A chat knowledge base credibility management method based on source anchoring and health check

Abstract

The application relates to the technical field of knowledge base scanning and health examination, and provides a chat knowledge base credibility management method based on source anchoring and health examination, which comprises the following steps: in response to a trigger event, analyzing a knowledge page of a chat knowledge base directory to be scanned, and performing multi-dimensional health problem examination on the knowledge page according to the obtained structured information; summarizing all detected health problems, calculating the health score of the chat knowledge base according to the severity level of each health problem, and generating a health report containing the health score and detailed information of each health problem; in the chat reply process, when hit knowledge is searched, the health report is loaded, the health level of the hit knowledge is marked according to the matching result of the path of the hit knowledge and the problem path in the health report; and corresponding downgrade constraint text is added in the reply generation prompt according to the health level, so as to control the output intensity of the reply generation model. The application can overcome the technical defect that knowledge management is disconnected from reply generation.

Description

Technical Field

[0001] This invention relates to the field of knowledge base scanning and health check technology, and in particular to a method for trust governance of chat knowledge bases based on source anchoring and health check. Background Technology

[0002] Enterprise desktop chat assistants typically need to maintain a local or team knowledge base to answer customer inquiries, internal collaboration questions, after-sales processes, invoicing information, pricing boundaries, and common operating instructions. After receiving a chat question, the system retrieves relevant information from the knowledge base and then provides the matched content to the model to generate a response. The quality of the knowledge base directly affects the reliability of the final response.

[0003] In practical use, the knowledge base will continue to grow. Original materials will be compiled into source pages, unconfirmed questions, standard responses, topic pages, and candidate responses; manually confirmed chat replies may also be accumulated as candidate knowledge. Over time, the knowledge base is prone to the following problems:

[0004] 1. Standard responses lack source paths; while the content can be retrieved, the original data or approval basis cannot be traced. 2. Standard responses lack evidence; the page has a conclusion but no evidence section, and the model may write unproven conclusions as established facts. 3. Long-term backlog of unresolved issues; a large number of unprocessed issues will be mixed into the knowledge base maintenance process, causing similar questions to repeatedly fail to obtain reliable answers. 4. Long-term unreviewed response candidates; when the candidate area continues to accumulate, genuine and frequently used responses cannot enter the formal knowledge area in a timely manner, which may also lead to misjudgment by maintenance personnel. 5. Conflicts exist between pages with the same name or topic; when multiple pages give different conclusions on time, price, process, contact person, or commitment boundaries, ordinary search algorithms may only select one based on similarity without indicating the risk of conflict. 6. Source pages become orphan pages; source materials exist but are not associated with standard responses or unresolved issues, and cannot be cited for a long time, resulting in the sinking of knowledge assets. 7. The generation chain lacks health status constraints; even if the knowledge is hit but the source is missing, conflicting, or an unresolved issue exists, the response generation may still output a definitive conclusion.

[0005] Chinese patent CN110046235A discloses a knowledge base evaluation method. This method calculates the health score of the knowledge base based on multi-dimensional indicators and their weighted scores, such as the text features of knowledge points (title length, sensitivity, semantic similarity), usage features (output rate, click rate, positive reviews, and conversion to human intervention), and knowledge coverage features (historical question coverage). Based on the evaluation results, the method recommends candidate knowledge points for covered issues to users or reminds them to correct low-quality knowledge points. This method only evaluates the quality of the knowledge base from superficial dimensions such as the content attributes of the knowledge points themselves (text length and sensitivity) and usage statistics (click-through rate and positive review rate). It does not pay attention to deeper knowledge governance issues such as the traceability of knowledge sources (whether the standard response has a source path and evidence section), knowledge structure issues (whether there is a backlog of questions to be confirmed and candidate responses, and whether the titles are duplicated), factual conflicts between knowledge (whether key fields such as time and price are consistent under the same topic), and the integrity of source citation (whether the source page is effectively associated). Moreover, its evaluation results are only used for offline knowledge point addition and correction recommendations, and it fails to connect with the online response generation process of the knowledge retrieval augmentation generation (RAG) system. It cannot perform real-time degradation constraints on the model output based on the health level of the hit knowledge when users ask and answer questions.

[0006] Therefore, how to provide a credibility governance method for chat knowledge bases, enabling the system not only to retrieve knowledge, but also to periodically or on-demand scan the knowledge base, identify problems such as missing sources, missing evidence, conflicts, backlogs, and orphan sources, generate health scores and problem lists, and mark health results on retrieval hits, and further control the strength of model expression before response generation, has become an urgent technical problem to be solved. Summary of the Invention

[0007] In view of this, in order to overcome the shortcomings of the prior art, the present invention aims to provide a chat knowledge base credibility governance method based on source anchoring and health check.

[0008] This invention provides a method for trust governance of chat knowledge bases based on source anchoring and health checks, the method comprising:

[0009] Step S1: In response to the triggering event, parse the knowledge pages of the chat knowledge base directory to be scanned, and perform a multi-dimensional health problem check on the knowledge pages based on the structured information obtained from the parsing.

[0010] Step S2: Summarize all detected health issues, calculate the health score of the chat knowledge base based on the severity level of each health issue, and generate a health report containing the health score and detailed information of each health issue;

[0011] Step S3: In the chat reply process, after the retrieved knowledge is obtained, the health report is loaded, and the health level is marked for the knowledge based on the matching results of the path of the knowledge and the question path in the health report.

[0012] Step S4: Add corresponding degradation constraint text to the response generation prompt based on the health level to control the output strength of the response generation model.

[0013] Optionally, in the chat knowledge base credibility governance method based on source anchoring and health check of the present invention, in step S1, the knowledge page includes the original data page, source page, standard reply page, question page to be confirmed, topic page, and candidate page to be reviewed. The structured information of the knowledge page is obtained by parsing in the following way: taking the root directory of the chat knowledge base as the scanning starting point, all readable text files under the root directory are traversed. Each file is classified into the corresponding page type according to the relative path of each file under the root directory and the name of its subdirectory. During the traversal, the pre-configured running status directory is skipped to avoid misidentifying non-knowledge files as knowledge pages. For each readable knowledge page, its title, body, and file path are extracted. For the standard reply page, the source path field and evidence segment content are extracted in addition. For knowledge pages that cannot be read or fail to be parsed, they are marked as unknown and corresponding unknown status question records are generated.

[0014] Optionally, in the chat knowledge base credibility governance method based on source anchoring and health checks of the present invention, step S1 involves performing a multi-dimensional health issue check on the knowledge page based on the structured information obtained through parsing, including performing source anchoring checks and evidence integrity checks on the standard response page in the following manner:

[0015] Read the metadata or header fields of the standard response page to determine whether there are non-empty source path fields used to trace the original data, candidate responses, or manual review results;

[0016] Read the evidence section area of ​​the standard response page and determine whether the evidence section area contains non-empty text content or referenceable file fragments;

[0017] When the source path field is missing or empty, and / or when the evidence section area is missing or empty, generate a health issue with a high severity level and mark the corresponding standard response page as unusable as strong evidence.

[0018] Optionally, in the chat knowledge base credibility governance method based on source anchoring and health checks of the present invention, step S1 involves performing a multi-dimensional health issue check on the knowledge page based on the structured information obtained through parsing, including performing a backlog check on issues to be confirmed and candidates to be reviewed in the following manner:

[0019] The number of pages with pending issues is counted. These pages contain questions that require manual confirmation but have not yet generated a formal standard response. When the number of pages with pending issues exceeds a preset first threshold, a backlog of pending issues is generated as a health issue with a severity level of "medium".

[0020] The number of candidate pages awaiting review is counted. These candidate pages contain candidate knowledge points extracted from historical chat replies but have not yet been reviewed and confirmed. When the number of candidate pages awaiting review exceeds a preset second threshold, a backlog of candidate health issues with a low severity level is generated.

[0021] Optionally, in the chat knowledge base credibility governance method based on source anchoring and health checks of the present invention, step S1 involves performing a multi-dimensional health issue check on the knowledge page based on the structured information obtained through parsing, including performing conflict detection on duplicate titles and factual conflicts on the same topic in the following manner:

[0022] The titles of each knowledge page are normalized. This normalization process includes at least removing leading and trailing whitespace characters, compressing consecutive whitespace characters, unifying capitalization, and removing characters that differ from the preset format. Knowledge pages with identical normalized strings are grouped into the same title group. When the same title group contains more than two knowledge pages, a duplicate title health issue with a severity level of medium is generated.

[0023] For knowledge page groups with the same standardized title or belonging to the same topic, use regular expressions or field positioning rules to extract fact fields from the body of each knowledge page. The fact fields include at least one of time, amount, process steps, contact person, and commitment boundary. Cross-compare the values ​​of the same type of fact field in different knowledge pages. When the comparison results are inconsistent, generate a content conflict health issue with a high severity level.

[0024] Optionally, in the chat knowledge base credibility governance method based on source anchoring and health checks of the present invention, step S1 involves performing a multi-dimensional health issue check on the knowledge page based on the structured information obtained from parsing. This includes performing an isolated reference check on orphan source pages that are not referenced in the following manner: extracting the unique identifier or path information of each source page; traversing the source path fields of all standard reply pages and all unconfirmed question pages, and determining whether the source path fields contain the unique identifier or path information of the source page; if no source path field of any standard reply page or unconfirmed question page points to the source page, marking the source page as an isolated reference and generating an orphan source page health issue with a low severity level.

[0025] Optionally, in the chat knowledge base credibility governance method based on source anchoring and health check of the present invention, in step S2, the health score of the chat knowledge base is calculated as follows: using the full score as the initial score, the initial score is deducted item by item according to the severity level of each health problem. Specifically, for each health problem with a high severity level, a first point is deducted and a first deduction limit is set; for each health problem with a medium severity level, a second point is deducted and a second deduction limit is set; and for each health problem with a low severity level, a third point is deducted and a third deduction limit is set. The first point is greater than the second point, the second point is greater than the third point, and the first, second, and third deduction limits decrease sequentially. The final health score after deduction is limited to between a preset minimum score and the full score.

[0026] Optionally, in the chat knowledge base credibility governance method based on source anchoring and health check of the present invention, in step S2, the health report includes both a structured health report and a human-readable health report; the structured health report is written into the status directory of the chat knowledge base in JSON format or an equivalent structured data format, and each health problem record includes at least a problem type field, a severity level field, a problem path field, a problem title field, a problem description field, and a suggested action field, and the problem path field uses a path representation consistent with the chat knowledge base file path; the human-readable health report is written into a designated directory of the chat knowledge base in markup language format, for providing manual viewing and export functions in the user interface.

[0027] Optionally, in the chat knowledge base credibility governance method based on source anchoring and health check of the present invention, in step S3, the health level of the hit knowledge is marked according to the matching result of the path of the hit knowledge and the problem path in the health report in the following manner: the path of the hit knowledge and the problem path of the health problem are respectively processed by path separator unification, case normalization and relative path to absolute path processing to obtain the normalized hit path string and problem path string; it is determined whether the hit path string and the problem path string satisfy any one of the following relationships: complete consistency, inclusion, or being included; if they are satisfied, the match is determined to be successful; if the match is successful, the health level of the corresponding hit knowledge is determined according to the severity level of the matched health problem, wherein, when a health problem with a high severity level is matched, it is marked as low health level; when a health problem with a medium severity level is matched, it is marked as pending confirmation; when a health problem with a low severity level is matched, it is marked as weak reference; if the hit knowledge does not match any health problem, it is marked as strong evidence; when the health report does not exist, has expired, or the hit path cannot be associated with any problem path, the health level of the corresponding hit knowledge is marked as unevaluated.

[0028] Optionally, in the chat knowledge base credibility governance method based on source anchoring and health check of the present invention, in step S4, the corresponding downgrade constraint text is added to the reply generation prompt according to the health level in the following manner:

[0029] When the health level of any hit knowledge is low, add a first control instruction to the pre-instruction segment of the response generation prompt to prohibit the knowledge content from being expressed as a confirmed fact and guide the model to output verification suggestions or follow-up questions.

[0030] When the health level of any hit knowledge is pending confirmation, add a second control instruction to the pre-instruction segment of the response generation prompt, requiring the model to output a confirmatory question to the user before giving a conclusion.

[0031] When the health level of any hit knowledge is a weak reference, add a third control instruction to the pre-instruction segment of the reply generation prompt to prohibit the model from promising specific time values, monetary values, process step sequences and contact information in the reply.

[0032] When the first, second, and third control commands are triggered simultaneously, they are merged in the order of priority: low health takes precedence over pending confirmation, and pending confirmation takes precedence over weak reference. Furthermore, the degradation constraint text and the independent response content risk interception strategy work together in the response generation process.

[0033] This invention provides a chat knowledge base credibility governance method based on source anchoring and health checks, which has the following beneficial technical effects:

[0034] By integrating knowledge base scanning, health issue identification, health report generation, health-related labeling, and response degradation constraints into a complete closed loop from offline governance to online constraints, the knowledge base quality evaluation results can be directly applied to the chat response generation process. This establishes a credibility control dimension independent of similarity matching in the retrieval enhancement generation system, effectively overcoming the technical deficiency of traditional solutions where knowledge governance and response generation are disconnected. Attached Figure Description

[0035] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart illustrating a chat knowledge base credibility governance method based on source anchoring and health checks according to an exemplary embodiment of the present invention.

[0037] Figure 2This is a schematic diagram illustrating the process of parsing and obtaining structured information using a chat knowledge base credibility governance method based on source anchoring and health checks according to an exemplary embodiment of the present invention.

[0038] Figure 3 This is a flowchart illustrating the process of performing source anchoring checks and evidence integrity checks in a chat knowledge base credibility governance method based on source anchoring and health checks according to an exemplary embodiment of the present invention.

[0039] Figure 4 A flowchart illustrating the process of performing backlog checks for a chat knowledge base credibility governance method based on source anchoring and health checks according to an exemplary embodiment of the present invention;

[0040] Figure 5 A flowchart illustrating the process of performing conflict detection for a chat knowledge base credibility governance method based on source anchoring and health checks according to an exemplary embodiment of the present invention;

[0041] Figure 6 A flowchart illustrating the process of performing isolated reference checks for a chat knowledge base credibility governance method based on source anchoring and health checks according to an exemplary embodiment of the present invention.

[0042] Figure 7 A flowchart illustrating the process of calculating the health score of a chat knowledge base according to an exemplary embodiment of the present invention, based on source anchoring and health checks.

[0043] Figure 8 This is a flowchart illustrating the process of labeling the health level of hit knowledge in a chat knowledge base credibility governance method based on source anchoring and health checks according to an exemplary embodiment of the present invention.

[0044] Figure 9 This is a flowchart illustrating the process of adding corresponding degradation constraint text to the reply generation prompt in the chat knowledge base credibility governance method based on source anchoring and health check according to an exemplary embodiment of the present invention. Detailed Implementation

[0045] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0046] It should be noted that, in the absence of conflict, the following embodiments and features can be combined with each other; and, based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0047] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0048] An exemplary embodiment of the present invention provides a method for trust governance of chat knowledge bases based on source anchoring and health checks. Figure 1 This is a flowchart illustrating a chat knowledge base credibility governance method based on source anchoring and health checks according to an exemplary embodiment of the present invention, as follows: Figure 1 As shown, in this embodiment, the method of the present invention is implemented in the following manner:

[0049] Step S1: In response to the triggering event, parse the knowledge pages of the chat knowledge base directory to be scanned, and perform a multi-dimensional health problem check on the knowledge pages based on the structured information obtained from the parsing.

[0050] In this embodiment, the knowledge page includes the original data page, the source page, the standard response page, the question page to be confirmed, the topic page, and the candidate page to be reviewed.

[0051] Figure 2 This is a schematic diagram illustrating the process of parsing and obtaining structured information using a chat knowledge base credibility governance method based on source anchoring and health checks according to an exemplary embodiment of the present invention. Figure 2 As shown, in this embodiment, the structured information of knowledge pages is obtained by parsing in the following way: Starting from the root directory of the chat knowledge base, all readable text files under the root directory are traversed. Each file is classified into its corresponding page type based on its relative path in the root directory and the name of its subdirectory. During the traversal, pre-configured running status directories are skipped to avoid misidentifying non-knowledge files as knowledge pages. For each readable knowledge page, its title, body, and file path are extracted. For standard reply pages, the source path field and evidence segment content are additionally extracted. For knowledge pages that cannot be read or fail to be parsed, they are marked as unknown and a corresponding unknown status issue record is generated. This embodiment achieves automated parsing of heterogeneous knowledge base structures and accurate classification of page types, providing a structured data foundation for subsequent differentiated health checks for different page types and reducing the workload of manually configuring page types.

[0052] In this embodiment, the multi-dimensional health issue check includes: performing source anchoring checks and evidence integrity checks on standard response pages; performing backlog checks on issues to be confirmed and candidates to be reviewed; performing conflict detection on duplicate titles and factual conflicts on the same topic; and performing isolated reference checks on orphan source pages that are not cited.

[0053] Figure 3 This is a flowchart illustrating the process of performing source anchoring checks and evidence integrity checks in a chat knowledge base credibility governance method based on source anchoring and health checks according to an exemplary embodiment of the present invention. Figure 3 As shown, in this embodiment, source anchoring checks and evidence integrity checks are performed on the standard response page in the following manner:

[0054] Read the metadata or header fields of the standard response page to determine whether there are non-empty source path fields used to trace the original data, candidate responses, or manual review results;

[0055] Read the evidence section area of ​​the standard response page and determine whether the evidence section area contains non-empty text content or referenceable file fragments;

[0056] When the source path field is missing or empty, and / or when the evidence section area is missing or empty, generate a health issue with a high severity level and mark the corresponding standard response page as unusable as strong evidence.

[0057] This embodiment establishes a credibility-based evaluation standard of "traceable source and replayable evidence" by performing a dual check on the source path field and evidence section content of the standard response page, and generating a high-severity health problem when either check fails. The source path solves the problem of tracing the origin of the conclusion, and the evidence section solves the problem of proving what the conclusion is based on. Together, they constitute the necessary conditions for the standard response to be cited as strong evidence, effectively preventing the model from outputting standard responses without evidence or with false evidence as certain facts.

[0058] Figure 4 A flowchart illustrating the backlog check process for a chat knowledge base credibility governance method based on source anchoring and health checks according to an exemplary embodiment of the present invention is shown below. Figure 4 As shown, in this embodiment, backlog checks are performed on issues to be confirmed and candidates to be reviewed in the following manner:

[0059] The number of pages with pending issues is counted. These pages contain questions that require manual confirmation but have not yet generated a formal standard response. When the number of pages with pending issues exceeds a preset first threshold, a backlog of pending issues is generated as a health issue with a severity level of "medium".

[0060] The number of candidate pages awaiting review is counted. These candidate pages contain candidate knowledge points extracted from historical chat replies but have not yet been reviewed and confirmed. When the number of candidate pages awaiting review exceeds a preset second threshold, a backlog of candidate health issues with a low severity level is generated.

[0061] This embodiment sets independent quantity thresholds for pages of questions awaiting confirmation and pages of candidate questions awaiting review, and generates health issues with differentiated severity levels. It includes intermediate knowledge growth states in the knowledge base that have not yet entered the formal standard response system within the scope of health evaluation, so that maintenance personnel can receive timely reminders when the backlog reaches the threshold. This prevents the knowledge growth chain from being blocked and the real high-frequency responses from being delayed in entering the database due to the long-term accumulation of questions awaiting confirmation and candidate responses.

[0062] Figure 5 A flowchart illustrating the conflict detection process for a chat knowledge base credibility governance method based on source anchoring and health checks according to an exemplary embodiment of the present invention is shown below. Figure 5 As shown, in this embodiment, conflict detection is performed on duplicate headings and factual conflicts on the same topic in the following manner:

[0063] The titles of each knowledge page are normalized. This normalization process includes at least removing leading and trailing whitespace characters, compressing consecutive whitespace characters, unifying capitalization, and removing characters that differ from the preset format. Knowledge pages with identical normalized strings are grouped into the same title group. When the same title group contains more than two knowledge pages, a duplicate title health issue with a severity level of medium is generated.

[0064] For knowledge page groups with the same standardized title or belonging to the same topic, use regular expressions or field positioning rules to extract fact fields from the body of each knowledge page. The fact fields include at least one of time, amount, process steps, contact person, and commitment boundary. Cross-compare the values ​​of the same type of fact field in different knowledge pages. When the comparison results are inconsistent, generate a content conflict health issue with a high severity level.

[0065] This embodiment achieves automatic detection of redundant and conflicting content in the knowledge base by performing normalization processing on knowledge page titles and identifying duplicate titles, as well as performing regular expression extraction and cross-comparison on key fact fields in knowledge pages with the same topic. Duplicate title detection solves the ambiguity problem of retrieval hits caused by pages with the same name, and fact conflict detection on the same topic ensures that inconsistencies in sensitive information such as prices and processes no longer rely solely on accidental human discovery, effectively reducing the risk of self-contradictory responses from the model output due to hitting conflicting knowledge.

[0066] Figure 6A flowchart illustrating the process of performing orphan reference checks in a chat knowledge base credibility governance method based on source anchoring and health checks according to an exemplary embodiment of the present invention is shown below. Figure 6 As shown, in this embodiment, orphan source pages that are not referenced are checked for isolated references in the following manner: the unique identifier or path information of each source page is extracted; the source path fields of all standard response pages and all pending issue pages are traversed to determine whether the source path fields contain the unique identifier or path information of the source page; if no source path field of any standard response page or pending issue page points to the source page, the source page is marked as an isolated reference and a low-severity orphan source page health issue is generated.

[0067] This embodiment automatically identifies orphan source pages that are not referenced by any page by extracting the unique identifier of each source page and traversing the source path field of all standard response pages and pages of questions to be confirmed. This allows original data that was originally buried in the knowledge base and could not be retrieved or used to be discovered and marked. Based on this, maintenance personnel can decide to supplement standard responses, improve reference associations, or archive expired data, thereby avoiding the source data being in an unreachable or unusable state for a long time.

[0068] Step S2: Summarize all detected health issues, calculate the health score of the chat knowledge base based on the severity level of each health issue, and generate a health report containing the health score and detailed information of each health issue.

[0069] Figure 7 This is a flowchart illustrating the process of calculating the health score of a chat knowledge base according to an exemplary embodiment of the present invention, which is based on source anchoring and health checks for trustworthiness governance. Figure 7 As shown, in this embodiment, the health score of the chat knowledge base is calculated as follows: The initial score is the maximum score. Each health issue is deducted progressively based on its severity level. For health issues with a high severity level, a first point is deducted and a first deduction limit is set. For health issues with a medium severity level, a second point is deducted and a second deduction limit is set. For health issues with a low severity level, a third point is deducted and a third deduction limit is set. The first point is greater than the second point, the second point is greater than the third point, and the first, second, and third deduction limits decrease sequentially. The final health score after deduction is limited to between a preset minimum score and the maximum score.

[0070] For example, the following point deduction rules can be adopted:

[0071] 1. 12 points will be deducted for each high-severity issue, with a maximum deduction of 48 points;

[0072] 2. 6 points will be deducted for each moderate to severe issue, with a maximum deduction of 30 points;

[0073] 3. 2 points will be deducted for each low-severity issue, with a maximum deduction of 16 points.

[0074] This embodiment sets progressively decreasing deduction scores and upper limits for health issues at three severity levels: high, medium, and low. This ensures that the final health score reflects the number of serious defects in the knowledge base while preventing a single low-risk issue from excessively lowering the overall score. This scoring rule guarantees that the health score, as a reliable metric for the knowledge base, has reasonable sensitivity and stability, making it easy for maintenance personnel to judge the dynamic evolution of the knowledge base quality by observing the trend of health score changes.

[0075] In practical applications, the health report in this embodiment includes both a structured health report and a human-readable health report. The structured health report is written to the status directory of the chat knowledge base in JSON format or an equivalent structured data format. Each health issue record includes at least an issue type field, a severity level field, an issue path field, an issue title field, an issue description field, and a suggested action field. The issue path field uses a path representation consistent with the chat knowledge base file path. The human-readable health report is written to a designated directory of the chat knowledge base in markup language format, which is used to provide manual viewing and export functions in the user interface.

[0076] Step S3: In the chat reply process, after the retrieved knowledge is obtained, the health report is loaded, and the health level is marked for the knowledge based on the matching results between the path of the knowledge and the problem path in the health report.

[0077] Figure 8 This is a flowchart illustrating the process of labeling the health level of hit knowledge in a chat knowledge base credibility governance method based on source anchoring and health checks according to an exemplary embodiment of the present invention. Figure 8As shown, in this embodiment, the health level of the hit knowledge is marked according to the matching results between the path of the hit knowledge and the question path in the health report in the following manner: The path of the hit knowledge and the question path of the health problem are respectively processed by path separator unification, case normalization, and relative path to absolute path processing to obtain the normalized hit path string and question path string; it is determined whether the hit path string and the question path string satisfy any one of the following relationships: complete consistency, inclusion, or being included. If they are satisfied, the match is considered successful; if the match is successful, the health level of the corresponding hit knowledge is determined according to the severity level of the matched health problem. Specifically, when a health problem with a high severity level is matched, it is marked as low health; when a health problem with a medium severity level is matched, it is marked as pending confirmation; when a health problem with a low severity level is matched, it is marked as weak reference; if the hit knowledge does not match any health problem, it is marked as strong evidence; when the health report does not exist, has expired, or the hit path cannot be associated with any question path, the health level of the corresponding hit knowledge is marked as unassessed.

[0078] This embodiment performs a three-tiered matching judgment based on complete consistency, inclusion, or contained relationship after normalizing the hit path and the problem path. It also accurately labels the hit knowledge as strong evidence, weak reference, pending confirmation, or low health according to the severity level of the matched health problem. At the same time, it separately labels cases where the report does not exist, has expired, or the path cannot be associated as unevaluated. This makes the health labeling process adaptable to the path differences under different operating systems and storage structures, ensuring the accuracy and reliability of the labeling results, and providing fine-grained health status input for response degradation.

[0079] Step S4: Add corresponding degradation constraint text to the response generation prompt based on the health level to control the output strength of the response generation model.

[0080] Figure 9 This is a flowchart illustrating the process of adding corresponding degradation constraint text to the reply generation prompt in the chat knowledge base credibility governance method based on source anchoring and health check according to an exemplary embodiment of the present invention. Figure 9 As shown, in this embodiment, the corresponding downgrade constraint text is added to the response generation prompt according to the health level in the following manner:

[0081] When the health level of any hit knowledge is low, add a first control instruction to the pre-instruction segment of the response generation prompt to prohibit the knowledge content from being expressed as a confirmed fact and guide the model to output verification suggestions or follow-up questions.

[0082] When the health level of any hit knowledge is pending confirmation, add a second control instruction to the pre-instruction segment of the response generation prompt, requiring the model to output a confirmatory question to the user before giving a conclusion.

[0083] When the health level of any hit knowledge is a weak reference, add a third control instruction to the pre-instruction segment of the reply generation prompt to prohibit the model from promising specific time values, monetary values, process step sequences and contact information in the reply.

[0084] When the first, second, and third control commands are triggered simultaneously, they are merged in the order of priority: low health takes precedence over pending confirmation, and pending confirmation takes precedence over weak reference. Furthermore, the degradation constraint text and the independent response content risk interception strategy work together in the response generation process.

[0085] This embodiment generates differentiated control instructions based on three health levels—low health, pending confirmation, and weak reference—to prohibit the output of confirmed facts, prioritize follow-up confirmation, and avoid committing specific values ​​and processes. When multiple instructions are triggered simultaneously, they are merged according to the priority order of low health over pending confirmation, and pending confirmation over weak reference, to achieve precise hierarchical control of the output strength of the generated model. This degradation constraint is independent of and complements the response content risk interception strategy, ensuring the credibility of the knowledge source while taking into account the security of the response content, effectively reducing the business risk of the model making incorrect commitments due to low-quality knowledge.

[0086] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0087] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for trust governance of chat knowledge bases based on source anchoring and health checks, characterized in that, The method includes: Step S1: In response to the triggering event, parse the knowledge pages of the chat knowledge base directory to be scanned, and perform a multi-dimensional health problem check on the knowledge pages based on the structured information obtained from the parsing. Step S2: Summarize all detected health issues, calculate the health score of the chat knowledge base based on the severity level of each health issue, and generate a health report containing the health score and detailed information of each health issue; Step S3: In the chat reply process, after the retrieved knowledge is obtained, the health report is loaded, and the health level is marked for the knowledge based on the matching results of the path of the knowledge and the question path in the health report. Step S4: Add corresponding degradation constraint text to the response generation prompt based on the health level to control the output strength of the response generation model.

2. The chat knowledge base credibility governance method based on source anchoring and health check according to claim 1, characterized in that, In step S1, the knowledge page includes the original data page, source page, standard reply page, question page to be confirmed, topic page, and candidate page to be reviewed. The structured information of the knowledge page is obtained by parsing in the following way: Starting from the root directory of the chat knowledge base, all readable text files in the root directory are traversed. Each file is classified into the corresponding page type according to its relative path in the root directory and the name of its subdirectory. During the traversal, the pre-configured running status directory is skipped to avoid misidentifying non-knowledge files as knowledge pages. For each readable knowledge page, its title, body, and file path are extracted. For the standard reply page, the source path field and evidence segment content are extracted in addition. For knowledge pages that cannot be read or fail to be parsed, they are marked as unknown and corresponding unknown status question records are generated.

3. The chat knowledge base credibility governance method based on source anchoring and health check according to claim 1, characterized in that, In step S1, a multi-dimensional health issue check is performed on the knowledge page based on the structured information obtained from the parsing, including source anchoring check and evidence integrity check on the standard response page in the following ways: Read the metadata or header fields of the standard response page to determine whether there are non-empty source path fields used to trace the original data, candidate responses, or manual review results; Read the evidence section area of ​​the standard response page and determine whether the evidence section area contains non-empty text content or referenceable file fragments; When the source path field is missing or empty, and / or when the evidence section area is missing or empty, generate a health problem with a high severity level and mark the corresponding standard response page as unusable as strong evidence.

4. The chat knowledge base credibility governance method based on source anchoring and health check according to claim 1, characterized in that, In step S1, a multi-dimensional health check is performed on the knowledge page based on the structured information obtained from the parsing, including backlog checks on issues awaiting confirmation and candidates awaiting review in the following manner: The number of pages with pending issues is counted. These pages contain questions that require manual confirmation but have not yet generated a formal standard response. When the number of pages with pending issues exceeds a preset first threshold, a backlog of pending issues is generated as a health issue with a severity level of "medium". The number of candidate pages awaiting review is counted. These candidate pages contain candidate knowledge points extracted from historical chat replies but have not yet been reviewed and confirmed. When the number of candidate pages awaiting review exceeds a preset second threshold, a backlog of candidate health issues with a low severity level is generated.

5. The chat knowledge base credibility governance method based on source anchoring and health check according to claim 1, characterized in that, In step S1, a multi-dimensional health check is performed on the knowledge page based on the structured information obtained from the parsing, including conflict detection for duplicate titles and factual conflicts on the same topic, in the following manner: The titles of each knowledge page are normalized. This normalization process includes at least removing leading and trailing whitespace characters, compressing consecutive whitespace characters, unifying capitalization, and removing characters that differ from the preset format. Knowledge pages with identical normalized strings are grouped into the same title group. When the same title group contains more than two knowledge pages, a duplicate title health issue with a severity level of medium is generated. For knowledge page groups with the same standardized title or belonging to the same topic, use regular expressions or field positioning rules to extract fact fields from the body of each knowledge page. The fact fields include at least one of time, amount, process steps, contact person, and commitment boundary. Cross-compare the values ​​of the same type of fact field in different knowledge pages. When the comparison results are inconsistent, generate a content conflict health issue with a high severity level.

6. The chat knowledge base credibility governance method based on source anchoring and health check according to claim 1, characterized in that, In step S1, a multi-dimensional health issue check is performed on the knowledge page based on the structured information obtained from parsing. This includes performing an orphan reference check on unreferenced orphan source pages in the following manner: extracting the unique identifier or path information of each source page; traversing the source path fields of all standard response pages and all pending issue pages to determine whether the source path field contains the unique identifier or path information of the source page; if no source path field of any standard response page or pending issue page points to the source page, the source page is marked as an orphan reference and an orphan source page health issue with a low severity level is generated.

7. The chat knowledge base credibility governance method based on source anchoring and health check according to claim 1, characterized in that, In step S2, the health score of the chat knowledge base is calculated as follows: using the full score as the initial score, the initial score is deducted item by item according to the severity level of each health issue. For each health issue with a high severity level, the first point is deducted and a first deduction limit is set; for each health issue with a medium severity level, the second point is deducted and a second deduction limit is set; and for each health issue with a low severity level, the third point is deducted and a third deduction limit is set. The first point is greater than the second point, the second point is greater than the third point, and the first, second, and third deduction limits decrease sequentially. The final health score after deduction is limited to between the preset minimum score and the full score.

8. The chat knowledge base credibility governance method based on source anchoring and health check according to claim 1, characterized in that, In step S2, the health report includes both a structured health report and a human-readable health report. The structured health report is written to the status directory of the chat knowledge base in JSON format or an equivalent structured data format. Each health issue record includes at least an issue type field, a severity level field, an issue path field, an issue title field, an issue description field, and a suggested action field. The issue path field uses the same path notation as the chat knowledge base file path. The human-readable health report is written to a designated directory of the chat knowledge base in markup language format, which is used to provide manual viewing and export functions in the user interface.

9. The chat knowledge base credibility governance method based on source anchoring and health check according to claim 1, characterized in that, In step S3, the health level of the hit knowledge is labeled according to the matching results between the path of the hit knowledge and the question path in the health report as follows: The path of the hit knowledge and the question path of the health problem are processed by unifying the path separator, normalizing case, and converting relative paths to absolute paths to obtain normalized hit path strings and question path strings; it is determined whether the hit path string and the question path string satisfy any one of the following relationships: complete consistency, inclusion, or being included. If they satisfy any of these, the match is considered successful; if the match is successful, the health level of the corresponding hit knowledge is determined according to the severity level of the matched health problem. Specifically, a high-severity health problem is labeled as low health, a medium-severity health problem is labeled as pending confirmation, and a low-severity health problem is labeled as weak reference. If the hit knowledge does not match any health problem, it is labeled as strong evidence; when the health report does not exist, has expired, or the hit path cannot be associated with any question path, the health level of the corresponding hit knowledge is labeled as unassessed.

10. The chat knowledge base credibility governance method based on source anchoring and health check according to claim 1, characterized in that, In step S4, add the corresponding downgrade constraint text to the response generation prompt according to the health level as follows: When the health level of any hit knowledge is low, add a first control instruction to the pre-instruction segment of the response generation prompt to prohibit the knowledge content from being expressed as a confirmed fact and guide the model to output verification suggestions or follow-up questions. When the health level of any hit knowledge is pending confirmation, add a second control instruction to the pre-instruction segment of the response generation prompt, requiring the model to output a confirmatory question to the user before giving a conclusion. When the health level of any hit knowledge is a weak reference, add a third control instruction to the pre-instruction segment of the reply generation prompt to prohibit the model from promising specific time values, monetary values, process step sequences and contact information in the reply. When the first, second, and third control commands are triggered simultaneously, they are merged in the order of priority: low health takes precedence over pending confirmation, and pending confirmation takes precedence over weak reference. Furthermore, the degradation constraint text and the independent response content risk interception strategy work together in the response generation process.

Images