A method and system for compliant version management and automatic upgrade of financial intelligent question answering

Abstract

This invention provides a method and system for compliance version management and automatic upgrade of financial intelligent question-and-answer software, relating to the field of version management and upgrade technology, including: compliance rule version management, compliance review, evidence fragment responsibility score management, automatic generation of compliance training data, and automatic fine-tuning and upgrade of the review model. This invention can assign a unique policy version number to each update of financial compliance rules, binding and storing it with the review model version, effective date, and rule difference information; retrieve evidence fragments using user questions and answers, and after the compliance review model determines the risk level, block high-risk content and trigger review for low-confidence results; backfill the evidence fragments with the review conclusions and calculate the responsibility score bound to the version; freeze the old score when the rule changes and re-accumulate and decay it in the new version; automatically collect data and generate triplet training samples in blocking and review scenarios; finally, filter data based on rule differences, fine-tune the adapter and bind it to the new version for deployment, and remove the adapter when the rule is rolled back.

Description

Technical Field

[0001] This invention relates to the field of version management and upgrade technology, and in particular to a method and system for compliant version management and automatic upgrade of financial intelligent question answering. Background Technology

[0002] In the financial sector, intelligent question answering commonly employs a Retrieval-Augmented Generation (RAG) technical architecture. This architecture retrieves evidence fragments from internal institutional documents such as private policy clauses, product descriptions, announcements, risk control guidelines, and training materials, serving as the basis for a large language model to generate answers. However, the retrieved fragments often contain outdated or invalid information, sensitive data, mismatched scope of application, or incomplete descriptions. This can easily lead to compliance risks such as the model output being interpreted as investment advice, lacking disclaimers, citing unpublished data, or referencing invalid clauses. Therefore, existing systems typically add a compliance review model before answer output to intercept or downgrade responses. However, financial industry compliance rules are subject to frequent changes. The same statement that is compliant under old rules may be non-compliant under new rules, and existing technologies struggle to adapt to version mismatch issues arising from dynamic rule changes. On one hand, existing compliance review model update schemes only focus on optimizing the model's own training set and updating it online, failing to link compliance review results with rule versions and backfill them into the evidence fragments. This fails to address the risk mismatch caused by rule changes in historical evidence fragments and also lacks a version-based risk governance mechanism for evidence fragments. On the other hand, existing RAG optimization techniques only focus on retrieval relevance and illusion suppression, failing to use compliance review conclusions as a supervisory signal to perform risk labeling and version management of evidence fragments. Furthermore, existing evidence attribution methods are only used to explain the contribution of evidence to the answer, lacking the ability to assess risk, link versions, and manage long-term evolution for financial compliance.

[0003] In summary, existing technologies cannot achieve version coordination among compliance rules, review models, and evidence fragments. When rules change, they are prone to erroneous approvals or interceptions, and it is difficult to provide traceable audit evidence. Therefore, they cannot meet the core requirements of financial intelligent question answering for compliance consistency, dynamic adaptability, and auditability. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a method and system for financial intelligent question-and-answer compliance version management and automatic upgrade, which can assign a unique policy version number to each update of financial compliance rules, and bind and store it with the review model version, effective time, and rule difference information; retrieve evidence fragments based on user questions and answers, block high-risk content after the compliance review model judges it, and trigger review for low-confidence results; fill the evidence fragments with the review conclusion and calculate the responsibility score bound to the version, freeze the old score when the rule changes, and re-accumulate and decay it in the new version; automatically collect data and generate triple training samples in blocking and review scenarios; finally, filter data according to rule differences, fine-tune the adapter and bind it to the new version for online, and remove the adapter when the rule is rolled back.

[0005] This invention provides a method for financial intelligent question-and-answer compliance version management and automatic upgrade. The method includes: Step 1, constructing a compliance rule base, assigning an independent policy version number to each update of financial compliance rules, and binding and storing each policy version with the review model version, effective time, and rule difference information to form a one-to-one correspondence between rules, models, and versions; Step 2, using user question-and-answer input as a query to obtain a set of evidence fragments, inputting the search results into a preset compliance review model, and outputting compliance judgment results, confidence levels, and risk labels, directly blocking the output of high-risk content, and triggering manual review for low-confidence results; Step 3, backfilling the compliance review conclusions into the corresponding evidence fragments, calculating and updating for each evidence fragment. Fragment-level responsibility scores are used to form version-bound risk scores. When compliance rules change, the responsibility scores of the old version are frozen, and the responsibility scores are re-accumulated and migrated and decayed under the new version. Step four: In the scenarios of blocking output and low-confidence review, the original questions and answers, retrieved evidence fragments, and model output content are automatically recorded. Based on the large model, risk trigger points and evidence fragments are aligned to generate triple training data. Step five: Based on rule differences, the associated rule IDs are extracted, the corresponding training data is filtered, and the adapter of the basic model is quickly fine-tuned. After the fine-tuning is up to standard, the new adapter is bound to the new version policy and automatically launched. When the rule is rolled back, the corresponding adapter is removed, so that the review model can be automatically upgraded with the version of the compliance rules.

[0006] This application also provides a financial intelligent question-and-answer compliance version management and automatic upgrade system. This system is applied to a method for financial intelligent question-and-answer compliance version management and automatic upgrade. The system includes: a compliance rule version management module, a compliance review module, an evidence fragment responsibility score management module, a compliance training data automatic generation module, and a review model automatic fine-tuning and upgrade module. The compliance rule version management module is used to build a compliance rule library, assigning an independent policy version number to each update of financial compliance rules, and binding and storing each policy version with the review model version, effective date, and rule difference information to form a one-to-one correspondence between rules, models, and versions. The compliance review module uses user question-and-answer input as a query to obtain a set of evidence fragments, inputs the search results into a preset compliance review model, and outputs compliance judgment results, confidence levels, and risk labels. High-risk content is directly blocked from output, and low-confidence results trigger manual intervention. The review and evidence fragment responsibility score management module is used to backfill compliance review conclusions to the corresponding evidence fragments, calculate and update fragment-level responsibility scores for each evidence fragment, forming version-bound risk scores. When compliance rules change, the old version responsibility scores are frozen, and the responsibility scores are re-accumulated and migrated and decayed under the new version. The compliance training data automatic generation module is used to automatically record the original Q&A, retrieve evidence fragments, and model output content in scenarios of blocked output and low-confidence review. Based on the large model, risk trigger points and evidence fragments are aligned to generate triple training data. The review model automatic fine-tuning and upgrade module is used to extract associated rule IDs based on rule differences, filter corresponding training data, and quickly fine-tune the adapter of the basic model. After the fine-tuning is up to standard, the new adapter is bound to the new version policy and automatically launched. When the rule is rolled back, the corresponding adapter is removed, so that the review model can be automatically upgraded with the compliance rule version.

[0007] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: By constructing a closed-loop system encompassing compliance rule version management, RAG retrieval compliance review, evidence fragment responsibility score control, automatic generation of triplet training data, and automatic fine-tuning and upgrading of the review model, problems such as version mismatch, inaccurate evidence risk, and lagging model updates caused by frequent changes in compliance rules in financial intelligent question answering can be effectively solved. By assigning an independent policy version number to each rule change and binding and storing version, model, rule differences, and effective time, precise coordination among rules, models, and versions is achieved, avoiding misjudgments and omissions due to version incompatibility. Based on the RAG retrieval and hierarchical compliance review mechanism, high-risk content is directly blocked, and low-confidence results trigger manual review, strictly controlling compliance bottom lines from the output end and significantly reducing the risk of releasing non-compliant content. By backfilling compliance review conclusions and forming version-bound fragment-level responsibility scores, freezing old scores when rules change and migrating and decaying them in the new version, the evidence risk status evolves synchronously with the rules, improving auditability and judgment consistency. Automatic data collection and generation of triplet training samples in blocking and low-confidence scenarios provide high-quality incremental data for model iteration, eliminating the need for extensive manual annotation. Leveraging lightweight adapters for rapid fine-tuning, version-bound deployment, and rule rollback / removal mechanisms, the review model automatically upgrades with compliance rules, quickly adapting to new regulations without altering the underlying model, significantly reducing iteration costs and deployment risks. The overall solution enhances the compliance stability, dynamic adaptability, and regulatory traceability of the financial intelligent question-answering system, ensuring continuous compliance, accurate judgments, and efficient operation even in scenarios with frequent rule changes, meeting the stringent compliance control requirements of the financial sector. Attached Figure Description

[0008] Figure 1 This is a flowchart of a financial intelligent question-and-answer compliance version management and automatic upgrade method provided by an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a financial intelligent question-and-answer compliance version management and automatic upgrade system provided in an embodiment of the present invention. Detailed Implementation

[0009] This invention provides a method for compliant version management and automatic upgrade of financial intelligent question answering systems, such as... Figure 1 The flowchart shown is a method for compliant version management and automatic upgrade of financial intelligent question answering. The processing flow of this method may include the following steps: Step 1: Build a compliance rule library, assign an independent policy version number to each update of financial compliance rules, and bind and store each policy version with the review model version, effective date, and rule difference information to form a one-to-one correspondence between rules, models, and versions.

[0010] It should be understood that a compliance rule library for the financial sector is pre-built and continuously maintained to uniformly store and manage all compliance requirements, regulatory rules, risk control standards, institutional clauses, and review criteria. When financial compliance rules are added, revised, or repealed due to regulatory updates or internal system adjustments, a unique and independent policy version number is assigned to this rule change to distinguish the compliance rule system at different times. Based on the comparison and parsing of the old and new rule texts, structured rule difference information containing added entries, deleted entries, modified content, and corresponding handling methods is automatically extracted and generated. A one-to-one binding relationship is established between this policy version number and the compliance review model version adapted to the current rule, the official effective date of the rule, and the structured rule difference information. The version number, model version, effective date, rule difference information, and related rule identifiers are uniformly and persistently stored in the rule management database, ultimately forming a version mapping system in which compliance rules, review models, and policy versions correspond to each other and are consistent.

[0011] It needs to be explained that the specific process for building a compliance rule base is as follows: A financial compliance rule base is pre-built and maintained, which stores all compliance requirements, regulatory rules, risk control standards and audit criteria in the financial field. If financial compliance rules are added, revised, or repealed, a unique and independent policy version number will be assigned to this rule change; Based on the comparison and analysis of the new and old rules, structured rule difference information including newly added entries, deleted entries, modified content and handling methods is generated; A one-to-one binding relationship is established between the policy version number and the corresponding compliance review model version, rule effective time, and structured rule difference information. The overall associated data of version number, model version, effective time, rule difference information, and associated rule identifier is persistently stored in the preset rule management database, thereby forming a version mapping system in which compliance rules, review models, and policy versions correspond one-to-one and are coordinated and consistent.

[0012] In this embodiment, a compliance rule base for the financial sector is pre-built and continuously maintained. This rule base adopts a hybrid storage architecture of structured database and vector library, uniformly collecting and managing all compliance requirements, regulatory clauses, risk control standards, prohibitive norms, and audit standards of the financial industry. Each rule is assigned a globally unique rule ID. When financial compliance rules are added, revised, or repealed due to updates in regulatory policies or internal system revisions, the system uses a rule change detection algorithm to automatically identify the differences and assign a globally unique and non-reusable policy version number (policy_version) to this rule change. Subsequently, based on semantic similarity algorithms and text difference algorithms, the old and new rule texts are compared and logically parsed sentence by sentence, automatically extracting added entries, deleted entries, modified content, and risk handling methods, generating structured and computable rule difference information (policy_diff). Furthermore, through a version association mapping algorithm, a strong one-to-one correspondence is established between the policy version number and the corresponding compliance review model version, rule effective time, structured rule difference information, and associated rule ID set. The above-mentioned associated data is persistently stored in the preset rule management database using a hash verification method to ensure that the data is tamper-proof, traceable, and auditable. Ultimately, a version mapping system is formed that accurately corresponds and coordinates the compliance rules, review models, and policy versions, providing unified and reliable version management support for subsequent compliance reviews, evidence responsibility score calculation, and automatic model upgrades.

[0013] Step 2: Using user question and answer input as the query, obtain a set of evidence fragments, input the search results into the preset compliance review model, and output compliance judgment results, confidence level and risk label. High-risk content is directly blocked from output, and low-confidence results are triggered for manual review.

[0014] It should be understood that the user-inputted question and answer text is used as a search query. An enhanced search architecture is used to perform a full-text and vector-based hybrid search on private domain materials such as internal financial institution regulations, product manuals, official announcements, risk control implementation guidelines, business training materials, and regulatory responses. This searches match highly relevant evidence fragments and form an evidence fragment set. This evidence fragment set is then input into a pre-trained compliance review model. The model evaluates the compliance, timeliness, scope of application, completeness of expression, and sensitive information of each evidence fragment, outputting the corresponding compliance judgment result, confidence score, and risk type label. When the judgment result is non-compliant and the risk level reaches a preset blocking threshold, the system directly blocks the question and answer output, returning no content to the user and simultaneously recording the blocking event. When the compliance judgment result is compliant but the confidence score is below the preset threshold, a manual review process is automatically triggered. The question and answer input, evidence fragments, model output, confidence score, and risk label are pushed to the compliance review end, where reviewers confirm the authenticity and compliance, thus achieving real-time compliance control over the question and answer output content.

[0015] It should be further explained that the specific process for obtaining a set of evidence fragments based on user question-and-answer input is as follows: The user-inputted question and answer text is used as a search query. Through search enhancement, relevant evidence fragments are matched and retrieved from the private domain data of financial institutions to form a set of evidence fragments corresponding to the current question and answer. The financial institution's internal private data includes internal rules and regulations, product manuals, official announcements, risk control implementation guidelines, business training materials, and regulatory responses. The acquired set of evidence fragments, along with user question and answer input, is input into a pre-trained compliance review model. Based on the compliance review model, the compliance, timeliness, scope of application, completeness of expression, and sensitive information of the evidence fragments are judged item by item. The model outputs the compliance judgment result, confidence score, and risk type label for the corresponding evidence fragments and question and answer content. The risk type label includes expired clauses, investment advice-like statements, missing disclaimers, unpublished data, prohibited statements, and mismatch of scope of application.

[0016] Using user question-and-answer input as the query, a set of evidence fragments is obtained, which also includes: If the compliance determination result is non-compliant and the risk level reaches the preset blocking threshold, the output of the question and answer will be blocked directly, no content will be returned to the user, and the blocking event, corresponding evidence fragments and triggering rule version will be recorded simultaneously. If the compliance determination result is compliant but the model output confidence score is lower than the preset threshold, it is determined that the model is unable to accurately identify risks, and the manual review process is automatically triggered. The user's Q&A input, evidence fragment set, model output content, confidence score and risk label are pushed to the compliance review end, where the reviewer confirms the authenticity and compliance of the content and provides feedback on the review conclusion. By archiving blocking records, low-confidence samples, manual review results, and corresponding evidence fragments in a unified manner, the risk of non-compliant content being exported is reduced.

[0017] In this embodiment, the user-inputted question-and-answer text is used as the search query. A search-enhanced generation architecture combined with a bidirectional semantic matching algorithm and a keyword-weighted search algorithm is employed to perform multi-dimensional searches and reordering of private domain materials such as internal financial institution regulations, product manuals, official announcements, risk control implementation guidelines, business training materials, and regulatory responses. This accurately matches and retrieves evidence fragments highly relevant to the current question-and-answer, forming a set of evidence fragments corresponding to the query intent. Subsequently, this set of evidence fragments, along with the user's question-and-answer input, is input into a pre-trained compliance review model. The model uses a multi-dimensional risk discrimination algorithm to quantitatively evaluate the compliance, timeliness, scope of application, completeness of expression, and sensitive information of each evidence fragment. It outputs the compliance judgment result, confidence score, and risk type label for the corresponding evidence fragments and question-and-answer content. The risk type label includes expired clauses, investment advice statements, and missing disclaimers. The system addresses several issues: mismatches in statements, unpublished data, prohibited expressions, and scope of application; when a compliance assessment determines non-compliance and the risk level calculated using a risk level weighting algorithm reaches a preset blocking threshold, the system directly blocks the Q&A output without returning any content to the user, and simultaneously records the blocking event, corresponding evidence fragments, and the version of the triggering rule; when a compliance assessment determines compliance but the model output confidence score is lower than the dynamic confidence threshold, it is determined that the model cannot accurately identify the risk, automatically triggering a manual review process. This process pushes the user's Q&A input, evidence fragment set, model output content, confidence score, and risk label to the compliance review end, where reviewers verify the authenticity and compliance of the content and provide feedback on the review conclusion; finally, the blocking record, low-confidence samples, manual review results, and corresponding evidence fragments are uniformly archived and stored, achieving full-process traceability and auditability, effectively reducing the risk of non-compliant content being output from the output end.

[0018] Step 3: Fill the corresponding evidence fragments with the compliance review conclusions, calculate and update the fragment-level responsibility score for each evidence fragment, and form a version-bound risk score. When the compliance rules change, freeze the old version responsibility score and re-accumulate and migrate the responsibility score under the new version.

[0019] It needs to be explained that after completing the compliance review and judgment of evidence fragments, a precise backfilling operation of the compliance review conclusion must be performed. This involves binding the judgment result obtained from this compliance review one-to-one to the corresponding original evidence fragment. Simultaneously, based on different compliance interception levels and compliance pass types, fixed base scores are configured, automatically calculating and updating the fragment-level responsibility score of each evidence fragment in real time. Furthermore, all risk responsibility scores are strongly bound to the current compliance rule version, forming a version-specific fixed risk score ledger. For compliance rule iteration and change scenarios, a dual mechanism of freezing old version scores and resetting new version scores will be implemented. This means locking and sealing all fragment responsibility scores accumulated under historical rule versions, preventing them from participating in subsequent natural decay. Simultaneously, in the new compliance rule version environment, score migration decay coefficients are configured according to the rule modification magnitude to complete the migration, conversion, and re-accumulation of historical responsibility scores, achieving a smooth transition of risk scores before and after rule iteration, traceability of compliance risks, and data continuity. Among these, the compliance review results can be divided into four judgment levels, each category... All results use a fixed base responsibility score as the benchmark for score calculation. Specifically, a single evidence fragment with direct blocking (serious compliance issues triggering system-mandated interception and prohibition of process flow) has a base responsibility score of 10 points; a single evidence fragment with manual review blocking (obvious compliance risks, requiring manual review and failing the review, resulting in process suspension) has a base responsibility score of 8 points; a single evidence fragment with low confidence but passing manual review (insufficient confidence in content compliance judgment, minor flaws, but confirmed by secondary manual verification to be free of violations) has a base responsibility score of 2 points; and a single evidence fragment that passes normally (the content of the evidence fragment fully complies with the current compliance rules and has no risks or flaws) has a base responsibility score of 0 points. Within the normal operating cycle where the compliance rule version remains unchanged and the rule content is stably effective, all generated evidence fragment responsibility scores will be subject to a periodic natural decay mechanism, with a uniform natural week as the minimum decay cycle. After each complete week, the current cumulative responsibility score of a single evidence fragment is automatically multiplied by a fixed decay coefficient of 0.The score decreases by 95 points month by month and week until it approaches zero. This mechanism is mainly used to weaken the weight of minor historical compliance flaws, preventing old, low-frequency risks from occupying risk assessment indicators for a long time, and aligning with the management needs of dynamically weakening compliance risks. When rules within the compliance system are adjusted, revised, or abolished, a version isolation mechanism will be immediately triggered to separate the old version data from the new version's accounting logic. Specifically, once a compliance rule is updated, the historical responsibility score of all evidence fragments under the previous rule version is immediately frozen. After freezing, the old version score will no longer participate in the weekly natural decay, will no longer incur additional deductions, and will not undergo any data changes. The historical compliance risk ledger will be fully preserved to meet the requirements of compliance traceability and responsibility auditing. After switching to the new compliance rules version, the original scores from the old version are no longer directly accumulated. Instead, four change levels are divided based on the scope and magnitude of the rule modifications, and a differentiated migration attenuation coefficient is matched. The frozen historical responsibility scores from the old version are then converted and migrated. Simultaneously, in the new version, the fragment responsibility scores are re-accumulated based on the newly added compliance audit results. When the rules remain almost unchanged (only minor adjustments such as wording optimization, formatting adjustments, and punctuation corrections without substantial differences in control, while the compliance judgment standards, interception logic, and audit scope remain completely unchanged), the historically frozen old scores are converted and migrated by multiplying the old scores by 0.9, serving as the initial stock risk score for the new version. These scores then participate in the normal weekly natural attenuation of 0.95. When minor modifications are made (minor adjustments to some clauses, slight adjustments to individual compliance judgment thresholds, increases or decreases in the scope of control for edge scenarios, while core compliance constraints and key interception rules remain unchanged), the historically frozen old scores are converted and migrated by multiplying the old scores by 0.6, and incorporated into the new version's score system for unified management and periodic decay. When major rule modifications are made (reconstruction of core compliance clauses, changes to key risk judgment standards, adjustments to interception mechanisms, large-scale iterations of core control scenarios, directly changing the core logic of compliance audits), the historically frozen old scores are significantly converted and migrated by multiplying the old scores by 0.2, weakening the historical risk weight under the old rule system. When a rule is abolished (the corresponding entire compliance rule is completely discontinued, invalidated, or deleted, and is no longer included in the compliance audit system), the old score is converted and migrated. All historical scores for related evidence fragments are immediately reset to zero, and the old score is multiplied by 0. The new version no longer retains the historical risk score corresponding to this rule. The entire process forms a complete closed loop. Each evidence fragment is first bound to a compliance review conclusion, matched with its corresponding base score, generating the initial responsibility score for this version. During the rule's stable period, a 0.95 coefficient decay is automatically applied weekly, dynamically updating the fragment's real-time risk score. When the rule is iterated and updated, the old version's score is statically archived and matched with four conversion coefficients (0.9, 0.6, 0.2, and 0) according to the modification level, completing the migration of historical scores to the new version. During the operation of the new version, newly added violation evidence fragments are normally added with their corresponding base score, and existing migrated scores participate in the normalized weekly decay.

[0020] It should be understood that after completing the compliance review and determining the final review conclusion, the compliance judgment results under the current policy version will be accurately backfilled into the corresponding evidence fragments used in this Q&A. Based on the compliance performance of the evidence fragments under the current version of the rules, the fragment-level responsibility score will be calculated and dynamically updated for each fragment, forming a risk score that is strongly bound to the policy version and can be quantified. When the financial compliance rules change and a new version of the policy is implemented, the system will automatically freeze the responsibility scores of all evidence fragments under the old policy version, keeping the old version scores unmodifiable and unoverwriteable for historical audits and retrospection. At the same time, based on the new version of the compliance rules system, the old version responsibility scores will be migrated and attenuated, and the accumulation and updating of responsibility scores will be restarted under the new version of the rules, so that the risk score of the evidence fragments is consistent with the currently effective compliance rules, and the evidence risk status evolves synchronously with the rule version.

[0021] It should be noted that the specific process for re-accumulating and migrating the responsibility score in the new version is as follows: After completing the compliance review and obtaining the compliance determination conclusion, the compliance review conclusion is backfilled into the corresponding evidence fragment called in this Q&A based on the currently effective policy version number. Based on the compliance performance of the evidence fragment under the current version of the rules, the fragment-level responsibility score is calculated and dynamically updated in real time for each independent evidence fragment. The responsibility score is used to quantitatively characterize the degree of compliance risk of the evidence fragment under the corresponding policy version, forming a versioned risk score that is strongly bound to the policy version, ensuring that the same evidence fragment has independent risk identification and responsibility quantification results under different policy versions. In the process of calculating and updating the responsibility score, different score weights and update ranges are assigned according to whether the evidence fragment triggers blocking, whether it is manually reviewed as a violation, and whether it belongs to a low-confidence risk sample, so that the responsibility score can truly reflect the actual compliance status of the evidence fragment. If the financial compliance rules change and a new policy version is implemented, the responsibility score of all evidence fragments under the old policy version will be automatically frozen to keep the old version data immutable and unoverwritten, so as to meet the requirements of audit traceability and historical verification. Based on the new version of the rules, the responsibility scores of the old version will be migrated and attenuated, and the accumulation and updating of responsibility scores will be restarted under the new version to ensure that the responsibility scores are consistent with the currently effective compliance rules. Through a collaborative mechanism of version binding, freezing, migration attenuation, and re-accumulation, the risk status of evidence fragments can evolve synchronously with the version of compliance rules.

[0022] In this embodiment, after completing the compliance review and obtaining the final compliance determination, the compliance review conclusion is accurately backfilled into the corresponding evidence fragment called in this Q&A based on the currently effective policy version number. Using the actual compliance performance of the evidence fragment under the current version's rules as the calculation benchmark, a responsibility score weighted iterative algorithm is employed to calculate and dynamically update the fragment-level responsibility score for each independent evidence fragment in real time. This responsibility score is used to quantify the compliance risk level of the evidence fragment under the corresponding policy version, forming a versioned risk score strongly bound to the policy version. This ensures that the same evidence fragment has independent and non-interfering risk identification and responsibility quantification results under different policy versions. During the responsibility score calculation and update process, differentiated score weights and update magnitudes are assigned based on whether the evidence fragment triggers output blocking, whether it has been manually reviewed and determined to be in violation, and whether it belongs to a low-confidence risk sample. This system ensures that the responsibility score accurately reflects the actual compliance status of evidence fragments. When financial compliance rules change and a new policy version is implemented, the system automatically locks and freezes the responsibility scores of all evidence fragments under the old policy version, keeping the old version data immutable and unalterable to meet the audit traceability and historical verification requirements in financial scenarios. Simultaneously, based on the new version rule system, an inter-version score migration attenuation algorithm is used to attenuate and smooth the migration of old version responsibility scores. Under the new version rules, the real-time accumulation, calculation, and dynamic updating of responsibility scores are restarted, ensuring that the risk score of evidence fragments remains highly consistent with the currently effective compliance rules. Through the collaborative mechanism of version binding, score freezing, migration attenuation, and re-accumulation, the risk status of evidence fragments evolves synchronously with the version of compliance rules, improving the consistency, accuracy, and auditability of retrieval and compliance judgment.

[0023] Step 4: In scenarios involving blocked output and low-confidence review, the original question and answer, retrieved evidence fragments, and model output content are automatically recorded. Based on the large model, risk trigger points and evidence fragments are aligned to generate triple training data.

[0024] It should be understood that when the system triggers the blocking of question and answer output or enters a low-confidence manual review scenario, it automatically collects and saves full-link data such as the user's original questions and answers, retrieved evidence fragments, model output content, compliance judgment results, confidence scores, risk tags, and the current effective policy version number. Then, it uses a large model to perform semantic alignment and correlation positioning of risk content and evidence fragments, accurately extracts text fragments that cause compliance risks, clarifies the correspondence between risk trigger points and corresponding evidence, and finally automatically generates triple training data consisting of violation fragments, evidence chunks, and corresponding rules according to a preset format, providing standardized samples for the iterative fine-tuning of the compliance review model.

[0025] It should be noted that the specific process for generating triplet training data is as follows: If the question-and-answer output is blocked and the low confidence level is manually reviewed, the training sample collection process will be started immediately to automatically capture and save the original data of the entire chain in the current business scenario. The original data of the entire chain includes the user's original question-and-answer input, all evidence fragments obtained through the retrieval enhancement generation method, the answer content to be output generated by the large language model, the judgment result output by the compliance review model, the confidence score, the risk label and the version number of the current effective policy, to ensure that the sample data is complete, traceable and reproducible. Based on the collected raw data from the entire chain, the risk triggering location, violation semantics, and evidence source are analyzed and aligned through a large language model. The specific text fragments that cause compliance risks are located and extracted, and the citation relationship, supporting relationship, and causal relationship between the risk triggering point and the corresponding evidence fragment are clarified, thus completing a one-to-one match between the risk point and the evidence block.

[0026] Generating triplet training data also includes: Training data consisting of triplets of violation fragments, evidence, and corresponding rules is automatically generated based on a preset structured format. The violation fragments are used to identify specific content in the model output that has compliance risks. The evidence is used to identify the retrieval source fragment that caused the violation. The corresponding rules are used to identify the compliance rule item triggered this time and the policy version number to which it belongs. Based on low-confidence unblocked samples, compliance is reviewed using a large language model combined with the current effective policy version. Once confirmed as genuine non-compliant samples, they are then included in the triple generation process. Triple generation is performed directly based on blocked samples; The generated triplet data is associated with version information, scenario identifiers, and risk types and stored to form a high-quality incremental training sample set for financial compliance review.

[0027] In this embodiment, when the system triggers a question-and-answer output blocking or low-confidence manual review scenario, the training sample collection process is immediately initiated. A full-link data capture algorithm automatically captures and encrypts the original data of the current business scenario. This original data comprehensively covers the user's original question-and-answer input, all evidence fragments obtained through retrieval enhancement generation, the output answer content generated by the large language model, the judgment result output by the compliance review model, confidence scores, risk labels, and the current effective policy version number. This ensures the integrity, traceability, and reproducibility of the sample data, providing reliable data support for subsequent model fine-tuning and problem tracing. Based on the collected original data, a large language model combined with a risk semantic alignment algorithm and an evidence association matching algorithm is used to deeply analyze and accurately align the risk trigger location, violation semantics, and evidence source. Through semantic similarity calculation and causal reasoning, the specific text fragments that trigger compliance risks are located and extracted. The citation, supporting, and causal relationships between the risk trigger point and the corresponding evidence fragment are clarified, achieving a one-to-one accurate match between the risk point and the evidence block, avoiding evidence... The system addresses training data failure caused by mismatch between data and risk points. It then automatically generates triple training data in a pre-defined structured format, consisting of violation fragments, evidence, and corresponding rules. Violation fragments precisely identify specific content in the model output that poses compliance risks; evidence clearly identifies the retrieval source fragment that triggered the violation; and corresponding rules precisely identify the triggered compliance rule entry and its policy version number. For low-confidence, unblocked samples, a second compliance review is performed using a large language model combined with the current effective policy version. Only samples confirmed as genuine non-compliant by a violation probability threshold judgment algorithm are included in the triple generation process, and invalid samples are removed. For blocked samples, the triple generation process is executed directly, improving sample generation efficiency. Finally, the generated triple data is associated with version information, scenario identifiers, and risk types for storage. High-quality samples are selected using data deduplication and quality verification algorithms, forming a high-quality incremental training sample set for financial compliance review. This provides standardized, highly adaptable training data support for rapid fine-tuning and version upgrades of the compliance review model, significantly improving model iteration efficiency and compliance judgment accuracy.

[0028] Step 5: Extract the associated rule IDs based on the rule differences, filter the corresponding training data, and quickly fine-tune the Adapter of the basic model. After the fine-tuning meets the requirements, bind the new Adapter with the new version of the policy and automatically go online. Remove the corresponding Adapter when the rule is rolled back, so that the review model can be automatically upgraded with the version of the compliance rule.

[0029] It should be understood that, based on the structured rule difference information formed after the compliance rules are updated, the scope and content of rule adjustments are accurately analyzed, and the unique identifiers of related rules affected by this rule revision are automatically extracted. Using these related rule IDs, highly matching dedicated training data is precisely selected from the standardized triplet training sample set, and irrelevant and redundant samples are eliminated to ensure that the training data highly conforms to the requirements of the new regulations. A basic large model with fixed parameters is used as the underlying capability foundation, keeping the core weights of the basic model unchanged. Targeted and efficient incremental fine-tuning training is conducted only for the lightweight adapter, shortening the training cycle and reducing computational power consumption. Key indicators during the training process are continuously monitored until the model loss indicators... Once core parameters such as compliance judgment accuracy reach the preset passing standards, the fine-tuning training is deemed successful. The lightweight adapter is then bound to the new policy version number, and the adapter module adapted to the new regulations is automatically deployed and launched, enabling the system to load the new compliance judgment logic in real time. When a version rollback is required due to compliance rule revocation, system rollback, or policy adjustment, the system automatically unbinds the corresponding adapter, takes it offline, and removes the calling permissions of the adapter module, quickly restoring the review capabilities to a historically stable version. This allows the compliance review model to follow the iterative updates of financial compliance rules, enabling one-click deployment and flexible rollback, completing a fully automated version upgrade and adaptation process.

[0030] It should be noted that the specific process for implementing automatic upgrades of the review model along with the compliance rule version is as follows: Once the new financial compliance rules are released and a new version policy version number and structured rule difference information are generated, the affected related rule IDs are automatically extracted based on the rule difference information. Based on the related rule IDs, incremental training data of violation fragments, evidence, and corresponding rules that are highly related to this rule change are selected from the pre-constructed triplet training dataset. Based on a pre-set basic model as the foundation, with the weights of the basic model remaining unchanged, the lightweight adapter module of the model is rapidly and incrementally fine-tuned to enable the adapter module to efficiently learn the judgment logic of the new version of compliance rules. During the fine-tuning process, the training loss value is monitored in real time. When the training loss value drops to the preset threshold and the accuracy meets the upper limit standard, the fine-tuning is deemed to have met the standard. The trained adapter is uniquely bound to the current new policy version number to establish a correspondence between policy version, adapter and model version. The bound adapter is then automatically published to the online inference environment to enable the updated compliance review capabilities under the new policy version.

[0031] To enable the review model to automatically upgrade with the version of compliance rules, it also includes: If any compliance rule needs to be rolled back due to failure to pass the trial, withdrawal by regulators, or adjustment of the system, the binding relationship between the old policy version and the corresponding adapter will be automatically released, and the loading and calling of the adapter in the online environment will be removed, restoring the compliance review model to the historical stable version. Through lightweight adapter fine-tuning, version binding, automatic release, and rule rollback removal mechanisms, the compliance review model can be dynamically and adaptively upgraded with policy versions, ensuring that review judgments are always consistent with the currently effective rules.

[0032] In this embodiment, after the new financial compliance rules are officially released and the system generates the new policy version number and structured rule difference information, based on the rule difference information, a rule correlation analysis algorithm is used to automatically extract the related rule IDs affected by this rule change. Then, through a training data precision screening algorithm, incremental training data highly related to the rule change, including violation fragments, evidence, and corresponding rules, are selected from a pre-constructed triplet training dataset according to the extracted related rule IDs. Redundant and irrelevant samples are removed to ensure the adaptability of the training data to the new version of the rules, reducing the computational power consumption and training cycle of fine-tuning. A preset basic large model is used as the underlying capability foundation to strictly maintain... While maintaining the core weights of the basic model, rapid incremental fine-tuning is performed only on the lightweight adapter module. Through a parameter-oriented iterative algorithm, the adapter module efficiently learns the judgment logic, risk thresholds, and scope of application of the new compliance rules, achieving rapid adaptation to the new regulations. During fine-tuning, a real-time training monitoring algorithm continuously monitors key indicators such as training loss and compliance judgment accuracy. When the training loss drops to a preset threshold and the judgment accuracy meets a preset launch standard, the fine-tuning is automatically deemed successful, and the process terminates. Subsequently, the trained adapter is uniquely bound to the current new policy version number, establishing a policy version-... The system establishes a one-to-one correspondence between the adapter and the model version, automatically deploying the bound adapter to the online inference environment to ensure the system enables updated compliance review capabilities under the new policy version, achieving real-time compliance determination after the new regulations are implemented. If any compliance rule needs to be rolled back due to trial failure, regulatory withdrawal, or system adjustments, the system automatically triggers a version rollback mechanism, unbinding the old policy version from the corresponding adapter and removing the adapter's loading and calling permissions from the online environment. This quickly restores the compliance review model to a historically stable version, ensuring continuity of compliance determination in rule rollback scenarios. Through a collaborative mechanism of lightweight adapter fine-tuning, unique version binding, automatic deployment, and rule rollback removal, combined with the synergistic effect of multiple algorithms, the compliance review model achieves dynamic adaptive upgrades and flexible rollbacks with policy versions. This ensures that the model's compliance determination logic remains highly consistent with the currently effective financial compliance rules, significantly improving model iteration efficiency, reducing manual intervention costs, and meeting the adaptation needs of frequent changes in financial compliance rules.

[0033] Figure 2This is a schematic diagram of the structure of a financial intelligent question-and-answer compliance version management and automatic upgrade system provided by an embodiment of the present invention. The system includes: a compliance rule version management module, a compliance review module, an evidence fragment responsibility score management module, a compliance training data automatic generation module, and a review model automatic fine-tuning and upgrade module. The compliance rule version management module is used to build a compliance rule library, assigning an independent policy version number to each update of financial compliance rules, and binding and storing each policy version with the review model version, effective date, and rule difference information to form a one-to-one correspondence between rules, models, and versions. The compliance review module uses user question-and-answer input as a query to obtain a set of evidence fragments, inputs the search results into a preset compliance review model, and outputs compliance judgment results, confidence levels, and risk labels. High-risk content is directly blocked from output, and low-confidence results trigger manual review. The evidence fragment responsibility score... The management module is used to backfill compliance review conclusions to the corresponding evidence fragments, calculate and update fragment-level responsibility scores for each evidence fragment, forming version-bound risk scores. When compliance rules change, the old version's responsibility scores are frozen, and the responsibility scores are re-accumulated and migrated and decayed under the new version. The automatic compliance training data generation module is used to automatically record the original Q&A, retrieve evidence fragments, and model output content in scenarios of blocked output and low-confidence review. Based on the large model, risk trigger points and evidence fragments are aligned to generate triple training data. The automatic fine-tuning and upgrading module of the review model is used to extract the associated rule IDs based on rule differences, filter the corresponding training data, and quickly fine-tune the adapter of the basic model. After the fine-tuning is successful, the new adapter is bound to the new version policy and automatically launched. When the rule is rolled back, the corresponding adapter is removed, so that the review model can be automatically upgraded with the compliance rule version.

[0034] In this embodiment, it consists of five functional modules working together: compliance rule version management module, compliance review module, evidence fragment responsibility score management module, compliance training data automatic generation module, and review model automatic fine-tuning and upgrade module. Each module has a clear division of labor and works together to complete the full-process compliance control and dynamic model iteration. The compliance rule version management module is responsible for building and maintaining the overall rule system. By establishing a standardized financial compliance rule library, it centrally collects various regulatory rules, internal control systems, and compliance review standards. Whenever financial compliance rules are added, revised, or repealed, a unique policy version number is independently assigned. Simultaneously, it collects and organizes rule differences, legally effective dates, and matching review model versions, linking and persistently storing multiple key information categories. This establishes a mapping relationship between compliance rules, review models, and policy versions, laying the foundation for collaborative version management across the entire system. The compliance review module targets front-end user Q&A interactions, using real-time user-inputted Q&A text as search criteria. It retrieves relevant business evidence fragments, integrates them into an evidence set, and then uniformly inputs them into a preset compliance review model for intelligent judgment. It accurately outputs compliance judgment conclusions, result confidence values, and standardized risk labels. For extremely high-risk content, it enforces output blocking. For low-confidence results with insufficient confidence or ambiguous judgment boundaries, it automatically pushes them to the compliance review port to initiate a manual review process, ensuring compliance is handled layer by layer. The system controls the compliance baseline for Q&A output; the evidence fragment responsibility score management module is responsible for the quantitative control of evidence risk. After a single compliance review process, the final review conclusion is accurately backfilled into the corresponding called evidence fragment. Combining dimensions such as violation type and risk level, the module quantifies and dynamically updates the fragment-level responsibility score, generating a unique risk score deeply bound to the policy version. When the compliance rules are iterated and updated, all historical responsibility scores under the old version are automatically locked and frozen to ensure audit traceability. At the same time, historical scores are migrated based on reasonable attenuation logic, and risk score accumulation statistics are re-conducted in the new rule version, realizing that evidence risk assessment is dynamically updated with the rules; the compliance training data automatic generation module focuses on the autonomous construction of incremental samples. In specific scenarios where Q&A content is blocked from output or enters low-confidence manual review, the module automatically collects and retains the user's original Q&A content, retrieves and matches the evidence fragments, model-generated text, and various discrimination parameters. Relying on the powerful semantic understanding capabilities of the large model, the module completes the accurate alignment and matching of risk trigger points and related evidence fragments, and autonomously and structurally generates standardized triplet training data, reserving high-quality sample resources for model optimization;The automatic fine-tuning and upgrade module for the review model is responsible for lightweight iteration and version adaptation of the model. It locates associated rule identifiers based on differences arising from rule changes, accurately selects matching specialized training data, and, while retaining the core weights of the basic large model, performs rapid incremental fine-tuning training only on the lightweight adapter. Once the model's discrimination accuracy and loss metrics meet preset criteria, the newly trained adapter is bound to the new policy version and automatically deployed online. If subsequent policy withdrawals or rule rollbacks are required, the corresponding adapter is promptly removed offline to quickly restore the historical model's capabilities. Ultimately, this enables the compliance review model to automatically upgrade and flexibly roll back in line with the iterative pace of financial compliance rules, continuously ensuring the timeliness and accuracy of compliance judgment logic.

Claims

1. A method for compliant version management and automatic upgrade of financial intelligent question-and-answer systems, characterized in that, The method includes: Step 1: Build a compliance rule library, assign an independent policy version number to each update of financial compliance rules, and bind and store each policy version with the review model version, effective time, and rule difference information to form a one-to-one correspondence between rules, models, and versions. Step 2: Using user question and answer input as the query, obtain a set of evidence fragments, input the search results into the preset compliance review model, and output the compliance judgment result, confidence level and risk label. High-risk content is directly blocked from output, and low-confidence results are triggered for manual review. Step 3: Fill the corresponding evidence fragments with the compliance review conclusions, calculate and update the fragment-level responsibility score for each evidence fragment, and form a version-bound risk score. When the compliance rules change, freeze the old version responsibility score and re-accumulate and migrate the responsibility score under the new version. Step 4: In the scenarios of blocking output and low-confidence review, the original question and answer, retrieved evidence fragments, and model output content are automatically recorded. Based on the large model, risk trigger points and evidence fragments are aligned to generate triple training data. Step 5: Extract the associated rule IDs based on the rule differences, filter the corresponding training data, and quickly fine-tune the Adapter of the basic model. After the fine-tuning meets the requirements, bind the new Adapter with the new version of the policy and automatically go online. Remove the corresponding Adapter when the rule is rolled back, so that the review model can be automatically upgraded with the version of the compliance rule.

2. The financial intelligent question and answer compliance version management and automatic upgrade method of claim 1, wherein, The specific process for building the compliance rule base is as follows: A financial compliance rule base is pre-built and maintained, which is used to store all compliance requirements, regulatory rules, risk control standards and audit criteria in the financial field. If financial compliance rules are added, revised, or repealed, a unique and independent policy version number will be assigned to this rule change; Based on the comparison and analysis of the new and old rules, structured rule difference information including newly added entries, deleted entries, modified content and handling methods is generated; A one-to-one binding relationship is established between the policy version number and the corresponding compliance review model version, rule effective time, and structured rule difference information. The overall associated data of version number, model version, effective time, rule difference information, and associated rule identifier is persistently stored in the preset rule management database, thereby forming a version mapping system in which compliance rules, review models, and policy versions correspond one-to-one and are coordinated and consistent.

3. The financial intelligent question and answer compliance version management and automatic upgrade method of claim 1, wherein, The specific process for obtaining a set of evidence fragments by using user question-and-answer input as the query is as follows: The user-inputted question and answer text is used as a search query. Through search enhancement, relevant evidence fragments are matched and retrieved from the private domain data of financial institutions to form a set of evidence fragments corresponding to the current question and answer. The financial institution's internal private data includes internal rules and regulations, product manuals, official announcements, risk control implementation guidelines, business training materials, and regulatory responses. The acquired set of evidence fragments, along with user question and answer input, is input into a pre-trained compliance review model. Based on the compliance review model, the compliance, timeliness, scope of application, completeness of expression, and sensitive information of the evidence fragments are judged item by item. The model outputs the compliance judgment result, confidence score, and risk type label for the corresponding evidence fragments and question and answer content. The risk type label includes expired clauses, investment advice-like statements, missing disclaimers, unpublished data, prohibited statements, and mismatch of scope of application.

4. The financial intelligent question and answer compliance version management and automatic upgrade method of claim 3, wherein, The method of obtaining a set of evidence fragments by using user question-and-answer input as a query also includes: If the compliance determination result is non-compliant and the risk level reaches the preset blocking threshold, the output of the question and answer will be blocked directly, no content will be returned to the user, and the blocking event, corresponding evidence fragments and triggering rule version will be recorded simultaneously. If the compliance determination result is compliant but the model output confidence score is lower than the preset threshold, it is determined that the model is unable to accurately identify risks, and the manual review process is automatically triggered. The user's Q&A input, evidence fragment set, model output content, confidence score and risk label are pushed to the compliance review end, where the reviewer confirms the authenticity and compliance of the content and provides feedback on the review conclusion. By archiving blocking records, low-confidence samples, manual review results, and corresponding evidence fragments in a unified manner, the risk of non-compliant content being exported is reduced.

5. The financial intelligent question and answer compliance version management and automatic upgrade method of claim 1, wherein, The specific process for re-accumulating and migrating attenuation of responsibility scores in the new version is as follows: After completing the compliance review and obtaining the compliance determination conclusion, the compliance review conclusion is backfilled into the corresponding evidence fragment called in this Q&A based on the currently effective policy version number. Based on the compliance performance of the evidence fragment under the current version of the rules, the fragment-level responsibility score is calculated and dynamically updated in real time for each independent evidence fragment. The responsibility score is used to quantitatively characterize the degree of compliance risk of the evidence fragment under the corresponding policy version, forming a versioned risk score that is strongly bound to the policy version, ensuring that the same evidence fragment has independent risk identification and responsibility quantification results under different policy versions. In the process of calculating and updating the responsibility score, different score weights and update ranges are assigned according to whether the evidence fragment triggers blocking, whether it is manually reviewed as a violation, and whether it belongs to a low-confidence risk sample, so that the responsibility score can truly reflect the actual compliance status of the evidence fragment. If the financial compliance rules change and a new policy version is implemented, the responsibility score of all evidence fragments under the old policy version will be automatically frozen to keep the old version data immutable and unoverwritten, so as to meet the requirements of audit traceability and historical verification. Based on the new version of the rules, the responsibility scores of the old version will be migrated and attenuated, and the accumulation and updating of responsibility scores will be restarted under the new version to ensure that the responsibility scores are consistent with the currently effective compliance rules. Through a collaborative mechanism of version binding, freezing, migration attenuation, and re-accumulation, the risk status of evidence fragments can evolve synchronously with the version of compliance rules.

6. The financial intelligent question-and-answer compliance version management and automatic upgrade method as described in claim 1, characterized in that, The specific process for generating triplet training data is as follows: If the question-and-answer output is blocked and the low confidence level is manually reviewed, the training sample collection process will be started immediately to automatically capture and save the original data of the entire chain in the current business scenario. The original data of the entire chain includes the user's original question-and-answer input, all evidence fragments obtained through the retrieval enhancement generation method, the answer content to be output generated by the large language model, the judgment result output by the compliance review model, the confidence score, the risk label and the version number of the current effective policy, to ensure that the sample data is complete, traceable and reproducible. Based on the collected raw data from the entire chain, the risk triggering location, violation semantics, and evidence source are analyzed and aligned through a large language model. The specific text fragments that cause compliance risks are located and extracted, and the citation relationship, supporting relationship, and causal relationship between the risk triggering point and the corresponding evidence fragment are clarified, thus completing a one-to-one match between the risk point and the evidence block.

7. The financial intelligent question and answer compliance version management and automatic upgrade method of claim 6, wherein, The generated triplet training data also includes: Training data consisting of triplets of violation fragments, evidence, and corresponding rules is automatically generated based on a preset structured format. The violation fragments are used to identify specific content in the model output that has compliance risks. The evidence is used to identify the retrieval source fragment that caused the violation. The corresponding rules are used to identify the compliance rule item triggered this time and the policy version number to which it belongs. Based on low-confidence unblocked samples, compliance is reviewed using a large language model combined with the current effective policy version. Once confirmed as genuine non-compliant samples, they are then included in the triple generation process. Triple generation is performed directly based on blocked samples; The generated triplet data is associated with version information, scenario identifiers, and risk types and stored to form a high-quality incremental training sample set for financial compliance review.

8. The financial intelligent question answering, compliance version management and automatic upgrade method of claim 1, wherein, The specific process for implementing automatic upgrades of the review model along with the version of compliance rules is as follows: Once the new financial compliance rules are released and a new version policy version number and structured rule difference information are generated, the affected related rule IDs are automatically extracted based on the rule difference information. Based on the related rule IDs, incremental training data of violation fragments, evidence, and corresponding rules that are highly related to this rule change are selected from the pre-constructed triplet training dataset. Based on a pre-set basic model as the foundation, with the weights of the basic model remaining unchanged, the lightweight adapter module of the model is rapidly and incrementally fine-tuned to enable the adapter module to efficiently learn the judgment logic of the new version of compliance rules. During the fine-tuning process, the training loss value is monitored in real time. When the training loss value drops to the preset threshold and the accuracy meets the upper limit standard, the fine-tuning is deemed to have met the standard. The trained adapter is uniquely bound to the current new policy version number to establish a correspondence between policy version, adapter and model version. The bound adapter is then automatically published to the online inference environment to enable the updated compliance review capabilities under the new policy version.

9. The financial intelligent question and answer compliance version management and automatic upgrade method of claim 8, wherein, The implementation of the review model to automatically upgrade with the version of compliance rules also includes: If any compliance rule needs to be rolled back due to failure to pass the trial, withdrawal by regulators, or adjustment of the system, the binding relationship between the old policy version and the corresponding adapter will be automatically released, and the loading and calling of the adapter in the online environment will be removed, restoring the compliance review model to the historical stable version. Through lightweight adapter fine-tuning, version binding, automatic release, and rule rollback removal mechanisms, the compliance review model can be dynamically and adaptively upgraded with policy versions, ensuring that review judgments are always consistent with the currently effective rules.

10. A financial intelligent question and answer compliance version management and automatic upgrade system, the financial intelligent question and answer compliance version management and automatic upgrade system is used to implement the financial intelligent question and answer compliance version management and automatic upgrade method of any one of claims 1-9, characterized in that, The system includes: a compliance rule version management module, a compliance review module, an evidence fragment responsibility score management module, a compliance training data automatic generation module, and a review model automatic fine-tuning and upgrade module. The compliance rule versioning management module is used to build a compliance rule library, assign an independent policy version number to each update of financial compliance rules, and bind and store each policy version with the review model version, effective time, and rule difference information to form a one-to-one correspondence between rules, models, and versions. The compliance review module is used to obtain a set of evidence fragments by taking user question and answer input as the query, input the search results into a preset compliance review model, and output compliance judgment results, confidence level and risk label. High-risk content is directly blocked from output, and low-confidence results are triggered for manual review. The evidence fragment responsibility score management module is used to backfill compliance review conclusions to the corresponding evidence fragments, calculate and update fragment-level responsibility scores for each evidence fragment, form version-bound risk scores, freeze the old version responsibility scores when compliance rules change, and re-accumulate and migrate responsibility scores in the new version. The automatic compliance training data generation module is used to automatically record the original questions and answers, retrieve evidence fragments, and model output content in scenarios of blocked output and low-confidence review. Based on the large model, risk trigger points and evidence fragments are aligned to generate triple training data. The automatic fine-tuning and upgrade module of the review model is used to extract the associated rule ID based on rule differences, filter the corresponding training data, quickly fine-tune the adapter of the basic model, and after the fine-tuning meets the requirements, bind the new adapter with the new version of the policy and automatically go online. When the rule is rolled back, the corresponding adapter is removed, so as to realize the automatic upgrade of the review model with the version of the compliance rule.

Images