An intelligent credit report interpretation and question answering system based on a large model agent

Abstract

The application relates to the field of semantic processing, and discloses an intelligent credit investigation report interpretation and question answering system based on a large model Agent, which comprises the following steps: reconstructing a credit event chain interwoven with account dimensions and time axis dimensions, and generating a focus index sequence according to the influence degree of entries and the user browsing track; splitting the corresponding credit event chain into explainable units to form an explanation fragment set with risk pointing marks; constructing a multi-round dialogue context cache area on the basis of the explanation fragment set, generating a dynamic explanation framework for a single credit record through an intention recognition and context completion mechanism; when it is identified that the user has doubts about the credit granting result or score change, generating a traceable influence link description after integrating the output results of various subtasks; and outputting an interactive interpretation result of the influence link description under the premise of maintaining data compliance boundaries. The application has the advantages of improving the report understanding.

Description

Technical Field

[0001] This invention relates to the field of semantic processing, specifically to an intelligent credit report interpretation and question-answering system based on a large-scale agent model. Background Technology

[0002] With the continuous expansion of the business scale of commercial banks, consumer finance institutions, and internet lending platforms, personal credit reports have become a fundamental data carrier in credit approval, risk assessment, and post-loan management. Existing credit reporting services typically present credit information to users in the form of structured text or standardized documents, including credit records, account status, overdue information, inquiry records, and public information. While the collection, modeling, and risk scoring capabilities of credit data have achieved a high degree of automation in the context of continuously upgrading intelligent risk control systems, the report presentation format for end users remains primarily static.

[0003] In practical application, it has been found that although existing credit reports are complete, when users view them on mobile devices or during online loan application processes, they often cannot immediately understand the cause, impact, and specific relationship to the current credit granting result for individual abnormal information. Existing systems typically only provide fixed-format explanations or frequently asked questions, failing to provide contextual explanations for specific items of interest to the user or to offer personalized explanations based on their overall credit status. This leads to users needing to contact human customer service or offline consultation when they have questions, increasing service costs and reducing user experience.

[0004] Therefore, it is necessary to design an intelligent credit report interpretation and question-answering system based on a large model agent to improve the comprehensibility of the report. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent credit report interpretation and question-answering system based on a large-scale agent model, which has the advantage of improving report comprehensibility and solves the problems mentioned in the background technology.

[0006] To achieve the aforementioned goal of improving report comprehensibility, this invention provides the following technical solution: an intelligent credit report interpretation and question-answering system based on a large-scale agent model, comprising: Message Chaining Module: Accesses original personal credit reports and formatted display files, parses credit account status fields, query source identifiers and repayment performance time series, reconstructs a credit event chain interwoven with account and time axis dimensions, and generates a focus index sequence based on the impact of items and user browsing history; Semantic Slicing Module: A semantic slicing mechanism is introduced around the focus index sequence to split the corresponding credit event chain into interpretable units. Combined with the fluctuation range of account quota utilization, the distribution of overdue days, and query density segments, a set of interpretable fragments with risk-indicating markers is formed. Conversation Management Module: Based on the set of explanatory fragments, a multi-turn dialogue context cache is built to record the user's follow-up question path, terminology comprehension deviations and reading pause nodes. Through intent recognition and context completion mechanisms, a dynamic explanatory framework for a single credit record is generated, and the consistency between the explanatory content and the original data fields is verified simultaneously. Source tracing and reasoning module: When it is detected that a user has doubts about the credit granting results or changes in the rating, the large model Agent triggers the decision source tracing task planning instruction, which is broken down into three sub-tasks: rating interval retrieval, rule hit record retrieval, and item weight comparison. Data collection is completed through internal interface calls. After integrating the output results of each sub-task, a traceable impact chain description is generated. Interactive Generation Module: Combining users' historical consultation behavior with the current conversation depth, the module adjusts the granularity and level of expression of responses, and outputs interactive interpretation results of the impact chain description while maintaining data compliance boundaries.

[0007] Preferably, the process of reconstructing the credit event chain that intertwines the account dimension and the time axis dimension is as follows: The interface adaptation component reads the message data from the credit reporting center and the formatted file displayed to users, and performs unified and standardized conversion of field encoding, data separators and time formats. Extract account number, account type, credit limit, current balance, repayment status, last repayment date, and query institution code to build an account primary key index; Based on the repayment and performance timeline, a monthly behavior record table is constructed for each account, mapping overdue identification, extension records, and account status change information to the corresponding time nodes; By using a cross-account time alignment mechanism, the behavior records of multiple accounts are spliced ​​along a unified timeline to form a credit event chain with a dual index structure of account dimension and time dimension.

[0008] Preferably, the process of generating a focus index sequence based on the influence of items and the user's browsing history is as follows: Based on the credit event chain, an item impact assessment function is constructed. The assessment function comprehensively considers the historical contribution weight of delinquency level, balance occupancy ratio and query frequency to credit granting decision. Collect user swipe duration, click-to-expand behavior, and highlighted actions in the report interface to form a browsing behavior trajectory vector; The item's influence is weighted and coupled with the browsing behavior trajectory vector to obtain the item's attention score. Sort the items by their attention score from highest to lowest, and generate an index sequence of attention points that reflects the current user concerns.

[0009] Preferably, the process of breaking down the corresponding credit event chain into interpretable units is as follows: Based on the focus index sequence, the corresponding credit event subchain is located, and the basic account information, time period and status change records associated with the credit event subchain are extracted. The standardized fields in the credit report are uniformly encoded, and terminology mapping rules for credit reporting business scenarios are constructed to map each field to the corresponding semantic label. A semantic slicing rule set is established based on term mapping rules. Context window scanning is performed on the credit event subchain to merge event records with consistent state features within consecutive time segments into the same semantic segment. Identify the time points when the state changes and divide the event records corresponding to the time points into independent boundary segments; The resulting semantic fragments and independent boundary fragments are structurally encapsulated to generate interpretable units containing semantic tags, time intervals, and account identification information.

[0010] Preferably, the process of forming a set of explanatory fragments with risk-indicating markers is as follows: Read the total credit limit and actual balance of the account corresponding to the interpretable unit within a continuous billing cycle, calculate the ratio sequence of the balance to the total credit limit, and extract the start and end time, the highest occupancy ratio, and the number of cycles required to recover to the normal range for the continuous high occupancy interval. Retrieve repayment records and status change indicators within the time period, calculate the actual number of overdue periods, the longest single overdue days, and the number of consecutive overdue periods, and distinguish between occasional single-period overdue and consecutive overdue situations across billing cycles. Synchronously retrieve credit inquiry records, count the number of loan approval inquiries, credit card approval inquiries, and the number of different financial institutions within a time window, and calculate the time interval between two adjacent inquiries; Risk labels are generated based on the peak balance occupancy ratio, consecutive overdue periods, longest overdue days, and number of approval-related queries according to the risk level corresponding to the numerical range. These labels are then attached to the corresponding semantic fragments to form a set of explanatory fragments with risk-indicating markers.

[0011] Preferably, the process of recording user follow-up questions, misunderstandings of terminology, and reading pause points is as follows: Assign a unique session identifier to each user question and record the set number of the explanation fragments corresponding to the question; Keyword extraction is performed on user input statements, and they are matched with risk-indicating markers and semantic tags to mark the hit relationship; The system tracks the time users spend on the explanation page and the number of repeated clicks to identify potential misunderstandings. The system then stores the question history, hit relationships, and misunderstanding results in the context cache.

[0012] Preferably, the process of synchronously verifying the consistency between the description content and the original data fields is as follows: The large model Agent is used to perform semantic parsing on the current session state data structure to identify the user's question intent category; By combining the semantic fragments and risk tags already recorded in the context cache, contextual completion is performed for missing time intervals or account information; The explanation logic is organized according to the preset explanation template, and a dynamic explanation framework is generated. The amount, date and status fields involved in the explanation content are verified item by item through the field backtracking interface.

[0013] Preferably, the process is broken down into three sub-tasks: score range retrieval, rule hit record retrieval, and item weight comparison. When the user intent category is determined to be the decision interpretation type, a decision tracing task identifier is generated and associated with the credit approval number and scoring timestamp corresponding to the current query; The decision tracing task is divided into a scoring interval retrieval subtask, a rule hit record retrieval subtask, and an item weight comparison subtask. The three subtasks are assigned corresponding data interface addresses and access permission tags, and a sequential execution queue is established based on data dependencies. The sequential execution queue is sent to the execution engine, where data retrieval, computation, and result encapsulation are completed in a predetermined order, before entering the data recycling and interpretation generation stage.

[0014] Preferably, the process of generating a traceable impact chain description after integrating the output results of each subtask is as follows: The scoring system interface is called to obtain the user's current score, the score range, and the upper and lower threshold scores of the corresponding range. The rule log interface is called to read the rule number, trigger time, trigger condition, and associated credit field identifier that were triggered during this credit assessment. Based on the feature importance ranking results returned by the scoring system, retrieve the feature contribution value corresponding to the credit field identifier, and combine it with the actual value of the current user's credit field to calculate the marginal impact score of the credit field on the final scoring result, which serves as the impact weight of the corresponding credit item; For credit entries with rule trigger records, the segmented adjustment values ​​corresponding to the rule hit are superimposed to form an entry-level comprehensive impact weight. The impact chain description is output according to the evaluation time order and the size of the comprehensive impact weight.

[0015] Preferably, the process of outputting interactive interpretation results that affect the description of the data link, while maintaining data compliance boundaries, is as follows: By analyzing the proportion of user attention to three types of content in historical consultation records—terminology explanation, rating mechanism, and improvement suggestions—a user understanding preference vector is constructed. Calculate the conversation depth level based on the current conversation round and the complexity of follow-up questions; select the corresponding explanation template and expansion level based on the understanding preference vector and the conversation depth level; and summarize the description of the influencing links. Before outputting, compliance checks are performed, sensitive information involving policy thresholds is masked, and interactive interpretation results affecting the link description are output.

[0016] Compared with existing technologies, this invention provides an intelligent credit report interpretation and question-answering system based on a large-scale agent model, which has the following advantages: This invention reconstructs the structure of original personal credit reports and formatted display files through a message chaining module, forming a credit event chain interwoven with account and time dimensions. This solves the problem that traditional credit reports only display information by category and fail to reflect causal relationships and temporal evolution logic. Furthermore, a semantic slicing module breaks down continuous credit behavior into interpretable units with semantic tags and risk indicators, enabling structured expression of key information such as fluctuations in account credit limit utilization, changes in overdue duration, and concentrated query frequency, significantly improving the clarity and comprehensibility of the explanation. Finally, a conversation management module constructs multi-turn dialogues. The system includes a text cache area that dynamically records and corrects user inquiries and misunderstandings of terminology, enabling layered and progressive explanations of individual credit records and avoiding inconsistencies or deviations from original fields. A source-tracing reasoning module breaks down the decision-making explanation process into three independently executable subtasks: score range retrieval, rule-hitting record retrieval, and item weight comparison, forming a structured description of the impact chain and enhancing the transparency and traceability of credit granting results and score change explanations. An interactive generation module adjusts the granularity of expression based on conversation depth and historical consultation behavior, outputting differentiated explanations while ensuring data compliance. This improves the accuracy, consistency, and credibility of credit report interpretation, reduces the frequency of human customer service intervention, and enhances users' efficiency in understanding credit results and their risk awareness. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the structure of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1: Please refer to Figure 1 As shown in this embodiment of the invention, an intelligent credit report interpretation and question-answering system based on a large-scale agent model includes: Message Chaining Module: Accesses original personal credit reports and formatted display files, parses credit account status fields, query source identifiers, and repayment performance time series, reconstructs a credit event chain interwoven with account and time axis dimensions, and generates a focus index sequence based on the impact of entries and user browsing history.

[0020] The process of reconstructing the credit event chain interwoven with the account dimension and the time axis dimension in the message chain construction module is as follows: The interface adaptation component reads both the credit reporting center's message data and the formatted files displayed to users, and performs standardized conversions of field encoding, data delimiters, and time formats. The interface adaptation component connects to the credit reporting message interface and the formatted file parsing interface. The credit reporting message data uses message structure parsing rules to locate fields, while the formatted files extract fields using layout label positioning rules or structured extraction templates. A unified character set conversion strategy is used to convert data from different sources into a unified character encoding format. Regular expression matching is used to identify data delimiters, and standard delimiters are recombined. Format detection logic is executed for time fields to determine whether they conform to preset date expression rules. Fields that do not conform to the rules are recombined or marked as abnormal. After conversion, field integrity verification rules are used to check for missing fields or illegal characters. If an abnormality is found, an abnormality log is recorded and a re-parsing process is triggered, thus ensuring that the input data structure is unified and computable. Extract account number, account type, credit limit, current balance, repayment status, most recent repayment date, and query institution code to build an account primary key index; in the standardized field set, locate the core fields of the account according to the preset field mapping table, perform uniqueness detection logic on the account number to determine if there are duplicate numbers, perform numerical range detection on the credit limit and current balance fields to verify that they are non-negative and within a reasonable range, and perform enumeration matching verification on the repayment status field to ensure that it belongs to the preset status set. Use the verified account number as the primary key to build the account index table, and establish an associated index between the account type and the query institution code. If multiple conflicting data from the same account number are detected, data overwriting or merging is performed according to the timestamp priority principle or the data source trust level, thereby forming a consistent account primary index structure; Based on the repayment performance time series, a monthly behavior record table is constructed for each account, mapping overdue identification, extension records, and account status change information to corresponding time nodes. The repayment performance time series field is extracted for each account, sorted according to timestamps, and divided into time windows by natural months. Behavior record entries are generated within each time window. The existence of the overdue identification field is checked and the number of overdue days is recorded. If an extension record exists, the extension type and the start and end time of the extension are marked. Time difference detection is performed on the account status change field. When the status value is detected to be different from the previous time node, it is marked as a status change node. For data of missing months, null value placeholder records are generated according to the completion rules and marked as months with no performance data, forming an account behavior record table containing month index, performance status, number of overdue days, and status change markers. By employing a cross-account time alignment mechanism, multi-account behavior records are concatenated along a unified timeline to form a credit event chain with a dual-index structure encompassing both account and time dimensions. The earliest and latest time nodes of all account behavior records are statistically analyzed to construct a unified timeline sequence. Timeline alignment is then performed on each account behavior record table, and placeholder data is generated for missing time nodes according to null value filling rules. A dual-index structure is established, where the first index is the account number and the second index is the time node. A two-dimensional index matrix stores the behavior status of each account at each time node. Consistency detection logic is executed during the concatenation process to ensure that there are no duplicate account records at the same time node. After concatenation, an integrity check is performed on the overall structure, including time continuity and account coverage checks. If the checks pass, a credit event chain with dual-index structure encompassing both account and time dimensions is generated.

[0021] The process of generating a focus index sequence in the message chaining module based on the impact of entries and the user's browsing trajectory is as follows: Based on the credit event chain, an item impact assessment function is constructed. The assessment function comprehensively considers the historical contribution weights of delinquency level, balance occupancy ratio, and query frequency to credit granting decisions. The delinquency level field, current balance and credit limit field, and query record count field corresponding to each credit event are extracted from the credit event chain. The delinquency level is graded and coded, for example, by mapping levels based on whether delinquency has occurred and the duration of delinquency. The balance occupancy ratio is calculated numerically, i.e., the current balance divided by the credit limit. The query frequency is statistically analyzed based on the cumulative number of queries within a preset time window. Based on historical credit granting sample data, the influence strength of each factor on the credit granting result is pre-calculated to form an influence weight parameter table. The delinquency level value, balance occupancy ratio value, and query frequency value are weighted and summed with their corresponding weights to obtain the item impact score. Simultaneously, a legality detection rule is set. If a field is missing or the value exceeds the preset range, an anomaly flag is triggered and a default weight is used to replace it, thereby ensuring that the impact calculation results are usable and stable. The system collects user data on swiping duration, click-to-expand actions, and highlighting operations in the report interface to form a browsing behavior trajectory vector. A behavior collection component is embedded in the report display interface to listen for page scrolling events, click-to-expand button events, and text highlighting operations. Each behavior record is assigned a unique timestamp and corresponding item identifier. The duration of user stay in each item area is calculated using time difference, and a minimum effective stay threshold is set to filter out invalid rapid swiping behaviors. The number of times click-to-expand actions are recorded, and the number of times highlighting operations are recorded. The duration of stay, number of expansions, and number of highlightings are aggregated and statistically analyzed by item to form a browsing behavior trajectory vector. Simultaneously, anomaly detection logic is executed; for example, repeated clicks within a very short period are identified as abnormal operations and removed to improve data reliability. The item's influence is weighted and coupled with the browsing behavior trajectory vector to obtain the item's attention score. A structured coupling weight parameter is set to adjust the ratio between the system's assessed influence and the user's subjective attention behavior. The weighted combination operation is performed according to the preset coupling calculation rules to obtain the item's attention score. A threshold detection mechanism is set during the calculation process. If an item has extremely low influence but significantly abnormal browsing behavior, a verification flag is triggered to avoid sorting errors caused by simple misoperation. After the calculation is completed, duplicate value detection and missing value filling are performed on all items to ensure the completeness of the scoring results. Sort the items by attention score from highest to lowest to generate a focus index sequence reflecting the current user's key concerns. Combine the attention scores and item identifiers of all items into an index pair set, and sort them in descending order of score value using a sorting rule. During the sorting process, items with the same score are further sorted according to time priority or risk level priority rules. This generates a focus index sequence containing item identifiers, score values, and corresponding time nodes. Before output, a consistency check is performed to verify that the number of index sequences matches the number of items in the current credit event chain and to check for duplicate indexes or missing items. After passing the check, the focus index sequence is written to the session context cache.

[0022] Semantic Slicing Module: A semantic slicing mechanism is introduced around the focus index sequence to split the corresponding credit event chain into interpretable units. Combined with the fluctuation range of account quota utilization, the distribution of overdue duration days, and query density segments, a set of interpretable fragments with risk-indicating markers is formed.

[0023] The process of splitting the corresponding credit event chain into interpretable units in the semantic slicing module is as follows: Based on the focus index sequence, the corresponding credit event subchain is located. The basic account information, time segment, and status change records associated with the credit event subchain are extracted. The standardized fields in the credit report are uniformly encoded, and a terminology mapping rule for credit reporting business scenarios is constructed to map each field to the corresponding semantic label. According to the entry identifier and time node recorded in the focus index sequence, the corresponding segment data in the credit event chain is located, and the credit event subchain is formed. The account number, account type, credit limit, balance status, repayment status, and status update time fields corresponding to the subchain are extracted. The field values ​​are uniformly encoded and converted. For example, the repayment status is converted from a numerical code to an internal status enumeration identifier. A terminology mapping table is established to map the enumeration identifier to the semantic expression of credit reporting business. For example, the overdue level code is mapped to semantic labels such as mild overdue and continuous overdue. During the mapping process, a legality check is performed to detect whether there are undefined codes or abnormal values ​​in the fields. If they exist, an abnormal mark is recorded and a preset catch-all semantic label is used. The credit event subchain data structure with semantic label fields is output. A semantic slicing rule set is established based on terminology mapping rules. Context window scanning is performed on the credit event subchain to merge event records with consistent state characteristics within consecutive time segments into the same semantic fragment. The semantic slicing rule set includes time continuity determination rules, state consistency determination rules, and maximum window length limit rules. The time continuity determination rules are used to detect whether the time difference between adjacent records is within a preset period. The state consistency determination rules are used to compare whether the semantic labels of adjacent records are consistent. A sliding window scanning method is used, traversing each credit event subchain one by one from the starting time node. When time continuity and consistent semantic labels are detected, the record is included in the current window. When time discontinuity or semantic label changes are detected, the current window is closed and a semantic fragment is generated. A maximum window span limit is set to avoid semantic generalization due to excessive span. Several semantic fragment sets are generated after the scan is completed. The system identifies time points where status changes occur and divides the event records corresponding to these time points into independent boundary segments. During the sliding window scanning process, it performs differential detection on the semantic labels of adjacent time points. When a change in repayment status, account status, or delinquency level is detected, the time point is marked as a status mutation node. Boundary identification is performed on the records before and after the mutation node, and the record containing the mutation node is divided into a separate boundary segment. At the same time, mutation rationality detection is performed, such as determining whether there are duplicate mutations or abnormal jump values. If an anomaly is found, the original data fields are called for verification. After verification and confirmation, an independent boundary segment is generated, and the mutation type and mutation direction information are recorded in the segment. The formed semantic segments and independent boundary segments are structurally encapsulated to generate interpretable units containing semantic labels, time intervals, and account identification information.

[0024] The process of forming a set of explanatory fragments with risk-indicating markers in the semantic slicing module is as follows: The system reads the total credit limit and actual balance of the account corresponding to the interpretable unit within a continuous billing cycle, calculates the ratio sequence of the balance to the total credit limit, and extracts the start and end times, the highest occupancy ratio, and the number of cycles required to recover to the normal range for consecutive high occupancy intervals. It reads the total credit limit field and the ending balance field for each billing cycle from the account index associated with the interpretable unit, calculates the balance occupancy ratio for each billing cycle (i.e., the balance divided by the total credit limit), performs interval normalization on the calculation results, and sets a high occupancy threshold. When the balance occupancy ratio exceeds the preset threshold, it is marked as a high occupancy state. A continuous detection logic scans the ratio sequence; when adjacent cycles meet the high occupancy condition, it is determined to be a consecutive high occupancy interval. The start and end cycles of this interval are recorded, and the highest occupancy ratio is extracted within the interval. When the ratio recovers to below the threshold, the number of cycles from the end of the interval to recovery to the normal range is counted. Anomaly detection is performed during the calculation process; for example, if the total credit limit is zero or missing, it is marked as an anomaly, and the data for that cycle is removed, thus ensuring the validity of the ratio calculation results. The system retrieves repayment records and status change flags within a given time period, calculates the actual number of overdue periods, the longest single overdue period, and the number of consecutive overdue periods, and distinguishes between occasional single-period overdue and consecutive overdue periods across billing cycles. Based on the time interval corresponding to the interpretable unit, it extracts the repayment records and repayment status fields within that time period, checks for overdue flags in each billing cycle, and counts the number of overdue billing periods. For each overdue period, it reads the overdue days field, calculates the longest single overdue days, and uses a time-series scanning method to detect consecutive overdue situations. When overdue flags exist in adjacent billing cycles, they are included in the consecutive overdue period count. If only a single billing cycle is overdue and the preceding and following cycles are in normal status, it is determined to be an occasional single-period overdue. If multiple consecutive billing cycles are overdue, it is determined to be a consecutive overdue period across cycles. At the same time, it sets up data consistency verification logic. If the number of overdue days is negative or exceeds the billing cycle length, it is marked as abnormal and the original message is called for review. The system synchronously retrieves credit inquiry records, statistically analyzes the number of loan approval inquiries, credit card approval inquiries, and the number of different financial institutions within a time window, and calculates the time interval between two adjacent inquiries. It then filters query data matching the interpretable unit time interval from the credit inquiry record table, distinguishing between loan approval and credit card approval queries based on the query type field, and statistically analyzes the number of inquiries for each type. It also statistically analyzes the number of different financial institutions using the institution code field, sorts all query records by timestamp, and calculates the time interval between two adjacent inquiries using a time difference calculation method. A minimum interval threshold is set to identify intensive query behavior within a short period. If the query type field is missing or not within the preset type set, the query is marked as invalid and excluded. The system outputs statistical results including the number of inquiries, the number of institutions, and the shortest query interval. Finally, it generates risk labels based on the risk level corresponding to the peak balance occupancy ratio, consecutive overdue periods, longest overdue days, and the number of approval inquiries, and attaches these labels to the corresponding semantic fragments, forming a set of interpretable fragments with risk-indicating tags.

[0025] The conversation management module builds a multi-turn dialogue context cache based on the set of explanatory fragments, records the user's follow-up question path, terminology misunderstanding deviations and reading pause nodes, and generates a dynamic explanatory framework for a single credit record through intent recognition and context completion mechanisms, while simultaneously verifying the consistency between the explanatory content and the original data fields.

[0026] The process of recording user follow-up questions, misunderstandings of terminology, and reading pause points in the conversation management module is as follows: Each user question is assigned a unique session identifier, and the set number of the explanatory fragments corresponding to the question is recorded. When a user initiates a question event, a unique session identifier is generated by combining the timestamp and the user identifier and written into the session control table. At the same time, an association index between the question and the fragment set is established based on the explanatory fragment set number displayed on the current page. To avoid duplicate recording, duplicate questions within the same time window are deduplicated. If the text similarity exceeds a preset threshold, they are merged into the same follow-up question node. The recorded fields include session identifier, question text, trigger time, associated fragment number, and current risk label status, thus forming the basic data for a traceable follow-up question path. Keyword extraction is performed on user input statements, and they are matched with risk indicator tags and semantic tags to mark the hit relationship. The user input text is segmented and stop word filtered to extract core keywords. Through a preset terminology dictionary and semantic tag mapping table, the keywords are matched with semantic tags and risk indicator tags in the explanation fragment. The matching methods include exact matching and similarity matching. When the similarity between the keyword and the tag is higher than a preset threshold, it is determined to be a hit, and the hit tag number and hit type are recorded. If no tag is matched, the semantic expansion mechanism is triggered, and a second matching is performed through the synonym mapping table to form a hit relationship record table between keywords and risk tags. The system tracks user dwell time and click frequency on explanation pages to identify potential comprehension biases. Question history, hit relationships, and bias assessment results are stored in a context cache. A behavior monitoring component is embedded in the explanation page to record the timestamps of user entry and exit from the segment page, calculate effective dwell time, set a minimum effective dwell threshold, filter rapid scrolling behavior, and track the number of times users repeatedly click and expand the same explanation segment. When dwell time exceeds a preset high threshold and is accompanied by repeated clicking behavior, it is identified as a potential comprehension bias. Simultaneously, considering hit relationships, if a user's question keywords are highly relevant to a risk label but are repeatedly searched, it is marked as a terminology comprehension bias. The session identifier, question text, hit label number, dwell time, click frequency, and bias assessment results are written to the context cache.

[0027] The process of synchronizing and verifying the consistency between the description content and the original data fields in the session management module is as follows: The large model Agent performs semantic parsing on the current session state data structure to identify the user's question intent category. It reads the question text, historical follow-up question paths, and hit tag records corresponding to the current session identifier from the context cache to construct a structured session state data object. This data object is then input into the large model Agent for semantic parsing, outputting intent category labels and confidence values. The preset intent category set includes types such as explanation requests, reason follow-ups, time confirmation, amount verification, and risk level inquiries. Confidence checks are performed on the parsing results. When the confidence level is lower than a preset threshold, a secondary confirmation mechanism is triggered. Keyword matching rules are used for auxiliary verification to determine the intent category and write it into the session state field. By combining semantic fragments and risk tags already recorded in the context cache, contextual completion is performed for missing time intervals or account information. Based on the identified intent category, the associated semantic fragment IDs and risk tag sets are retrieved from the context cache to check whether the current question contains a complete account ID or time interval field. If missing, it is completed by default according to the most recently hit fragment ID. For questions spanning time periods, time coverage detection logic is executed. If the time involved in the question is not within the current fragment interval, the credit event chain data is called for extended querying. At the same time, the legality of the completed information is verified, such as checking whether the account ID exists in the current user dataset and whether the time interval exceeds the range of credit records. The explanation logic is organized according to the preset explanation template to generate a dynamic explanation framework. The amount, date, and status fields involved in the explanation content are verified item by item through the field backtracking interface.

[0028] Source tracing and reasoning module: When it is detected that a user has doubts about the credit granting results or changes in the score, the large model agent triggers the decision source tracing task planning instruction, which is broken down into three sub-tasks: score interval retrieval, rule hit record retrieval, and item weight comparison. Data collection is completed through internal interface calls, and after integrating the output results of each sub-task, a traceable impact chain description is generated.

[0029] The process of breaking down the source tracing and reasoning module into three sub-tasks—score range retrieval, rule hit record retrieval, and item weight comparison—is as follows: When the user intent category is determined to be a decision interpretation type, a decision tracing task identifier is generated and associated with the credit approval number and scoring timestamp corresponding to the current query. The determined intent category label is received from the session management module. When the intent category is detected to belong to the preset decision interpretation type set, the tracing task initialization process is triggered. A unique decision tracing task identifier is generated by using the timestamp and user identifier. At the same time, the approval number and scoring generation timestamp corresponding to the current query are read from the credit approval record table. Existence verification and time validity check are performed. If the approval number does not exist or the scoring time exceeds the data retention period, an exception flag is returned and the task is terminated. After the verification is passed, the task identifier, approval number and scoring timestamp are bound together to form a structured tracing task object. The decision tracing task is divided into three sub-tasks: a scoring interval retrieval sub-task, a rule hit record retrieval sub-task, and an item weight comparison sub-task. Each sub-task is assigned a corresponding data interface address and access permission flag, and a sequential execution queue is established based on data dependencies. After the tracing task object is generated, it is divided into three independent sub-task objects according to preset task decomposition rules. The scoring interval retrieval sub-task is assigned a scoring rule library interface address for retrieving scoring interval boundary values ​​and risk level descriptions; the rule hit record retrieval sub-task is assigned a rule execution log interface address for retrieving rule trigger records during approval; and the item weight comparison sub-task is assigned a scoring rule library interface address for retrieving rule hit record retrieval sub-tasks for retrieving rule trigger records during approval. The parameter storage interface address is used to obtain the item weights and values ​​of the items participating in the scoring calculation. An access permission flag is assigned to each interface, and the permission verification logic is executed to confirm that the current session has the access permission. A sequential execution queue is established according to the data dependency relationship. The scoring interval retrieval and rule hit record retrieval can be executed in parallel. The item weight comparison subtask depends on the scoring identification information output by the aforementioned subtask. After the queue is generated, an integrity check is performed to confirm that all three types of subtasks have been registered and the interface addresses are valid. The sequential execution queue is sent to the execution engine, and the data retrieval, calculation processing and result encapsulation are completed in the predetermined order, entering the data recycling and interpretation generation stage.

[0030] The process of generating a traceable impact chain description after integrating the output results of each subtask in the source tracing and reasoning module is as follows: The scoring system interface is called to obtain the user's current score, the score range, and the upper and lower threshold scores of the corresponding range. The rule log interface is called to read the rule number, trigger time, trigger condition, and associated credit field identifier that were triggered during this credit assessment. Based on the feature importance ranking results returned by the scoring system, the feature contribution value corresponding to the credit field identifier is retrieved. Combined with the actual value of the current user's credit field, the marginal impact score of the credit field on the final scoring result is calculated as the influence weight of the corresponding credit entry. The feature importance ranking list is obtained from the scoring system interface, which contains each feature identifier and its importance coefficient. Based on the credit field identifier returned by the rule log interface, the feature importance ranking list is matched and located. The scoring parameter interface is called to retrieve the contribution value calculation rule of the corresponding feature, and the actual value of the current user on the credit field is read. According to the preset contribution calculation formula, the feature importance coefficient and the difference between the actual value of the user are combined and calculated to obtain the marginal impact score of the credit field on the scoring result. Boundary detection logic is executed during the calculation process. For example, if the actual value is missing or exceeds the legal range, the default correction coefficient is used. The calculated marginal impact score is written into the entry influence weight table. For credit entries with rule-triggered records, the corresponding segment adjustment values ​​for rule hits are superimposed to form an entry-level comprehensive impact weight. The impact chain description is output according to the evaluation time order and the size of the comprehensive impact weight. The selected set of rule-triggered records is traversed one by one. The corresponding segment adjustment value is queried in the rule parameter table according to the rule number. The rule adjustment value is superimposed with the marginal impact score to generate an entry-level comprehensive impact weight. During the superposition process, a sign consistency check is performed to ensure that the direction of the bonus and deduction items is correct. After the weight calculation of all entries is completed, they are sorted by time according to the rule trigger time and the occurrence time of the credit field, and at the same time, they are sorted by priority according to the size of the comprehensive impact weight.

[0031] Interactive Generation Module: Combining users' historical consultation behavior with the current conversation depth, the module adjusts the granularity and level of expression of responses, and outputs interactive interpretation results of the impact chain description while maintaining data compliance boundaries.

[0032] The process of outputting interactive interpretation results affecting the link description in the interactive generation module, while maintaining data compliance boundaries, is as follows: By analyzing the proportion of user attention to three types of content in historical consultation records—terminology explanation, rating mechanism, and improvement suggestions—a user understanding preference vector is constructed. Calculate the conversation depth level based on the current conversation round and the complexity of follow-up questions; select the corresponding explanation template and expansion level based on the understanding preference vector and the conversation depth level; and summarize the description of the influencing links. Before outputting, compliance checks are performed, sensitive information involving policy thresholds is masked, and interactive interpretation results affecting the link description are output.

[0033] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0034] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart credit report interpretation and question-answering system based on a large-scale agent model, characterized in that, include: Message Chaining Module: Accesses original personal credit reports and formatted display files, parses credit account status fields, query source identifiers and repayment performance time series, reconstructs a credit event chain interwoven with account and time axis dimensions, and generates a focus index sequence based on the impact of items and user browsing history; Semantic Slicing Module: A semantic slicing mechanism is introduced around the focus index sequence to split the corresponding credit event chain into interpretable units. Combined with the fluctuation range of account quota utilization, the distribution of overdue days, and query density segments, a set of interpretable fragments with risk-indicating markers is formed. Conversation Management Module: Based on the set of explanatory fragments, a multi-turn dialogue context cache is built to record the user's follow-up question path, terminology comprehension deviations and reading pause nodes. Through intent recognition and context completion mechanisms, a dynamic explanatory framework for a single credit record is generated, and the consistency between the explanatory content and the original data fields is verified simultaneously. Source tracing and reasoning module: When it is detected that a user has doubts about the credit granting results or changes in the rating, the large model Agent triggers the decision source tracing task planning instruction, which is broken down into three sub-tasks: rating interval retrieval, rule hit record retrieval, and item weight comparison. Data collection is completed through internal interface calls. After integrating the output results of each sub-task, a traceable impact chain description is generated. Interactive Generation Module: Combining users' historical consultation behavior with the current conversation depth, the module adjusts the granularity and level of expression of responses, and outputs interactive interpretation results of the impact chain description while maintaining data compliance boundaries.

2. The intelligent credit report interpretation and question-answering system based on a large-scale agent model as described in claim 1, characterized in that, The process of reconstructing the credit event chain that intertwines the account dimension and the time axis dimension is as follows: The interface adaptation component reads the message data from the credit reporting center and the formatted file displayed to users, and performs unified and standardized conversion of field encoding, data separators and time formats. Extract account number, account type, credit limit, current balance, repayment status, last repayment date, and query institution code to build an account primary key index; Based on the repayment and performance timeline, a monthly behavior record table is constructed for each account, mapping overdue identification, extension records, and account status change information to the corresponding time nodes; By using a cross-account time alignment mechanism, the behavior records of multiple accounts are spliced ​​along a unified timeline to form a credit event chain with a dual index structure of account dimension and time dimension.

3. The intelligent credit report interpretation and question-answering system based on a large-scale agent model as described in claim 2, characterized in that, The process of generating a focus index sequence based on the impact of items and user browsing history is as follows: Based on the credit event chain, an item impact assessment function is constructed. The assessment function comprehensively considers the historical contribution weight of delinquency level, balance occupancy ratio and query frequency to credit granting decision. Collect user swipe duration, click-to-expand behavior, and highlighted actions in the report interface to form a browsing behavior trajectory vector; The item's influence is weighted and coupled with the browsing behavior trajectory vector to obtain the item's attention score. Sort the items by their attention score from highest to lowest, and generate an index sequence of attention points that reflects the current user concerns.

4. The intelligent credit report interpretation and question-answering system based on a large-scale agent model as described in claim 3, characterized in that, The process of breaking down the corresponding credit event chain into interpretable units is as follows: Based on the focus index sequence, the corresponding credit event subchain is located, and the basic account information, time period and status change records associated with the credit event subchain are extracted. The standardized fields in the credit report are uniformly encoded, and terminology mapping rules for credit reporting business scenarios are constructed to map each field to the corresponding semantic label. A semantic slicing rule set is established based on term mapping rules. Context window scanning is performed on the credit event subchain to merge event records with consistent state features within consecutive time segments into the same semantic segment. Identify the time points when the state changes and divide the event records corresponding to the time points into independent boundary segments; The resulting semantic fragments and independent boundary fragments are structurally encapsulated to generate interpretable units containing semantic tags, time intervals, and account identification information.

5. The intelligent credit report interpretation and question-answering system based on a large-scale agent model according to claim 4, characterized in that, The process of forming a set of explanatory fragments with risk-indicating markers is as follows: Read the total credit limit and actual balance of the account corresponding to the interpretable unit within a continuous billing cycle, calculate the ratio sequence of the balance to the total credit limit, and extract the start and end time, the highest occupancy ratio, and the number of cycles required to recover to the normal range for the continuous high occupancy interval. Retrieve repayment records and status change indicators within the time period, calculate the actual number of overdue periods, the longest single overdue days, and the number of consecutive overdue periods, and distinguish between occasional single-period overdue and consecutive overdue situations across billing cycles. Synchronously retrieve credit inquiry records, count the number of loan approval inquiries, credit card approval inquiries, and the number of different financial institutions within a time window, and calculate the time interval between two adjacent inquiries; Risk labels are generated based on the peak balance occupancy ratio, consecutive overdue periods, longest overdue days, and number of approval-related queries according to the risk level corresponding to the numerical range. These labels are then attached to the corresponding semantic fragments to form a set of explanatory fragments with risk-indicating markers.

6. The intelligent credit report interpretation and question-answering system based on a large-scale agent model according to claim 5, characterized in that, The process of recording user follow-up questions, misunderstandings of terminology, and reading pause points is as follows: Assign a unique session identifier to each user question and record the set number of the explanation fragments corresponding to the question; Keyword extraction is performed on user input statements, and they are matched with risk-indicating markers and semantic tags to mark the hit relationship; The system tracks the time users spend on the explanation page and the number of repeated clicks to identify potential misunderstandings. The system then stores the question history, hit relationships, and misunderstanding results in the context cache.

7. The intelligent credit report interpretation and question-answering system based on a large-scale agent model according to claim 1, characterized in that, The process of synchronously verifying the consistency between the description content and the original data fields is as follows: The large model Agent is used to perform semantic parsing on the current session state data structure to identify the user's question intent category; By combining the semantic fragments and risk tags already recorded in the context cache, contextual completion is performed for missing time intervals or account information; The explanation logic is organized according to the preset explanation template, and a dynamic explanation framework is generated. The amount, date and status fields involved in the explanation content are verified item by item through the field backtracking interface.

8. The intelligent credit report interpretation and question-answering system based on a large-scale agent model according to claim 1, characterized in that, The process can be broken down into three sub-tasks: score range retrieval, rule hit record retrieval, and item weight comparison. When the user intent category is determined to be the decision interpretation type, a decision tracing task identifier is generated and associated with the credit approval number and scoring timestamp corresponding to the current query; The decision tracing task is divided into a scoring interval retrieval subtask, a rule hit record retrieval subtask, and an item weight comparison subtask. The three subtasks are assigned corresponding data interface addresses and access permission tags, and a sequential execution queue is established based on data dependencies. The sequential execution queue is sent to the execution engine, where data retrieval, computation, and result encapsulation are completed in a predetermined order, before entering the data recycling and interpretation generation stage.

9. The intelligent credit report interpretation and question-answering system based on a large-scale agent model according to claim 1, characterized in that, The process of generating a traceable impact chain description after integrating the output results of each subtask is as follows: The scoring system interface is called to obtain the user's current score, the score range, and the upper and lower threshold scores of the corresponding range. The rule log interface is called to read the rule number, trigger time, trigger condition, and associated credit field identifier that were triggered during this credit assessment. Based on the feature importance ranking results returned by the scoring system, retrieve the feature contribution value corresponding to the credit field identifier, and combine it with the actual value of the current user's credit field to calculate the marginal impact score of the credit field on the final scoring result, which serves as the impact weight of the corresponding credit item; For credit entries with rule trigger records, the segmented adjustment values ​​corresponding to the rule hit are superimposed to form an entry-level comprehensive impact weight. The impact chain description is output according to the evaluation time order and the size of the comprehensive impact weight.

10. The intelligent credit report interpretation and question-answering system based on a large-scale agent model according to claim 9, characterized in that, The process of outputting interactive interpretation results that affect the link description while maintaining data compliance boundaries is as follows: By analyzing the proportion of user attention to three types of content in historical consultation records—terminology explanation, rating mechanism, and improvement suggestions—a user understanding preference vector is constructed. Calculate the conversation depth level based on the current conversation round and the complexity of follow-up questions; select the corresponding explanation template and expansion level based on the understanding preference vector and the conversation depth level; and summarize the description of the influencing links. Before outputting, compliance checks are performed, sensitive information involving policy thresholds is masked, and interactive interpretation results affecting the link description are output.

Images