An intelligent assistance method for postgraduate project system training scenarios

Abstract

This invention discloses an intelligent assistance method for postgraduate project-based training scenarios, comprising the following steps: collecting user training process data from a training management platform to determine the user's training progress status; constructing a knowledge resource set, and labeling the visible stage of each resource in the knowledge resource set based on the correspondence between resource content features and training progress status; performing candidate resource filtering on the knowledge resource set based on the user's permission level and training progress status to obtain a candidate resource set; receiving the user's query statement, and sorting the candidate resource set by the semantic similarity between the query statement and the candidate resource set and the BM25 retrieval score to obtain the retrieval context fragments corresponding to the top N candidate resources; injecting the query statement and retrieval context fragments into a prompt word template, driving a large language model to generate structured output containing intent identifiers; and calling a business interface to execute the business operation corresponding to the intent identifier.

Description

Technical Field

[0001] This invention relates to the field of educational informatization technology, specifically to an intelligent assistance method for postgraduate project-based training scenarios. Background Technology

[0002] In recent years, with the rapid development of artificial intelligence technology, the application of Retrieval-Augmented Generation (RAG) technology in the education field has experienced explosive growth, especially in scenarios such as educational question-answering systems and teaching aids. The current mainstream technical implementation path follows a three-stage process: user question → information retrieval → generation model outputs textual answers. This paradigm effectively alleviates the knowledge lag and illusion problems of traditional generation models, providing a new technological foundation for personalized education services.

[0003] For the construction of RAG systems in educational scenarios, existing technical solutions mainly revolve around three dimensions: knowledge base construction, retrieval algorithm optimization, and generative model fine-tuning. At the data level, researchers and developers are attempting to integrate multimodal educational resources such as course materials, academic literature, and instructional video subtitles. At the retrieval level, hybrid retrieval strategies combining sparse and dense vector retrieval are gradually becoming mainstream. At the generative level, instruction fine-tuning and alignment techniques are used to improve the answer quality of large language models in educational professional fields. Some advanced systems have also introduced user profile tags to achieve preliminary personalized content filtering, which to some extent improves the disconnect between general question-answering systems and specific educational scenarios.

[0004] However, existing educational RAGs and online education Q&A systems still have significant limitations in terms of deep integration of technical architecture and business. First, at the data coupling level, existing solutions mostly use general question-and-answer patterns or coarse-grained personalization based on simple tags, failing to establish a deep, structured coupling mechanism between graduate student training process data and knowledge retrieval. This makes it difficult to combine dynamic process data such as students' training stages, research project participation, and credit achievement progress for targeted reasoning and precise matching. Second, at the output level, Q&A results are usually limited to static text answers, lacking actionability. They cannot directly trigger business processes within the platform such as lab reservations, course selection and withdrawal, visual viewing of training progress, and personalized achievement path planning, making it difficult to form a closed-loop guidance mechanism. Furthermore, most solutions only provide text answers or resource link lists, lacking visualized learning paths and next-step action guidance deeply integrated with platform business actions. This prevents the transformation of knowledge retrieval results into actionable training guidance plans, hindering the intelligent teaching assistant's transition from an information provider to a learning facilitator. Summary of the Invention

[0005] This invention proposes an intelligent assistance method for postgraduate project-based training scenarios, which solves the problem that existing technologies fail to deeply couple postgraduate training process data with knowledge retrieval, resulting in the inability to form a closed-loop training guidance.

[0006] To address the aforementioned technical problems, this invention provides an intelligent assistance method for postgraduate project-based training scenarios, comprising the following steps:

[0007] Step S1: Collect user's cultivation process data from the cultivation management platform to determine the user's cultivation progress status; construct a knowledge resource set, and mark the visible stage of each resource in the knowledge resource set based on the correspondence between the resource content characteristics and the cultivation progress status, so that each resource in the knowledge resource set is associated with at least one cultivation progress status;

[0008] Step S2: Based on the user's permission level and the training progress status, perform candidate resource filtering on the knowledge resource set to obtain a candidate resource set;

[0009] Step S3: Receive the user's query statement, sort the candidate resource set by the semantic similarity between the query statement and the candidate resource set and the BM25 retrieval score, and obtain the retrieval context fragments corresponding to the top N candidate resources;

[0010] Step S4: Inject the query statement and retrieval context fragment into the prompt word template to drive the large language model to generate structured output containing intent identifiers;

[0011] Step S5: Call the business interface to execute the business operation corresponding to the intent identifier.

[0012] Preferably, step S1, which involves collecting user training process data from the training management platform to determine the user's training progress status, includes the following steps: collecting user training process data from the training management platform, constructing a user status feature vector that includes skill learning completion rate, project participation status, experimental course completion rate, permission level, and credit achievement rate, and determining the user's training progress status based on the user status feature vector;

[0013] The skill learning completion rate, lab course completion rate, and credit achievement rate are within a certain range. The project participation status is a continuous dimension, while the project participation status is a discrete dimension representing three states: not participating, participating, and completed. The permission level is a discrete dimension representing three user roles: student, tutor, and administrator.

[0014] After normalizing the discrete dimension, it is combined with the continuous dimension to form a normalized vector. A weight vector is introduced to sum the weighted values ​​of all dimensions in the normalized vector except for the permission level, to obtain a comprehensive score. Based on the comparison between the comprehensive score and a preset threshold, the cultivation progress status is divided into three discrete states:

[0015] ;

[0016] In the formula, For users The progress of cultivation; and These are the lower threshold and the upper threshold, respectively. For users Overall score;

[0017] When the overall score crosses a preset threshold, triggering a state switch, the weight vector is updated using a smooth transition:

[0018] ;

[0019] in, This is the transition weight vector; This is the current weight vector; The target weight vector; This is the smoothing coefficient.

[0020] Preferably, the construction of the knowledge resource set in step S1 includes the following steps:

[0021] Step S11: Collect and segment the system documents, course descriptions, experimental project descriptions, process guidelines, and frequently asked questions of the training management platform to obtain a resource segment set;

[0022] Step S12: Vectorize and index each resource in the resource block set;

[0023] Step S13: Label each resource with a set of metadata fields, wherein the set of metadata fields includes at least the resource unique identifier, resource type, set of visible roles, set of visible stages, validity period, and adaptation condition vector;

[0024] The labeling rules for the visible stage set are as follows: Based on the content characteristics and applicable scenarios of the resources, determine in which training progress stages the resources are visible; resources for basic learning are labeled as visible in the basic preparation stage, resources for project practice and experimental participation are labeled as visible in the process advancement stage, resources for project completion and achievement confirmation are labeled as visible in the task improvement stage, and general resources are labeled as visible in all stages.

[0025] Each dimension of the adaptation condition vector corresponds one-to-one with the user state feature vector.

[0026] Preferably, step S1 further includes periodically fine-tuning the weight vector based on user feedback, including the following steps: collecting user feedback on the recommendation results and quantifying each feedback instance into a feedback score:

[0027] ;

[0028] In the formula, For users The feedback score at time t; For feedback coefficients, ; A binary indicator variable representing user clicks on recommended results; A binary indicator variable representing the user's adoption of the recommendation result; A binary indicator variable representing whether users ignore the recommendation results; A binary indicator variable representing users' negative ratings of recommendation results; A binary indicator variable representing a user's appeal against the recommendation results. coefficient The symbol is determined based on the outcome of the appeal; when the appeal is successful... Take a positive value when the appeal is rejected. Take the negative value; A binary indicator variable representing the successful execution of the business operation corresponding to the recommendation result; A binary indicator variable representing the failure of the business operation corresponding to the recommendation result;

[0029] The update formula for the weight vector is:

[0030] ;

[0031] In the formula, , These are the weights corresponding to the i-th user state feature at time t+1 and time t, respectively; The learning rate; This is the normalized value of the i-th user state feature in this recommendation.

[0032] Preferably, the candidate resource filtering in step S2 includes hard filtering and soft filtering performed sequentially. The hard filtering filters the knowledge resource set based on user role, training progress status, and current timestamp.

[0033] ;

[0034] In the formula, This is the candidate set after hard filtering; For users The level of access control; For users The progress of cultivation; , , , For resources The set of visible roles, the set of visible stages, the valid start date, and the valid end date; t is the current timestamp;

[0035] The soft filter expresses business rules as a set of configurable constraints. Constraint consistency screening is performed on the hard-filtered candidate set:

[0036] ;

[0037] In the formula, This is the candidate set after soft filtering; For constraint sets; Indicates the resource and users Enforcement constraints The verification function.

[0038] Preferably, step S3 involves ranking the candidate resource set based on the semantic similarity between the query statement and the candidate resource set, and the BM25 retrieval score, including the following steps:

[0039] Step S31: Construct semantic embedding vectors for the query statement and each resource in the candidate resource set. Calculate the cosine similarity between the query statement and each resource based on the semantic embedding vectors. Perform translational normalization on the cosine similarity to obtain a semantic score. The expression for the semantic score is:

[0040] ;

[0041] ;

[0042] In the formula, For semantic score; For query statement With candidate resources Cosine similarity between them; , These are the minimum and maximum values ​​of the cosine similarity in the current candidate resource set, respectively. To prevent division by zero of small constants; This is a query statement; Candidate resources; This is the semantic embedding vector of the query statement; This represents the semantic embedding vector of the candidate resource;

[0043] Step S32: Calculate the BM25 keyword retrieval score for each resource in the query statement and the candidate resource set, and perform saturation normalization on the BM25 keyword retrieval score:

[0044] ;

[0045] In the formula, The BM25 keyword retrieval score after saturation normalization; For query statement With candidate resources BM25 keyword search scores between; It is a smoothing constant;

[0046] Step S33: The semantic score and the BM25 keyword retrieval score after saturation normalization are weighted and fused according to the weight parameters to obtain the hybrid retrieval score:

[0047] ;

[0048] In the formula, For mixed search scores; Configurable weight parameters related to resource type;

[0049] Step S34: The final ranking score is obtained by adding a user state matching correction term to the mixed retrieval score.

[0050] ;

[0051] In the formula, The final ranking score; This is a correction factor; The weighted matching degree between the user state feature vector and the resource adaptation condition vector.

[0052] Preferably, the weighted matching degree The calculation method is as follows:

[0053] Resource The adaptation condition vector is Each dimension corresponds one-to-one with the user state feature vector;

[0054] For continuous dimensions :

[0055] ;

[0056] In the formula, For the first Normalized values ​​of individual user state features;

[0057] For discrete dimensions When the user's status meets the resource requirements ,otherwise ;

[0058] Since the permission level dimension is already used as an independent filtering condition, the weighted matching degree calculation excludes the permission level dimension. The expression for the weighted matching degree is:

[0059] ;

[0060] In the formula, This is the weight vector corresponding to the user state feature vector.

[0061] Preferably, the prompt word template in step S4 includes at least:

[0062] User development stage and permission information: The status summary text is automatically generated by concatenating the structured fields of the user status feature vector according to the template;

[0063] Retrieve context fragments and number identifiers;

[0064] Output format constraints: The large language model is required to return the output results according to a predefined JSON structure, wherein the JSON structure shall at least include an explanatory answer field, a recommended resource list field, an executable action list field, a state write-back suggestion field, an evidence citation field, the intent identifier field, and an intent confidence field;

[0065] Prohibited Items Declaration: The large language model is required not to output unauthorized content and not to construct non-existent resources or time periods.

[0066] Preferably, step S4 further includes generating, verifying, and performing fault-tolerant processing on the structured output:

[0067] The output of the large language model is subjected to format validation and value range validation. The format validation includes checking whether the necessary fields exist and are of the correct type. The value range validation includes checking whether the intent confidence is within the range of [0,1], whether the intent identifier belongs to the intent set defined by the system, and whether the resource identifier referenced in the list of executable actions exists in the candidate resource set.

[0068] When the verification fails, a format correction instruction is appended to the prompt word template and the large language model is called again, with a maximum of a preset number of retries; if it still fails after retries, a downgraded response containing only explanatory answers is returned, without triggering business operations.

[0069] Preferably, step S5 includes the following steps:

[0070] Step S51: Compare and determine the intent confidence in the structured output with the configurable trigger threshold corresponding to the intent identifier. Set trigger thresholds for different intent types according to risk level. When the intent confidence does not reach the trigger threshold, it is downgraded to the only recommendation mode.

[0071] Step S52: After the intent confidence level passes the threshold, perform the following checks on the business operation parameters in sequence: permission check, time window check, capacity and quota check, conflict check, approval status check and idempotency check. If any check fails, return the reason for failure and alternative suggestions.

[0072] Step S53: When the intent identifier belongs to the write operation type intent, generate a one-time confirmation token that binds the user identifier, intent identifier and business operation parameter summary, and call the business interface to execute the business operation after the user confirms;

[0073] Step S54: After the business operation is successfully executed, the user state feature vector is written back and updated. The continuous dimension is updated incrementally and pruned to the [0,1] interval, while the discrete dimension is transitioned using a state machine. The comprehensive score and training progress status are recalculated simultaneously.

[0074] The advantages of this invention include at least the following:

[0075] 1. Encode process data such as training stages, project participation, and credit achievement into computable structured features, and dynamically determine the training progress status based on a comprehensive scoring mechanism, so that knowledge retrieval can accurately match the user's actual training situation;

[0076] 2. A dual filtering mechanism based on permission level and training status is introduced to remove resources that users do not have permission to access or that are not applicable at the current stage before semantic retrieval, which significantly improves retrieval efficiency and result usability, while avoiding the business risks of recommending high-level resources to lower-grade students or leaking undisclosed project information.

[0077] 3. By using the intent confidence threshold to determine low-confidence intents caused by model illusions, and combining this with the legality verification of business operation parameters, a safety gate is established before automatic execution. This not only preserves the flexibility of the large model, but also ensures the rigor and traceability of key operations in postgraduate training. Compared with the plain text output of existing technologies, this significantly reduces business risks. Attached Figure Description

[0078] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0079] Figure 2 This is a flowchart of resource filtering and semantic matching in an embodiment of the present invention;

[0080] Figure 3This is a flowchart illustrating the update process of the state vector in an embodiment of the present invention. Detailed Implementation

[0081] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0082] This invention provides an intelligent assistance method for postgraduate project-based training scenarios. This method is deployed in the server environment of a postgraduate training management platform. The platform includes at least a business database, a knowledge base and vector index service, a generative reasoning service, a rule and permission service, an audit log service, and a visualization front-end. The business database includes at least several tables such as a skills learning record table, a project participation record table, an innovative experimental course reservation and attendance table, a grade and credit mapping table, and a user role and organizational domain permission table. The core process of this method is as follows: Figure 1 As shown, it includes the following steps:

[0083] Step S1: Collect user's cultivation process data from the cultivation management platform to determine the user's cultivation progress status; construct a knowledge resource set, and mark the visible stage of each resource in the knowledge resource set based on the correspondence between resource content characteristics and cultivation progress status, so that each resource in the knowledge resource set is associated with at least one cultivation progress status.

[0084] Specifically, user training process data is collected from the training management platform to construct a user status feature vector that includes skill learning completion rate, project participation status, experimental course completion rate, permission level, and credit achievement rate. The user's training progress status is determined based on the user status feature vector.

[0085] The system retrieves data from the business database of the training management platform, categorized by user identifier. Periodically or event-driven extraction of basic feature data related to the cultivation process is used to construct user state feature vectors. This vector contains five dimensions, defined as follows:

[0086] ;

[0087] in This indicates the completion rate of core experimental skills learning, with a value ranging from [0,1]. Its value equals the number of core skill learning tasks completed by the user plus the total number of tasks in the course. The ratio increases with each completed task; this dimension increases. The increment; This represents the project participation status. It is a discrete dimension with a value set of {0, 1, 2}, which correspond to three statuses: not participating, participating, and completed, respectively. This indicates the completion rate of the open innovation experiment course, with a value ranging from [0,1]. Its value is equal to the number of times the user has checked in for completed experiment courses and the total number of experiment courses. The ratio; This represents the permission level, which is a discrete dimension with a value set of {1,2,3}, corresponding to the three user roles of student, tutor and administrator, respectively. This represents the percentage of students who have met the credit requirements, and its value ranges from [0,1]. It is calculated by periodically obtaining actual credit data from the academic affairs system.

[0088] Of the five dimensions mentioned above, , and As a continuous dimension, its values ​​naturally fall within the range of [0,1]. and For discrete dimensions, normalization mapping is required to ensure that features of different dimensions can be weighted and fused. The normalization mapping rules are as follows: for continuous dimensions, the original values ​​are directly retained, i.e. For the project participation status dimension, the following approach is adopted: Normalization mapping is performed using the following method; for the permission level dimension, a normalized mapping is adopted. Perform a normalization mapping. After the above mapping, the normalized vector is obtained:

[0089] ;

[0090] All dimensions in the normalized vector fall within the range of [0,1], providing a unified numerical basis for subsequent weighted fusion calculations.

[0091] The system introduces a weight vector. The system introduces a weight vector. The weight components satisfy the following conditions: and The constraints are as follows. The formula for calculating the overall score is:

[0092] .

[0093] Permission Level It does not participate in the weighted summation of the overall score, but is used as a discrete filtering condition in the subsequent hard filtering stage, specifically for three purposes:

[0094] 1. Used in the resource hard filtering process to determine whether a user role belongs to the set of visible roles of a resource;

[0095] 2. Used for permission verification in the intent triggering stage to check whether the user meets the minimum permission level required for the operation;

[0096] 3. It participates in the calculation of the matching degree between user status and resource adaptation conditions as a binary matching item in the discrete dimension.

[0097] The comprehensive score plays the following roles in subsequent processes: Regarding resource type priority, the system prioritizes resource categories more suitable for the current training progress based on the comprehensive score's range. For example, lower scores prioritize basic learning resources, medium scores prioritize project participation or lab courses, and higher scores prioritize completion or finalization resources. Regarding ranking correction, the degree of matching the comprehensive score's status is added as a ranking correction factor to the original search score, improving the ranking position of resources more closely aligned with the current progress. Regarding controlling the strength of action suggestions, when the comprehensive score is low, the system prioritizes suggestions such as viewing rules and completing basic tasks. Only when the comprehensive score reaches the corresponding range are actionable action suggestions such as booking, course selection, and submission provided, reducing the probability of inappropriate actions being triggered prematurely.

[0098] Based on the comparison between the comprehensive score and a preset threshold, the system divides the user's training progress status into three discrete states:

[0099] ;

[0100] In the formula, For users The progress of cultivation; , These are the lower and upper thresholds, respectively. The recommended default values ​​are... =0.3、 =0.7, which can be configured and adjusted by the system administrator according to the actual training pace of the department.

[0101] The "Basic Preparation" status indicates that the user has not yet met the basic requirements for core learning and credits, and is still in the introductory learning stage. The "Progress" status indicates that the user has acquired basic skills and is progressing with tasks such as project participation and experimental practice. The "Task Completion" status indicates that the user's status in all dimensions is close to meeting the requirements and has entered the final submission stage. The training progress status here does not necessarily come from the school's official training process fields, but is an auxiliary classification result generated by the system after approximate identification based on existing business data such as learning task completion records, project participation records, open innovation experimental course participation records, and credit completion ratios. It is mainly used for resource filtering, sorting correction, and path guidance control.

[0102] The system pre-sets weight templates for different training progress stages to reflect the differentiated impact of each dimension on recommendation decisions at different stages. In the basic preparation stage, the weights of dimensions related to skill learning completion and credit ratio are appropriately increased; the recommended configuration is as follows: During the progress phase, appropriately increase the weighting of dimensions such as project participation and participation in experimental classes; a recommended configuration is [insert configuration here]. When the task is in a complete state, appropriately increase the weight of dimensions related to completion and achievement; a recommended configuration is [configuration value missing]. During system initialization, the weight template corresponding to the basic preparation state is used as the initial weight vector by default.

[0103] When a user's overall score crosses a preset threshold range, triggering a state switch, the system updates the weight vector using a smooth transition method to avoid sudden changes in recommendation results due to abrupt weight shifts.

[0104] ;

[0105] in, This is the transition weight vector; This is the current weight vector; The target weight vector; This is the smoothing coefficient; a default value of 0.5 is recommended, with a range of [0.3, 0.8]. The larger the value, the smoother the transition. The smaller the value, the more agile the transition. This smooth transition mechanism ensures that weight switching is triggered by changes in the progress status driven by existing business data in the system, rather than by drastic changes in recommendation results caused by small fluctuations in the overall score around the threshold.

[0106] Step S1 also includes a process of periodically fine-tuning the weight vector based on user feedback. The system collects user feedback on the recommendation results, including types such as clicks, acceptance, ignore, negative reviews, appeals, successful execution, and failed execution, and quantifies each feedback instance into a feedback score:

[0107] ;

[0108] In the formula, For users The feedback score at time t; For feedback coefficients, This is used to distinguish the strength of different feedback behaviors; A binary indicator variable representing user clicks on recommended results; A binary indicator variable representing the user's adoption of the recommendation result; A binary indicator variable representing whether users ignore the recommendation results; A binary indicator variable representing users' negative ratings of recommendation results; A binary indicator variable representing a user's appeal against the recommendation results. coefficient The symbol is determined based on the outcome of the appeal; when the appeal is successful... Take a positive value when the appeal is rejected. If a negative value is used, the appeal will be reviewed by the administrator and marked as either successful or rejected. A binary indicator variable representing the successful execution of the business operation corresponding to the recommendation result; A binary indicator variable representing the failure of the business operation corresponding to the recommendation result.

[0109] The weight vector is updated using a small-step adjustment method, and the update formula is as follows:

[0110] ;

[0111] In the formula, , These are the weights corresponding to the i-th user state feature at time t+1 and time t, respectively; The learning rate is recommended to be set to a default value of 0.01, with a range of [0.001, 0.05]. A smaller value should be used for weekly updates to avoid weight oscillations. This is the normalized value of the i-th user state feature in this recommendation.

[0112] After the update, each Cut to a value not less than zero, i.e. Then, normalization is performed to ensure the total weight remains constant. In engineering implementation, the system uses a periodic strategy of daily data collection and weekly updates for weight fine-tuning, rather than real-time updates with each interaction. This leverages feedback loops to optimize parameters while avoiding short-term fluctuations that could destabilize the system. When the feedback sample size is insufficient, the system does not perform automatic fine-tuning but maintains the current stable template. For convergence, the system uses engineering criteria, such as determining that parameters have entered a stable state and stopping updates when the parameter change is below a preset threshold for multiple consecutive update cycles, or when the main retrieval indicators do not show significant improvement.

[0113] To support the scalability of the support vector dimension, the system allows the basic state vector to be expanded to... The system can be expanded to include research direction label vectors, capability gap vectors, and time availability vectors. These expanded dimensions are integrated into the system through a unified data source, feature extraction, quantization mapping, and update mechanism interface, without affecting the existing calculation process of the basic five-dimensional vectors. This expansion mechanism allows the system to adapt to the diverse training needs of different departments and professional directions.

[0114] This step constructs a knowledge resource set for subsequent retrieval and filtering. The construction of the knowledge resource set is completed during the system initialization phase and is incrementally maintained during resource updates. The construction result provides the data foundation for candidate resource filtering in step S2 and semantic matching and ranking in step S3.

[0115] Specifically, the system records the knowledge resource set as It encompasses various types of resources, including policy documents, course descriptions, experimental project descriptions, process guidelines, and frequently asked questions (FAQs). The construction of the knowledge resource collection includes three stages: resource acquisition and segmentation, vectorized encoding and index construction, and metadata annotation.

[0116] Resource Acquisition and Segmentation: The system collects various training-related resources from the document library of the training management platform. It segments policy documents, course descriptions, experimental project descriptions, process guidelines, etc., into blocks of fixed length, maintaining a certain overlap between adjacent blocks to reduce semantic loss across segments. The block length is 256 tokens, and the overlap length is 64 tokens.

[0117] Vectorization encoding and index construction: The block results are vectorized and indexed for semantic similarity calculation in subsequent steps.

[0118] Metadata annotation: The system annotates each resource Binding metadata field collection It must contain at least the following fields: unique resource identifier Resource types Visible character set Visible stage set Validity period Related business action types Action Interface Entry Prerequisites Capacity field Conflict determination field Maintenance department and responsible person Version number and update time and adaptation condition vector wait.

[0119] Labeling rules for visible stage sets: Visible stage sets It serves as a crucial link between the knowledge resource set and the user's training progress status. Based on the content characteristics and applicable scenarios of the resources, the system uses the three training progress statuses determined in step S1 as the labeling criteria, and labels each resource with a set of visible stages according to the following rules:

[0120] (1) Resources for basic learning are visible in the basic preparation stage: including introductory learning guides, core skills learning tutorials, training program descriptions, interpretations of basic rules and regulations, etc. These resources mainly serve users whose skills learning completion rate and credit achievement rate have not yet met the standards;

[0121] (2) Resources for project practice and experimental participation are marked as visible in the process advancement stage: including project participation guidelines, instructions for open innovation experimental courses, laboratory reservation rules, project mid-term inspection requirements, etc.

[0122] (3) Resources for project completion submission and achievement confirmation are marked as visible during the task improvement stage: including project completion report templates, credit achievement confirmation process, project acceptance standards, and results submission guidelines.

[0123] (4) General resources are marked as visible at all stages: including FAQs, platform operation manuals, contact information and other resources that do not depend on a specific stage of training progress.

[0124] The above annotation rules are configured by resource maintenance personnel based on the actual content and applicable objects of the resources when they are added to the database. The system provides annotation templates and recommended rules to assist in annotation.

[0125] Step S2: Filter the knowledge resource set based on permission level and training progress status to obtain a candidate resource set.

[0126] This step, based on the user's permission level in the user's state feature vector and the training progress status determined in step S1, performs candidate resource filtering on the knowledge resource set constructed in step S1, resulting in a candidate resource set. Since each resource in the knowledge resource set has been labeled with metadata such as the set of visible roles, the set of visible stages, and the validity period in step S1, the system can use this metadata to match the user's permission level and training progress status, thereby eliminating resources that the user is not authorized to access or that are not applicable at the current stage before semantic retrieval. The filtering process includes two stages: hard filtering and soft filtering, executed sequentially.

[0127] like Figure 2 As shown, hard filtering filters the knowledge resource set based on user role, training progress status, and current timestamp, reducing the interference of irrelevant resources on retrieval and generation from the source. The formal definition of hard filtering is:

[0128] ;

[0129] In the formula, This is the candidate set after hard filtering; For users The permission level is determined by the permission system; For users The cultivation progress status is determined by the auxiliary classification results obtained from the comprehensive score calculation in step S1; , , , For resources The set of visible roles, the set of visible stages, the valid start date, and the valid end date; t is the current timestamp. Only resources that simultaneously meet the three conditions of being visible in roles, visible in stages, and valid in time can pass the hard filter and enter the candidate set.

[0130] Building upon hard filtering, the system further performs soft filtering. Soft filtering expresses business rules as a set of configurable constraints. Perform constraint consistency screening on the candidate set after hard filtering:

[0131] ;

[0132] In the formula, This is the candidate set after soft filtering; It is a strongly constrained set; Indicates the resource and users Enforcement constraints The verification function, when resources In users Under the condition of satisfying constraints Returns True if true, otherwise returns False.

[0133] The constraint sets include, but are not limited to: capacity constraints (to determine if the remaining slots for a resource are greater than zero), time window constraints (to determine if the current time is within an available booking period), conflict constraints (to determine if the resource conflicts with the user's existing course selection or booking records), and prerequisite constraints (to determine if the user meets the resource's required completion rate or completion threshold). The constraint sets are managed through a configuration-based approach and are not fixed, hard-coded logic. Configuration can be performed by academic affairs or administrative personnel, resource maintenance personnel, or system administrators. Different roles maintain different categories of rules, including teaching rules, resource rules, and system execution rules. The system maintains the constraint sets through a rule configuration table. Each rule includes at least the following fields: rule type, rule expression, priority, start / stop status, and scope of effect.

[0134] When conflicts arise between constraints, the system prioritizes stronger constraints over weaker ones. Permissions, time windows, capacity, and conflicts are considered strong constraints, primarily controlling whether resources can be executed; preferences and popularity are considered weak constraints, only affecting sorting weights. When conflicts occur between similar strong constraints, they are then sorted according to the priority field.

[0135] Through a two-stage filtering mechanism of hard filtering and soft filtering, the system not only ensures the correctness of resource visibility, but also guarantees the executability of resources at the current moment, effectively improving the input quality of the subsequent semantic matching and sorting stage.

[0136] Step S3: Receive the user's query statement, sort the candidate resource set by the semantic similarity between the query statement and the candidate resource set and the BM25 retrieval score, and obtain the retrieval context fragments corresponding to the top N candidate resources.

[0137] This step receives the user's query and performs semantic similarity calculation between the candidate resource set obtained in step S2 and the query, obtaining the retrieval context fragments corresponding to the top N candidate resources with the highest scores. This calculation process employs a hybrid retrieval strategy combining semantic vector retrieval and BM25 keyword retrieval, and overlays user state matching correction terms to obtain more accurate ranking results, such as... Figure 2 As shown, it specifically includes the following four sub-steps.

[0138] Step S31: Process the query statement entered by the user. Constructing semantic embedding vectors The query can originate from the user's natural language question, task description, or a system-generated query template. Simultaneously, the candidate resource set... Each resource Constructing resource semantic embedding vectors This vector is generated by concatenating the resource's title, summary, text fragments, and tag fields using an encoding model. The cosine similarity between the query and each resource is calculated based on the semantic embedding vector.

[0139] ;

[0140] Since the cosine similarity score ranges from [−1, 1], it undergoes a translation and normalization process to facilitate fusion calculations with other scores, resulting in a semantic score:

[0141] .

[0142] In the formula, For semantic score; For query statement With candidate resources Cosine similarity between them; , These are the minimum and maximum values ​​of the cosine similarity in the current candidate resource set, respectively. To prevent small constants from being divided by zero.

[0143] After translation and normalization, The value range is [0,1].

[0144] Step S32: For the query statement With each resource in the candidate resource set Calculate the BM25 keyword retrieval score. The BM25 algorithm measures the degree of matching between keywords in the query and the resource text based on factors such as term frequency, inverse document frequency, and document length normalization. Since the original BM25 score has a variable and potentially large range, the system performs saturation normalization on it.

[0145] ;

[0146] In the formula, The BM25 keyword retrieval score after saturation normalization; For query statement With candidate resources BM25 keyword search scores between; For smoothing constants, a recommended value range is [1.0, 3.0], with the default value being... =1.5. After this transformation The value range is [0,1), and the normalization result is also 0 when the original score of BM25 is 0, thus maintaining the monotonicity of the score.

[0147] Step S33: The semantic score and the BM25 keyword retrieval score after saturation normalization are weighted and fused according to the weight parameters to obtain the hybrid retrieval score:

[0148] ;

[0149] In the formula, For mixed search scores; For configurable weight parameters related to resource type, the default value is... The focus is on semantic retrieval. The adaptive mechanism here is essentially a rule-based mechanism that selects different parameter configurations according to resource type. For example, for resources with stronger keyword features such as rules and regulations and process descriptions, the semantic weight can be appropriately reduced to around 0.5, while for resources such as course descriptions, project descriptions, and FAQs, the semantic weight can be appropriately increased.

[0150] Step S34: Add a user state matching correction term to the mixed search score to obtain the final ranking score:

[0151] ;

[0152] In the formula, The final ranking score; To correct the error, a default value of 0.2 is recommended, with a range of [0.1, 0.5]. The weighted matching degree between the user state feature vector and the resource adaptation condition vector.

[0153] Weighted matching degree The calculation method is as follows. Let the resources... The adaptation condition vector is Each dimension corresponds one-to-one with the user's state feature vector. ∈[0,1] represents the lower limit of skill completion for resource recommendations. ∈{0,1,2} represents the project participation status of resource requirements. ∈[0,1] represents the lower limit of the completion rate of innovative lessons based on resource suggestions. ∈{1,2,3} represents the minimum permission level required for the resource. ∈[0,1] represents the lower limit of the credit achievement rate for resource recommendations.

[0154] For continuous dimensions The matching score calculation rules are as follows: when the normalized value of the user's status reaches or exceeds the lower limit of the resource suggestion, the matching score is 1; when the user's status is below the lower limit of the resource suggestion and the lower limit is greater than 0, the matching score is proportionally reduced; when there are no prerequisite requirements for the resource, the matching score is 1. The specific formula is:

[0155] ;

[0156] For discrete dimensions When the user's status meets the resource requirements ,otherwise .

[0157] Since the permission level dimension is already used as an independent filtering condition, the weighted matching degree calculation excludes the permission level dimension. The expression for the weighted matching degree is:

[0158] ;

[0159] In the formula, This is the weight vector corresponding to the user state feature vector.

[0160] The value range is [0,1], and the closer the value is to 1, the better the user's status matches the resource suitability conditions. This correction item does not replace the original search score, but is used to improve the ranking position of resources that are more closely matched to the current progress.

[0161] Step S35: Press The candidate set after soft filtering is sorted, and the top N candidate resources with the highest sorting scores are selected to form the Top-N set. The value of N is determined by a combination of a base value, a candidate set size adjustment, and upper and lower bound control.

[0162] ;

[0163] in The recommended default value is 5; As a lower bound, we take 3 to ensure that at least 3 candidate resources are returned; The upper limit is recommended to be 15, in order to avoid prompts being too long and exceeding the context window limit of the large language model.

[0164] exist Based on this, the system extracts the most relevant fragments to the query statement from the hit resource blocks as the retrieval context, i.e., the retrieval evidence set. The extraction of evidence fragments employs a combination of fixed-length block segmentation and overlapping sliding windows, with a single fragment length capped at 512 tokens to control the total length of the prompts. The specific values ​​of these parameters can be adjusted based on the context window capacity of the large language model. This retrieval evidence set will serve as the core input for constructing enhanced prompts in subsequent step S4.

[0165] Step S4: Inject the user state feature vector, query statement, and retrieval context fragment into the prompt word template to drive the large language model to generate structured output containing intent identifiers.

[0166] This step injects the user state feature vector constructed in step S1, the user's query statement, and the retrieval context fragment obtained in step S3 into the prompt word template, driving the large language model to generate structured output containing intent identifiers, intent confidence, and business operation parameters.

[0167] The system will use the user state vector User Inquiry 1. Retrieve evidence set and business rules constraints The prompts are uniformly injected into the prompt template to form controlled generated input. Each prompt must contain at least four parts:

[0168] The first part contains user training stage and permission information. The system automatically generates a status summary text based on the user's structured fields and a template, namely:

[0169] ;

[0170] in This represents a templated generation function; This is the set of status fields already acquired by the system. This summary converts structured fields such as user roles and permissions, progress status, learning completion status, project participation, lab participation, and credit percentage into a unified text summary to constrain the output range. The status summary is a template mapping result from structured fields to prompt text; it does not add new business fields or rely on model inference.

[0171] The second part is the retrieved evidence fragments, namely the retrieved evidence set obtained in step S3. The content of each segment and its numbering are used to reduce the illusion phenomenon of large language models.

[0172] The third part is the output format constraint, which requires the large language model to strictly follow the predefined JSON structure to return the output results, ensuring that the output is parsable, executable, and auditable.

[0173] Part Four is a statement of prohibited items, which explicitly requires that large language models must not output unauthorized content or construct non-existent resources or time periods, in order to reduce the risks of illusion and unauthorized access.

[0174] In the retrieval and filtering module, the system introduces a data anonymization proxy layer. This is used in constructing semantic embedding vectors. Previously, the system identified and masked user privacy information, such as student ID, real name, and specific scores. Now, it uses desensitized feature labels for vector space retrieval to ensure a more comprehensive evidence set. It does not contain any raw sensitive data and meets the compliance requirements for educational data.

[0175] The output generated by the large language model adopts a fixed JSON structure format, which includes at least the following fields: the answer field is the interpretive natural language answer; the recommendations field is the recommended resource list, each item containing Resource_ID, title, recommendation reason and associated action type; the next_actions field is the list of executable actions, each item containing ActionType, ActionEndpoint, business operation parameters params and a flag indicating whether secondary confirmation is required_confirm; the state_update_hint field is the suggestion for state write-back, indicating the affected state dimension and incremental value; the evidence field is the identifier of the retrieved evidence fragment; the Intent field is the intent identifier, indicating the type of user intent identified by the system; and the Confidence field is the intent confidence, with a value range of [0,1].

[0176] By constraining the output format of the aforementioned fixed fields, the system transforms the output of the large language model from natural language responses into parsable, executable, and auditable intermediate results, providing stable structured input for subsequent intent recognition, parameter validation, and process triggering.

[0177] Step S4 also includes the process of generating, verifying, and handling errors in the structured output, specifically including the following steps.

[0178] First, format validation is performed, which involves parsing the JSON output of the large language model to check if the necessary fields (answer, Intent, Confidence, next_actions) exist and are of the correct type.

[0179] Secondly, value range validation is performed, which checks whether the value of the Confidence field is within the range of [0,1], whether the value of the Intent field belongs to the system's defined intent set, and whether the Resource_ID referenced in next_actions exists in the candidate resource set.

[0180] When format validation or value range validation fails, the system appends a format correction instruction to the prompt word template and re-invokes the large language model, retrying a maximum of a preset number of times. If the retry still fails, the system only returns a downgraded response containing a text explanation, that is, only retains the content of the answer field, does not trigger any business process, and prompts the user "The operation cannot be performed automatically at present. Please operate manually or try again later," without proceeding to step S5.

[0181] In addition, the system includes a confidence calibration mechanism. When there is a significant deviation between the historical execution success rate of a certain intent type and the average confidence output by the model, the system performs linear calibration on the confidence.

[0182] ;

[0183] Here, 'a' and 'b' are obtained by fitting historical data. This calibration mechanism allows the confidence score to more accurately reflect the probability of actual successful execution.

[0184] Step S5: Call the business interface to execute the business operation corresponding to the intent identifier.

[0185] This step performs a threshold determination on the intent confidence in the structured output generated in step S4, verifies the legality of the business operation parameters, and after all passes, calls the business interface of the training management platform to execute the business operation corresponding to the intent identifier. After successful execution, the user state feature vector is written back and updated to form a complete closed loop.

[0186] The system determines the trigger conditions for Intent and Confidence in the structured output. The trigger conditions are defined as follows:

[0187] ;

[0188] In the formula, For the confidence level of the intention; Configurable trigger thresholds for different intent types are set in tiers according to risk level: query intents have lower risk, so the threshold is set to [value missing]. Recommendation-type intents require a certain degree of stability; the threshold is set to... Actions such as booking appointments, selecting courses, and submitting applications directly impact real business data; threshold values ​​are set accordingly. . This indicates the results of parameter completeness and validity verification. The above thresholds are initial, feasible configuration values ​​given based on risk stratification, and can be gradually adjusted based on operational data.

[0189] Once the intent confidence level passes the threshold, the system performs multiple checks on the business operation parameters according to a fixed priority: The first step is permission check, which checks whether the user's current role and organizational domain permissions allow them to execute the action; the second step is time window check, which checks whether the time information in the operation parameters is within the executable time period; the third step is capacity and quota check, which checks whether the remaining quota of the target resource is greater than zero; the fourth step is conflict check, which checks whether the target operation conflicts with the user's existing appointments or course schedules; the fifth step is approval status check, which checks whether the preconditions for approval are met if the operation requires approval; and the sixth step is idempotency check, which checks whether the same action has already been executed to avoid duplicate submissions.

[0190] If any verification step fails, the system does not perform a write operation but instead enters a failure degradation strategy: it prioritizes returning a description of the failure reason and actionable alternative suggestions, such as alternative time periods, alternative resources, or parameter completion hints. Alternative solutions are primarily derived from rule-based business queries and candidate set re-filtering rather than being freely generated by a large language model. Specifically, when time window, capacity, or conflict verification fails, the system prioritizes re-querying for executable time periods in the available time schedule corresponding to the current resource; when the current resource is entirely unexecutable, the system excludes the failed resource from the candidate set and selects the next best resource as a replacement.

[0191] ;

[0192] ;

[0193] In the formula, For failure resources; As an alternative resource; This represents the set difference operation.

[0194] Only after alternative resources or time periods have been determined through rule retrieval and business queries does the system invoke the generated model to provide explanatory details. Therefore, the alternative solutions themselves originate from verifiable business data rather than model-generated content. The system also records the reasons for verification failures in the audit log for traceability.

[0195] When the intent identifier belongs to the write operation category, i.e., the intent type that adds, modifies, or deletes data on the training management platform, such as appointment, course selection, cancellation, and application submission, the system generates a one-time confirmation token and requires the user to confirm again before executing the business operation. This one-time confirmation token is bound to the user identifier, intent identifier, a digest of the business operation parameters, and an expiration date. Its structure includes `confirm_token`, which is generated by BASE64 encoding of the user identifier, intent identifier, parameters, and timestamp using a hash-based message authentication code (HMAC); the token's expiration time `expires_at`; and `action_digest`, which is an operation digest generated by SHA256 hashing of the user identifier, intent identifier, and parameters. The system only calls the backend business interface to execute the operation after user confirmation to avoid accidental triggering and unauthorized execution.

[0196] For write operations, after all validations pass and the user confirms, the system calls the business interface of the cultivation management platform to execute the business operation corresponding to the intent identifier. To ensure the auditability and rollbackability of the interface call, the system generates a unique execution number exec_id for each business operation and records the input parameters, validation results, interface response, and write-back results.

[0197] The system records the execution chain of business operations as sequentially executed sub-steps. When the execution link is in the first If a step fails, perform compensation operations on the successful preceding sub-steps in reverse order:

[0198] ;

[0199] In the formula, For sub-steps Compensation actions include releasing occupied slots, deleting temporary records, or restoring the previous state. Compensation is triggered by conditions such as API call failure, status write-back failure, log write failure, call timeout, and abnormal interruption during the execution chain. For timeout scenarios, the system sets a uniform timeout threshold. After a timeout, the current execution is marked as pending, and a limited number of retries are performed. If the retries still fail, the compensation process begins and the compensation result is recorded.

[0200] After a business operation is successfully executed, the system triggers a write-back update of the user's state feature vector. The state vector update process is as follows: Figure 3 As shown, record the time window The set of events generated by internal business operations is The update function for the state vector is defined as follows: .

[0201] For continuous dimensions, the system uses an incremental update and pruning approach:

[0202] ;

[0203] In the formula, and They are respectively Time and Time of the first The values ​​of each continuous dimension; For time windows A set of events generated by internal business operations; For the event For the first Incremental contribution in each dimension; This means cropping the result to the [0,1] interval to ensure that the value is reasonable.

[0204] The specific mapping relationship between events and dimension increments is as follows: When completing a learning task, the skill learning completion rate... Increase ,in This represents the total number of tasks for the course, with all other dimensions remaining unchanged; when joining a project, the project participation status... The migration from 0 to 1 remains unchanged, with all other dimensions remaining the same; the project participation status changes upon successful project completion. Transfer from 1 to 2, while simultaneously achieving the required credit rate Increase the percentage of credits corresponding to this project When completing the lab session sign-in, the completion rate of the innovation session is... Increase , The total number of lab sessions remains the same, while other dimensions remain unchanged; the percentage of students who meet the credit requirements is the percentage of students who meet the credit requirements when the grade assessment or credit entry event is triggered. The system retrieves actual values ​​from the academic affairs system and directly overwrites them. When multiple events occur simultaneously, the system prioritizes merging and then updating incremental values ​​for continuous dimensions; if there is a clear sequential dependency between events, they are written sequentially according to their timestamps.

[0205] For discrete dimensions, the system does not accumulate values ​​but updates them using a state machine transition method. For example, the transition path for project participation status is 0 (not participating) → 1 (participating) → 2 (project closed), and the transition is irreversible; changes in permission levels are triggered by permission change events in the management system.

[0206] After the write-back update is completed, the system will synchronously recalculate the overall score. And update the cultivation progress status. If the updated overall score crosses the threshold range, a smooth transition update of the weight vector is triggered. The system also generates a state change log for each write-back operation, recording the state values ​​before and after the change, the triggering event, and the timestamp, for auditing and backtracking purposes.

[0207] During operation, the system adaptively adjusts trigger thresholds by analyzing historical execution data for each intent type. When a certain type of action consistently experiences complaints or false triggers, the system appropriately increases its trigger threshold to enhance security. The threshold update formula is:

[0208] ;

[0209] ;

[0210] In the formula, For small step size update coefficients, a value of 0.005 is recommended, with a range of [0.001, 0.02]. Threshold adjustment needs to be more conservative than weight update. The false trigger rate for this type of intent over the most recent statistical period is calculated by taking into account both the appeal rate and the execution failure rate. To set the upper limit for the target false trigger rate, recommended values ​​are 0.05 for query operations, 0.03 for recommendation operations, and 0.01 for write operations. Greater than hour, When the threshold is greater than 0, the threshold becomes more stringent; when Less than hour, If the threshold is less than 0, the threshold should be lowered and relaxed appropriately.

[0211] The threshold constraints for each intent type are as follows: query threshold range [0.4, 0.8], recommended initial value 0.6; recommendation threshold range [0.5, 0.9], recommended initial value 0.7; write operation threshold range [0.7, 0.95], recommended initial value 0.85. This adaptive adjustment mechanism allows the system to gradually optimize trigger sensitivity based on actual operating conditions while ensuring security.

[0212] The following complete embodiment illustrates the collaborative working process of each step of the present invention.

[0213] Suppose a graduate student user Enter the following query into the system: "What is my next step to achieve the target? Can you recommend the lab classes I should book?"

[0214] In step S1, the system reads the user's culture process data from the culture management platform and constructs a user status feature vector. Assuming the user's skill learning completion rate... Project participation status (Participation in progress), completion rate of innovation course Permission levels (Students), Credit Completion Rate After normalization mapping, a normalized vector is obtained, which adopts the weight template of the process advancement stage. Calculate the overall score:

[0215] =0.20×0.6+0.35×0.5+0.25×0.25+0.20×0.55=0.4675.

[0216] The overall score falls within the range of [0.3, 0.7), therefore the cultivation progress status is determined to be... (Process Progress). The system detected the user's completion rate of the innovative lesson. If the value is significantly below the threshold, this information will be used for resource matching and path guidance in subsequent steps.

[0217] In step S2, the system is based on permission levels. (Student role) and progress of training (Process Advancement) Perform hard filtering on the knowledge resource set to remove resources that are not visible to student roles, are not suitable for the process advancement stage, or are outdated, resulting in a candidate set after hard filtering. Then, soft filtering is performed to further exclude resources that are full, have closed time windows, or conflict with the user's existing course schedule, resulting in the final set of candidate resources. .

[0218] In step S3, the system performs semantic embedding encoding on the query "What's the next step for me to achieve the target? Can you recommend the lab classes I should book?", and... The system calculates the cosine similarity of the semantic vectors of each resource and performs translational normalization. Simultaneously, it calculates the BM25 keyword retrieval score and performs saturation normalization, then fuses them according to weights to obtain a hybrid retrieval score. The system further calculates the weighted matching degree between the user's state and the adaptation condition vectors of each resource. Since the user's completion rate of the innovation course is low, the matching degree correction term with innovation experiment course resources will significantly improve the ranking position of such resources. The system selects the Top-N resources with the highest final ranking scores and their corresponding retrieval evidence fragments to constitute the retrieval context.

[0219] In step S4, the system injects a user status summary (containing information such as the user's role as a student, progress status as "progressing," and low completion rate of the innovation course), the query statement, and retrieval evidence fragments from Top-N resources into a prompt word template, driving the large language model to generate structured JSON output. In the model output, the Intent field is "Reservation," the Confidence field is 0.92, next_actions contains the operation parameters for reserving an innovation experiment course, such as the course identifier and recommended time slot, and state_update_hint indicates the completion rate of the innovation course after successful reservation. The corresponding increment will be added after check-in. The system performs JSON format validation and value range validation on the output to confirm that all field types are correct and values ​​are valid.

[0220] In step S5, the system first determines that the confidence level of 0.92 is greater than the write operation threshold of 0.85, thus passing the threshold check. Subsequently, it sequentially performs permission checks: confirming the student role has the permission to book the lab class; time window check; confirming the recommended time slot is within the bookable range; capacity check; confirming there are remaining slots for the target lab class; conflict check; and confirming that the recommended time slot does not conflict with the user's existing schedule. All checks pass. Since this operation is a write operation intent, the system generates a one-time confirmation token and displays it to the user. After the user confirms, the system calls the booking interface to complete the lab class booking. After successful booking, the system triggers a status write-back, recording the booking event but not updating it immediately. Users need to sign in and update their status, and the audit log will be updated synchronously. The system displays the path nodes and jump entries to the user in the visual path: appointment → sign in → completion → achievement, forming a complete closed loop from asking questions to action execution and status update.

[0221] The above steps constitute the complete implementation of the intelligent assistance method for postgraduate project-based training scenarios of this invention. This method achieves end-to-end intelligent assistance from user inquiry to knowledge retrieval, from intent recognition to process triggering, and from operation execution to state update, through structured construction and dynamic updating of user state feature vectors, two-level resource filtering based on permissions and training stages, mixed semantic and keyword retrieval and sorting, controlled enhanced prompt word-driven structured output generation, business operation execution under multiple verification and security confirmation mechanisms, and event-driven state write-back closed loop. While ensuring access control security and execution reliability, it significantly improves resource matching accuracy and operation execution efficiency in postgraduate training management scenarios.

[0222] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; only preferred embodiments of the present invention are illustrated. The descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. As long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0223] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the appended claims.

Claims

1. An intelligent assistance method for postgraduate project-based training scenarios, characterized in that, Includes the following steps: Step S1: Collect user's cultivation process data from the cultivation management platform to determine the user's cultivation progress status; Construct a knowledge resource set, and label the visible stage of each resource in the knowledge resource set based on the correspondence between the resource content characteristics and the cultivation progress status, so that each resource in the knowledge resource set is associated with at least one cultivation progress status. Step S2: Based on the user's permission level and the training progress status, perform candidate resource filtering on the knowledge resource set to obtain a candidate resource set; Step S3: Receive the user's query statement, sort the candidate resource set by the semantic similarity between the query statement and the candidate resource set and the BM25 retrieval score, and obtain the retrieval context fragments corresponding to the top N candidate resources; Step S4: Inject the query statement and retrieval context fragment into the prompt word template to drive the large language model to generate structured output containing intent identifiers; Step S5: Call the business interface to execute the business operation corresponding to the intent identifier.

2. The intelligent assistance method for postgraduate project-based training scenarios according to claim 1, characterized in that: Step S1, which involves collecting user cultivation process data from the cultivation management platform to determine the user's cultivation progress status, includes the following steps: The training process data of users is collected from the training management platform to construct a user status feature vector, which includes skill learning completion rate, project participation status, experimental course completion rate, permission level and credit achievement rate. The training progress status of users is determined based on the user status feature vector. The skill learning completion rate, lab course completion rate, and credit achievement rate are within a certain range. The project participation status is a continuous dimension, while the project participation status is a discrete dimension representing three states: not participating, participating, and completed. The permission level is a discrete dimension representing three user roles: student, tutor, and administrator. After normalizing the discrete dimension, it is combined with the continuous dimension to form a normalized vector. A weight vector is introduced to sum the weighted values ​​of all dimensions in the normalized vector except for the permission level, to obtain a comprehensive score. Based on the comparison between the comprehensive score and a preset threshold, the cultivation progress status is divided into three discrete states: ； In the formula, For users The progress of cultivation; and These are the lower threshold and the upper threshold, respectively. For users Overall score; When the overall score crosses a preset threshold, triggering a state switch, the weight vector is updated using a smooth transition: ； in, This is the transition weight vector; This is the current weight vector; The target weight vector; This is the smoothing coefficient.

3. The intelligent assistance method for postgraduate project-based training scenarios according to claim 2, characterized in that: The construction of the knowledge resource set in step S1 includes the following steps: Step S11: Collect and segment the system documents, course descriptions, experimental project descriptions, process guidelines, and frequently asked questions of the training management platform to obtain a resource segment set; Step S12: Vectorize and index each resource in the resource block set; Step S13: Label each resource with a set of metadata fields, wherein the set of metadata fields includes at least the resource unique identifier, resource type, set of visible roles, set of visible stages, validity period, and adaptation condition vector; The labeling rules for the visible stage set are as follows: Based on the content characteristics and applicable scenarios of the resources, determine in which training progress stages the resources are visible; resources for basic learning are labeled as visible in the basic preparation stage, resources for project practice and experimental participation are labeled as visible in the process advancement stage, resources for project completion and achievement confirmation are labeled as visible in the task improvement stage, and general resources are labeled as visible in all stages. Each dimension of the adaptation condition vector corresponds one-to-one with the user state feature vector.

4. The intelligent assistance method for postgraduate project-based training scenarios according to claim 2, characterized in that: Step S1 also includes periodically fine-tuning the weight vector based on user feedback, including the following steps: collecting user feedback on the recommendation results and quantifying each feedback instance into a feedback score: ； In the formula, For users The feedback score at time t; For feedback coefficients, ; A binary indicator variable representing user clicks on recommended results; A binary indicator variable representing the user's adoption of the recommendation result; A binary indicator variable representing whether users ignore the recommendation results; A binary indicator variable representing users' negative ratings of recommendation results; A binary indicator variable representing a user's appeal against the recommendation results. coefficient The symbol is determined based on the outcome of the appeal; when the appeal is successful... Take a positive value when the appeal is rejected. Take the negative value; A binary indicator variable representing the successful execution of the business operation corresponding to the recommendation result; A binary indicator variable representing the failure of the business operation corresponding to the recommendation result; The update formula for the weight vector is: ； In the formula, , These are the weights corresponding to the i-th user state feature at time t+1 and time t, respectively; The learning rate; This is the normalized value of the i-th user state feature in this recommendation.

5. The intelligent assistance method for postgraduate project-based training scenarios according to claim 1, characterized in that: The candidate resource filtering in step S2 includes hard filtering and soft filtering performed sequentially. The hard filtering filters the knowledge resource set based on user role, training progress status, and current timestamp. ； In the formula, This is the candidate set after hard filtering; For users The level of access control; For users The progress of cultivation; , , , For resources The set of visible roles, the set of visible stages, the valid start date, and the valid end date; t is the current timestamp; The soft filter expresses business rules as a set of configurable constraints. Constraint consistency screening is performed on the hard-filtered candidate set: ； In the formula, This is the candidate set after soft filtering; For constraint sets; Indicates the resource and users Enforcement constraints The verification function.

6. The intelligent assistance method for postgraduate project-based training scenarios according to claim 2, characterized in that: Step S3 involves ranking the candidate resource set based on the semantic similarity between the query statement and the candidate resource set, and using the BM25 retrieval score. This includes the following steps: Step S31: Construct semantic embedding vectors for the query statement and each resource in the candidate resource set. Calculate the cosine similarity between the query statement and each resource based on the semantic embedding vectors. Perform translational normalization on the cosine similarity to obtain a semantic score. The expression for the semantic score is: ； ； In the formula, For semantic score; For query statement With candidate resources Cosine similarity between them; , These are the minimum and maximum values ​​of the cosine similarity in the current candidate resource set, respectively. To prevent division by zero of small constants; This is a query statement; Candidate resources; This is the semantic embedding vector of the query statement; This represents the semantic embedding vector of the candidate resource; Step S32: Calculate the BM25 keyword retrieval score for each resource in the query statement and the candidate resource set, and perform saturation normalization on the BM25 keyword retrieval score: ； In the formula, The BM25 keyword retrieval score after saturation normalization; For query statement With candidate resources BM25 keyword search scores between; It is a smoothing constant; Step S33: The semantic score and the BM25 keyword retrieval score after saturation normalization are weighted and fused according to the weight parameters to obtain the hybrid retrieval score: ； In the formula, For mixed search scores; Configurable weight parameters related to resource type; Step S34: The final ranking score is obtained by adding a user state matching correction term to the mixed retrieval score. ； In the formula, The final ranking score; This is a correction factor; The weighted matching degree between the user state feature vector and the resource adaptation condition vector.

7. The intelligent assistance method for postgraduate project-based training scenarios according to claim 6, characterized in that: The weighted matching degree The calculation method is as follows: Resource The adaptation condition vector is Each dimension corresponds one-to-one with the user state feature vector; For continuous dimensions : ； In the formula, For the first Normalized values ​​of individual user state features; For discrete dimensions When the user's status meets the resource requirements ,otherwise ; Since the permission level dimension is already used as an independent filtering condition, the weighted matching degree calculation excludes the permission level dimension. The expression for the weighted matching degree is: ； In the formula, This is the weight vector corresponding to the user state feature vector.

8. The intelligent assistance method for postgraduate project-based training scenarios according to claim 2, characterized in that: The prompt word template mentioned in step S4 includes at least: User development stage and permission information: The status summary text is automatically generated by concatenating the structured fields of the user status feature vector according to the template; Retrieve context fragments and number identifiers; Output format constraints: The large language model is required to return the output results according to a predefined JSON structure, wherein the JSON structure shall at least include an explanatory answer field, a recommended resource list field, an executable action list field, a state write-back suggestion field, an evidence citation field, the intent identifier field, and an intent confidence field; Prohibited Items Declaration: The large language model is required not to output unauthorized content and not to construct non-existent resources or time periods.

9. The intelligent assistance method for postgraduate project-based training scenarios according to claim 8, characterized in that: Step S4 also includes generating, verifying, and performing fault-tolerance processing on the structured output: The output of the large language model is subjected to format validation and value range validation. The format validation includes checking whether the necessary fields exist and are of the correct type. The value range validation includes checking whether the intent confidence is within the range of [0,1], whether the intent identifier belongs to the intent set defined by the system, and whether the resource identifier referenced in the list of executable actions exists in the candidate resource set. When the verification fails, a format correction instruction is appended to the prompt word template and the large language model is called again, with a maximum of a preset number of retries. If the retry still fails, a downgraded response containing only explanatory answers will be returned, without triggering any business operations.

10. The intelligent assistance method for postgraduate project-based training scenarios according to claim 2, characterized in that: Step S5 includes the following steps: Step S51: Compare and determine the intent confidence in the structured output with the configurable trigger threshold corresponding to the intent identifier. Set trigger thresholds for different intent types according to risk level. When the intent confidence does not reach the trigger threshold, it is downgraded to the only recommendation mode. Step S52: After the intent confidence level passes the threshold, perform the following checks on the business operation parameters in sequence: permission check, time window check, capacity and quota check, conflict check, approval status check and idempotency check. If any check fails, return the reason for failure and alternative suggestions. Step S53: When the intent identifier belongs to the write operation type intent, generate a one-time confirmation token that binds the user identifier, intent identifier and business operation parameter summary, and call the business interface to execute the business operation after the user confirms; Step S54: After the business operation is successfully executed, the user state feature vector is written back and updated. The continuous dimension is updated incrementally and pruned to the [0,1] interval, while the discrete dimension is transitioned using a state machine. The comprehensive score and training progress status are recalculated simultaneously.

Images