the use of fixed-length slice retrieval, logically related information across pages and chapters is broken. For example, the related information of "project leading unit + achievement transformation and application" in the tender document is split, resulting in the lack of recall of core review points. 2. Limited scoring accuracy: It lacks in-depth integration of bidding business knowledge, relies solely on keyword matching, and cannot identify key qualifiers such as "lead" and "national level". Furthermore, the external model has black box risks and is prone to errors when parsing complex bid documents (such as image certificates and multi-level tables). 3. Low processing efficiency: The retrieval and scoring process is lengthy, cross-system data interaction is frequent, the average processing time for 100-page tender documents exceeds 500 seconds, and it relies excessively on high-end computing power, resulting in high deployment costs; 4. Poor scenario adaptability: The rule base is fixed. When facing different categories (such as the financial status of insulators, the innovation capability of compensation devices) review details, it is necessary to manually adjust the rules and slicing strategies, resulting in low reusability. 5. Weak collaboration capabilities: A multi-role collaboration mechanism has not been established. The transmission of review rules, expert feedback, and supplier Q&A still rely on manual processes, resulting in long process cycles and a high risk of information discrepancies.

[0018] To address this issue, the industry has gradually introduced artificial intelligence technology, resulting in three main technological paths: 1. RAG Slicing Retrieval-Based Bidding Evaluation System: This system constructs a knowledge base by converting the full text of the bid documents into text, slicing and vectorizing them, and then combining this with a large-scale model for semantic retrieval and scoring. Representative technologies include a combination of traditional vector databases and general-purpose large-scale models, with the core reliance on fragmented information matching for evaluation.

[0019] Representative technologies include: vector databases (such as Milvus and FAISS), general large models (such as GPT-3.5 and LLaMA), text slicing tools (such as LangChain's TextSplitter), and semantic similarity calculation tools (such as Sentence-BERT).

[0020] Limitations: It heavily relies on fixed slice lengths, which can easily fragment cross-page tables and chapter-related information, leading to incomplete recall of key review points; it has weak processing capabilities for non-textual information such as images and complex charts in tender documents, requiring additional reliance on multimodal models with unstable results; it lacks the injection of professional knowledge in the evaluation field, resulting in low accuracy in recognizing key qualifiers such as "leading a project" and "national-level qualification," which can easily lead to scoring bias; and the knowledge base construction relies on external storage, making it difficult to balance data security and retrieval efficiency.

[0021] 2. External Large-Scale Model-Based Assisted Evaluation Scheme: This scheme calls APIs from third-party large-scale models such as Baidu and Alibaba to transform evaluation requirements into prompt words, and then the external model generates the evaluation results. A representative technology is the use of general large-scale model API calls combined with simple rule validation, with the core reliance on the semantic understanding capabilities of the external model.

[0022] Key technical issues include: reliance on external platforms for core review capabilities, posing a technical black-box risk; poor adaptability to specialized scenarios such as "insulator financial indicator verification" and "compensation device innovation capability scoring"; significant impact of network stability on the call chain, leading to response delays or call failures in high-concurrency scenarios; inability to deeply integrate with bidding business rules (such as score cap control and logical association of review items), resulting in generated results often exceeding the scoring criteria; data upload to external platforms, posing a risk of leakage of trade secrets and corporate financial data in bid documents; and high long-term costs due to billing based on the number of calls.

[0023] 3. Evaluation system based on traditional rule engine: It matches and verifies structured data (such as financial indicators and qualification certificate numbers) in the bid documents by manually pre-setting the review rule base. It relies on fixed rules and template filling and is suitable for objective review items in simple scenarios.

[0024] Representative technologies include: rule engines (such as Drools and JBossRules), structured data query tools (such as SQL and ExcelVLOOKUP), fixed template generation tools (such as JasperReports), and form validation tools (such as Apache CommonsValidator). Limitations: It heavily relies on predefined rules and structured templates, and cannot parse unstructured text (such as technical solution descriptions and research content); rule base updates require manual coding and adjustment by technical personnel, resulting in long response times and poor adaptability when facing new review details (such as scoring items related to new power systems); it lacks logical reasoning ability and cannot judge complex review scenarios such as "the relevance between the patent and the project" and "the reasonableness of financial data"; multi-role collaboration relies on manual transmission of rule documents, which can easily lead to misunderstandings of the rules and inconsistent results in different review stages; it only supports fixed categories of bidding projects, and expanding to new categories (such as power services and insulation materials) requires the redevelopment of an entire set of rules, resulting in high maintenance costs.

[0025] The existing solutions described above still have the following key shortcomings when dealing with document generation in the bidding and tendering field: 1. Insufficient information integrity: Due to the use of fixed-length slice retrieval, logically related information across pages and chapters is broken. For example, the related information of "project leading unit + achievement transformation and application" in the tender document is split, resulting in the lack of recall of core review points. 2. Limited scoring accuracy: It lacks in-depth integration of bidding business knowledge, relies solely on keyword matching, and cannot identify key qualifiers such as "lead" and "national level". Furthermore, the external model has black box risks and is prone to errors when parsing complex bid documents (such as image certificates and multi-level tables). 3. Low processing efficiency: The retrieval and scoring process is lengthy, cross-system data interaction is frequent, the average processing time for 100-page tender documents exceeds 500 seconds, and it relies excessively on high-end computing power, resulting in high deployment costs; 4. Poor scenario adaptability: The rule base is fixed. When facing different categories (such as the financial status of insulators, the innovation capability of compensation devices) review details, it is necessary to manually adjust the rules and slicing strategies, resulting in low reusability. 5. Weak collaboration capabilities: A multi-role collaboration mechanism has not been established. The transmission of review rules, expert feedback, and supplier Q&A still rely on manual processes, resulting in long process cycles and a high risk of information discrepancies.

[0026] The substantive improvement proposed by this invention to solve the above-mentioned technical problems lies in: 1. Hybrid retrieval mechanism to solve information integrity problem: The hybrid retrieval method of "structured positioning + fragmented filling" is adopted. First, the document chapter is located by using the semantic alignment capability of the large model to extract complete chapter information to avoid contextual fragmentation. Then, RAG slicing retrieval is used to supplement detailed information to ensure full recall of cross-page tables and related text.

[0027] 2. Autonomous decision-making intelligent agents improve scoring accuracy: Construct a multi-role intelligent agent consisting of "planner, information retrieval officer, information analyst, and reviewer", inject the experience and business rules of evaluation experts, and strengthen the identification of key limiting words through prompt word templates. The consistency between scoring results and expert conclusions is improved to over 90%.

[0028] 3. Improved efficiency through simplified call paths and optimized scheduling: Self-developed core algorithms build independent services, simplifying the call path from four layers ("business system → knowledge base → multimodal model → external platform → large model") to two layers ("ECP / ESS → autonomous decision-making agent → large model"), reducing the processing time of 100-page tender documents to less than 360 seconds; through multi-node load distribution, single-node multi-threaded processing, and on-demand calling of multiple models, the concurrency capability of a single node is increased to 20-30 projects.

[0029] 4. Enhanced business configuration for improved scenario adaptability: The platform supports configuring prompt word templates and case studies through a visual interface. For example, when adding a new category such as "Power Service Review", only the review details and 3-5 sample files need to be uploaded. The system will automatically generate an adaptation strategy, eliminating the need for repeated development. The update and iteration cycle is shortened to 1-2 days, and reusability is improved by more than 80%.

[0030] 5. Security is ensured by independent controllability and compliant design: The self-developed content extraction algorithm and decision-making intelligence agent have core capabilities that do not rely on external platforms. Data storage adopts "memory vector caching + local encrypted storage" to avoid leakage of trade secrets. The review results can be traced back to specific chapters and information sources, and reports containing the scoring calculation process and deviation explanations are generated to meet regulatory compliance requirements.

[0031] 6. Multi-role collaborative optimization process: Establish a multi-role interaction channel for bidding agents (strategy configuration), evaluation experts (result review), and suppliers (question response). Through structured message queues, rules are issued in real time and feedback is synchronized, shortening the process cycle by 30% and avoiding information transmission deviations.

[0032] This invention achieves complete capture of bid document information through hybrid retrieval, relies on multi-role intelligent agents to collaboratively implement review logic, combines business configuration to adapt to diverse bid evaluation scenarios, simultaneously optimizes the call chain to improve efficiency, and ensures compliance with secure storage and traceability design, forming an integrated optimization closed loop covering information extraction, intelligent review, scenario adaptation, efficiency, and security. Figure 2 As shown.

[0033] To better understand the present invention, the following description, in conjunction with the accompanying drawings and embodiments, will further illustrate the content of the present invention.

[0034] Example 1: A smart evaluation assistance method based on a large language model, such as Figure 1 As shown, it includes: Step 1: Initialize and map rules based on the received evaluation tasks to generate corresponding evaluation strategies; Step 2: Based on the review strategy and the evaluation documents in the initialized evaluation task, extract all information using the pre-built large language model semantic positioning, and construct a two-layer review information database based on the all information; Step 3: Based on the aforementioned two-layer review information database, a multi-role intelligent agent collaborative mechanism is used to carry out division of labor, cooperation, and information exchange to generate scoring results and explanations of the basis; Step 4: Based on the scoring results and the explanation of the basis, generate a review report containing the traceability basis and feed it back to the business system. Update the intelligent agent experience base corresponding to the multi-role intelligent agent collaboration mechanism according to the feedback results.

[0035] In this embodiment, step 1 involves directly connecting to the ECP / ESS (E-procurement & C-commerce Platform / E-bid Submission System) business system to receive the bid evaluation task. Based on this task, initialization and rule mapping are performed to lay the foundation for the subsequent construction of the review information database. Specifically, this includes: Receive review guidelines documents and sample tender documents uploaded by users through a visual interface; The review rules are automatically parsed using a large language model to generate a review strategy adapted to the review task. The review strategy includes prompt word templates, case sets, and scoring rules.

[0036] Specifically, the business system interaction and strategy loading include: connecting to the ECP / ESS business system, receiving the evaluation task (including project type, review details, and bid documents), loading the pre-configured review strategy (prompt word template, case set, and scoring rules), and completing the task initialization and rule mapping.

[0037] In this embodiment, after initializing the received bid evaluation task and generating the corresponding review strategy through rule mapping in step 1, the bid documents in the initialized bid evaluation task can be used to extract various types of information such as text, tables, and images from the bid documents through layout detection, structured positioning, and fragmentation filling. This constructs a complete and secure review information database, providing a foundation for subsequent multi-role intelligent agent collaborative bid evaluation. Specifically, this includes: The layout of the bid documents in the bidding task is detected by using photovoltaic character recognition algorithm and document structure parsing algorithm, which identifies the title level, table boundaries and image positions, and generates a structured directory index. By leveraging the semantic alignment capabilities of a pre-built large language model, the core review items in the review strategy are matched with the structured directory index to locate and extract the corresponding complete text information, which includes chapter text, table data, and image association descriptions. For the detailed information not covered by structured localization, a dynamic length slicing strategy and RAG technology are used for vectorized retrieval to fill in the gaps and obtain fragmented detailed information. The complete text information and fragmented detailed information are used as the total information, and a two-layer review information database is obtained based on the total information.

[0038] In one specific embodiment, to address the information fragmentation problem caused by traditional content extraction relying solely on slice retrieval and ignoring document structure, the implementation logic is as follows: Figure 3As shown, the process specifically includes a three-level extraction workflow: "layout detection - structured positioning - fragmented gap filling" to achieve full information capture. First, relying on OCR and document structure parsing algorithms, it identifies layout information such as heading levels, table boundaries, and image positions in the tender documents, generating a structured table of contents index. Second, through the semantic alignment capabilities of a large language model, it matches core review items in the review rules (such as "leading national-level projects" and "number of invention patents") with the table of contents index, locating the corresponding chapters and extracting the entire chapter's text, table data, and image association descriptions, avoiding contextual fragmentation. Finally, for detailed information not covered by structured positioning, a dynamic length slicing strategy (adaptively adjusting the slice length based on text semantic breakpoints) is adopted, and RAG technology is used for vectorized retrieval to fill gaps, forming a two-layer review information database of "complete chapters + detailed fragments." This workflow supports processing tender documents in multiple formats such as Word and PDF. Non-text information (such as qualification certificate images) is converted into searchable vectors through multimodal feature extraction, ensuring no information is missed.

[0039] In this embodiment, after constructing the two-layer review information database based on step 2 above, the database can be used to perform task division and information interaction through a structured message queue, relying on a multi-role intelligent agent set (comprehensive review intelligent agent, tender document parsing intelligent agent, bid document review intelligent agent, indicator calculation intelligent agent, and risk identification intelligent agent). This simulates expert review logic to complete scoring and risk identification, specifically including: Based on the multi-role intelligent agent collaborative mechanism, the tender document parsing intelligent agent breaks down the review details in the review strategy into a standardized review task list and marks the key qualifying words. Based on the multi-role intelligent agent collaborative mechanism, the tender document review intelligent agent calls the two-layer review information database to verify the objective items in the task list of the bid evaluation task and output the basic verification results. Based on the multi-role intelligent agent collaborative mechanism, the indicator calculation intelligent agent combines the preset score rules to calculate and score the preset quantitative review items and identify potential calculation errors. The risk identification agent in the multi-role intelligent agent collaboration mechanism monitors scenarios of scoring exceeding standards or data logic contradictions and outputs risk warnings. Based on the comprehensive review agent in the multi-role intelligent agent collaborative mechanism, a pre-set evaluation expert experience database is integrated, the output results of each intelligent agent are summarized, and a scoring result and explanation of the basis are generated.

[0040] In one specific embodiment, such as Figure 4As shown, the multi-role intelligent agent collaboration mechanism of the present invention is based on the "planner-executor" architecture. It simulates the expert review logic through the division of labor and cooperation of multiple intelligent agents, and the core interaction relies on a structured message queue to achieve asynchronous communication. The tender document parsing agent first analyzes the review rules, breaking them down into actionable scoring items (such as "2 points for meeting financial indicators" and "1-3 points for innovative technology application"), and marks them with key qualifiers ("national level" and "leading role"), generating a standardized review task list. The tender document review agent calls the review information database to verify the objective items in the task list item by item (such as qualification certificate number and financial data integrity), and outputs the basic verification results. The indicator calculation agent combines preset scoring rules to automatically score quantitative items (such as the number of patents and the level of the project), and simultaneously identifies potential errors such as "misjudging a utility model patent as an invention patent". The risk identification agent issues warnings for scenarios such as "scoring beyond the scoring standard" and "data logic contradictions", generating risk prompts. The comprehensive review agent summarizes the outputs of all agents, integrates the experience database of evaluation experts (such as historical review cases and dispute handling rules), and generates the final scoring results and explanations of the basis. The entire process is achieved through a structured message queue to realize task distribution and result feedback, ensuring zero-delay collaboration.

[0041] In this embodiment, after generating the scoring results and explanations based on the aforementioned step 3, a review report containing traceability can be generated based on the scoring results and explanations. The traceable review results (including score details, information sources, and deviation explanations) are then sent back to the ECP / ESS business system to support expert review and result export, forming a closed loop of "review-feedback-optimization".

[0042] By combining the above technologies, the goal is to improve the accuracy of bid evaluation to over 90%, increase the efficiency of single-node review by over 28%, and significantly enhance the adaptability to multiple product categories and scenarios.

[0043] In this embodiment, to ensure the security and compliance of the core algorithm, a dual-mode storage mechanism can be used to store the data, which specifically includes: The review report is stored in a dual-mode system, combining in-memory vector caching with local encrypted storage. The traceability basis includes the specific chapter position of each scoring item in the tender document and fragments of the original information source.

[0044] In this embodiment, to improve the efficiency of reviewing tender documents, a multi-node load balancing and single-node multi-threaded processing approach can be adopted to reduce document processing time. Specifically, this includes: A high-concurrency processing architecture is constructed by employing a multi-node load distribution strategy and a single-node multi-threaded parallel processing technology. By simplifying the call chain into a two-layer architecture and combining it with the high-concurrency processing architecture, single-node concurrent processing is achieved. The two-layer architecture consists of a business system, an autonomous decision-making intelligent agent, and a large model.

[0045] The present invention specifically includes the following beneficial effects: 1. Significantly improved information completeness: Through a hybrid retrieval mechanism of "structured positioning + fragmented gap filling", the recall rate of key content such as cross-page tables and chapter-related information reaches over 99%, completely solving the problem of missing core review points caused by traditional slice retrieval. For example, related information such as "national-level project leadership + achievement transformation" can be fully captured.

[0046] 2. Significantly improved scoring accuracy: Multi-role intelligent agents are infused with bidding business knowledge and expert experience, achieving a key qualifier recognition accuracy of over 95%, and improving the consistency between scoring results and expert manual review to over 90%, effectively avoiding errors such as "misjudgment of patent type" and "scoring beyond the standard," reducing the review deviation rate by 60%.

[0047] 3. Significantly optimized bid evaluation efficiency: The call chain is simplified from 4 layers to 2 layers. Combined with multi-node load distribution and single-node multi-threaded processing, the average processing time of 100-page bid documents is reduced to less than 360 seconds. The concurrent processing capacity of a single node is increased to 20-30 projects. The overall bid evaluation process cycle is shortened by 30%. It does not rely on high-end GPU computing power, and the deployment cost is reduced by 40%.

[0048] 4. Enhanced flexibility in scenario adaptability: Supports visual configuration of review strategies. When adding new categories or updating review rules, only 3-5 sample files need to be uploaded, and the system can automatically generate an adaptation strategy within 1-2 days. No repeated development is required, and scenario reusability is increased by more than 80%, adapting to the evaluation needs of multiple categories.

[0049] 5. Highly efficient and smooth multi-role collaboration: Through structured message queues, real-time information exchange is achieved among bidding agents, evaluation experts, and suppliers. Review rules are transmitted without deviation, expert feedback is responded to synchronously, manual transmission costs are reduced by 70%, and the process bottleneck rate is reduced to below 5%, completely solving the inefficiency problem caused by traditional manual transmission.

[0050] 6. Data security and compliance assurance: Core technologies are independently developed and do not rely on external platforms. Data adopts a dual mode of "memory vector caching + local encrypted storage", and the risk of leakage of trade secrets and financial data is zero. The review results can be traced back to specific chapters and information sources, fully meeting regulatory compliance requirements, with a compliance rate of 100%.

[0051] Example 2: Based on the same inventive concept, this invention also provides a smart evaluation assistance system based on a large language model, such as... Figure 5 As shown, it includes: The mapping module is used to initialize and map rules based on the received evaluation tasks, and generate corresponding evaluation strategies. The extraction module is used to extract full information based on the review strategy and the evaluation documents in the initialized evaluation task using a pre-built large language model for semantic localization, and to build a two-layer review information database based on the full information. The collaboration module is used to perform division of labor and information interaction based on the two-layer review information database through a multi-role intelligent agent collaboration mechanism, and generate scoring results and explanations of the basis. The generation module is used to generate a review report containing traceability basis based on the scoring results and the explanation, and feed it back to the business system. The module also updates the agent experience base corresponding to the multi-role agent collaboration mechanism based on the feedback results.

[0052] Preferably, the extraction module is further used for: The layout of the bid documents in the bidding task is detected by using photovoltaic character recognition algorithm and document structure parsing algorithm, which identifies the title level, table boundaries and image positions, and generates a structured directory index. By leveraging the semantic alignment capabilities of a pre-built large language model, the core review items in the review strategy are matched with the structured directory index to locate and extract the corresponding complete text information, which includes chapter text, table data, and image association descriptions. For the detailed information not covered by structured localization, a dynamic length slicing strategy and RAG technology are used for vectorized retrieval to fill in the gaps and obtain fragmented detailed information. The complete text information and fragmented detailed information are used as the total information, and a two-layer review information database is obtained based on the total information.

[0053] Preferably, the collaborative module is further configured to: Based on the multi-role intelligent agent collaborative mechanism, the tender document parsing intelligent agent breaks down the review details in the review strategy into a standardized review task list and marks the key qualifying words. Based on the multi-role intelligent agent collaborative mechanism, the tender document review intelligent agent calls the two-layer review information database to verify the objective items in the task list of the bid evaluation task and output the basic verification results. Based on the multi-role intelligent agent collaborative mechanism, the indicator calculation intelligent agent combines the preset score rules to calculate and score the preset quantitative review items and identify potential calculation errors. The risk identification agent in the multi-role intelligent agent collaboration mechanism monitors scenarios of scoring exceeding standards or data logic contradictions and outputs risk warnings. Based on the comprehensive review agent in the multi-role intelligent agent collaborative mechanism, a pre-set evaluation expert experience database is integrated, the output results of each intelligent agent are summarized, and a scoring result and explanation of the basis are generated.

[0054] Preferably, the mapping module is further configured to: Receive review guidelines documents and sample tender documents uploaded by users through a visual interface; The review rules are automatically parsed using a large language model to generate a review strategy adapted to the review task. The review strategy includes prompt word templates, case sets, and scoring rules.

[0055] Preferably, the system further includes: The efficiency optimization module is used to build a high-concurrency processing architecture by employing multi-node load distribution strategies and single-node multi-threaded parallel processing technology. By simplifying the call chain into a two-layer architecture and combining it with the high-concurrency processing architecture, single-node concurrent processing is achieved. The two-layer architecture consists of a business system, an autonomous decision-making intelligent agent, and a large model.

[0056] Preferably, the system further includes: The security and compliance module is used to store the review report in a dual mode, combining memory vector caching with local encrypted storage. The traceability basis includes the specific chapter position of each scoring item in the tender document and fragments of the original information source.

[0057] Example 3 like Figure 6 As shown, the present invention also provides an electronic device, which may be a computer device, a microcontroller device, a smart mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, processor, and transceiver component are connected via a bus; the memory can be used to store executable programs, and an exemplary executable program may include instructions; the processor is used to execute the instructions stored in the memory. The memory can also be used to store data, which can be accessed and / or modified when instructions are executed.

[0058] The processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, and it is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the storage medium to implement the corresponding method flow or corresponding function, so as to realize the steps of the intelligent evaluation assistance method based on a large language model in the above embodiments.

[0059] Example 4 Based on the same inventive concept, this invention also provides a readable storage medium, specifically an electronic device readable storage medium (Memory). This readable storage medium is a memory device within an electronic device used to store programs and data. It is understood that the storage medium here can include both built-in storage media within the electronic device and extended storage media supported by the electronic device. The storage medium provides storage space, which stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more executable programs (including program code). It should be noted that the storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. Loading and executing one or more instructions stored in the storage medium by the processor can implement the steps of the intelligent evaluation assistance method based on a large language model in the above embodiments.

[0060] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0061] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0062] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0063] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0064] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of the claims of the present invention pending approval.

Claims

1. A smart evaluation assistance method based on a large language model, characterized in that, include: Initialize and map rules based on the received evaluation tasks, and generate corresponding evaluation strategies; Based on the aforementioned review strategy and the bid evaluation documents in the initialized bid evaluation task, the full information is extracted using a pre-built large language model for semantic localization, and a two-layer review information database is constructed based on the full information. Based on the aforementioned two-layer review information database, a multi-role intelligent agent collaborative mechanism is used to carry out division of labor and information interaction, generating scoring results and explanations of the basis; Based on the scoring results and the explanation of the basis, a review report containing the traceability basis is generated and fed back to the business system. The intelligent agent experience base corresponding to the multi-role intelligent agent collaboration mechanism is updated according to the feedback results.

2. The method according to claim 1, characterized in that, The evaluation strategy and the evaluation documents in the initialized evaluation task are used to extract full information using a pre-built large language model, and a two-layer evaluation information database is constructed based on the full information. include: The layout of the bid documents in the bidding task is detected by using photovoltaic character recognition algorithm and document structure parsing algorithm, which identifies the title level, table boundaries and image positions, and generates a structured directory index. By leveraging the semantic alignment capabilities of a pre-built large language model, the core review items in the review strategy are matched with the structured directory index to locate and extract the corresponding complete text information, which includes chapter text, table data, and image association descriptions. For the detailed information not covered by structured localization, a dynamic length slicing strategy and RAG technology are used for vectorized retrieval to fill in the gaps and obtain fragmented detailed information. The complete text information and fragmented detailed information are used as the total information, and a two-layer review information database is obtained based on the total information.

3. The method according to claim 1, characterized in that, The process of generating scoring results and explanations based on the two-layer review information database, through a multi-role intelligent agent collaborative mechanism, involves division of labor, cooperation, and information exchange, including: Based on the multi-role intelligent agent collaborative mechanism, the tender document parsing intelligent agent breaks down the review details in the review strategy into a standardized review task list and marks the key qualifying words. Based on the multi-role intelligent agent collaborative mechanism, the tender document review intelligent agent calls the two-layer review information database to verify the objective items in the task list of the bid evaluation task and output the basic verification results. Based on the multi-role intelligent agent collaborative mechanism, the indicator calculation intelligent agent combines the preset score rules to calculate and score the preset quantitative review items and identify potential calculation errors. The risk identification agent in the multi-role intelligent agent collaboration mechanism monitors scenarios of scoring exceeding standards or data logic contradictions and outputs risk warnings. Based on the comprehensive review agent in the multi-role intelligent agent collaborative mechanism, a pre-set evaluation expert experience database is integrated, the output results of each intelligent agent are summarized, and a scoring result and explanation of the basis are generated.

4. The method according to claim 1, characterized in that, The process of loading a pre-configured review strategy based on the received evaluation task, initializing the evaluation task, and mapping rules includes: Receive review guidelines documents and sample tender documents uploaded by users through a visual interface; The review rules are automatically parsed using a large language model to generate a review strategy adapted to the review task. The review strategy includes prompt word templates, case sets, and scoring rules.

5. The method according to claim 1, characterized in that, The method further includes: A high-concurrency processing architecture is constructed by employing a multi-node load distribution strategy and a single-node multi-threaded parallel processing technology. By simplifying the call chain into a two-layer architecture and combining it with the high-concurrency processing architecture, single-node concurrent processing is achieved. The two-layer architecture consists of a business system, an autonomous decision-making intelligent agent, and a large model.

6. The method according to claim 1, characterized in that, After generating the review report containing the traceability basis based on the scoring results and the explanation, the process also includes: The review report is stored in a dual-mode system, combining in-memory vector caching with local encrypted storage. The traceability basis includes the specific chapter position of each scoring item in the tender document and fragments of the original information source.

7. A smart evaluation and evaluation assistance system based on a large language model, characterized in that, include: The mapping module is used to initialize and map rules based on the received evaluation tasks, and generate corresponding evaluation strategies. The extraction module is used to extract full information based on the review strategy and the evaluation documents in the initialized evaluation task using a pre-built large language model for semantic localization, and to build a two-layer review information database based on the full information. The collaboration module is used to perform division of labor and information interaction based on the two-layer review information database through a multi-role intelligent agent collaboration mechanism, and generate scoring results and explanations of the basis. The generation module is used to generate a review report containing traceability basis based on the scoring results and the explanation, and feed it back to the business system. The module also updates the agent experience base corresponding to the multi-role agent collaboration mechanism based on the feedback results.

8. The system according to claim 7, characterized in that, The extraction module is further used for: The layout of the bid documents in the bidding task is detected by using photovoltaic character recognition algorithm and document structure parsing algorithm, which identifies the title level, table boundaries and image positions, and generates a structured directory index. By leveraging the semantic alignment capabilities of a pre-built large language model, the core review items in the review strategy are matched with the structured directory index to locate and extract the corresponding complete text information, which includes chapter text, table data, and image association descriptions. For the detailed information not covered by structured localization, a dynamic length slicing strategy and RAG technology are used for vectorized retrieval to fill in the gaps and obtain fragmented detailed information. The complete text information and fragmented detailed information are used as the total information, and a two-layer review information database is obtained based on the total information.

9. An electronic device, characterized in that, include: At least one processor and memory; The memory and processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, the intelligent evaluation assistance method based on a large language model as described in any one of claims 1 to 6 is implemented.

10. A readable storage medium, characterized in that, It contains an execution program, which, when executed, implements the intelligent evaluation assistance method based on a large language model as described in any one of claims 1 to 6.

Images