An Incremental Self-Healing Update Method and System for Insurance Knowledge Base
By identifying stable anchor points and minimum reconstruction intervals in the insurance knowledge base, only locally changed knowledge units are re-blocked and vectorized, and duplicate knowledge is cleaned up and conflicts are handled. This solves the problem of duplicate processing and conflicts caused by local modifications to the insurance knowledge base, and improves update efficiency and retrieval stability.
Patent Information
- Application Number
- CN202611562967.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-10
- Publication Date
- 2026-11-17
AI Technical Summary
When making partial modifications to documents, existing insurance knowledge bases struggle to accurately define the scope of knowledge reconstruction. This leads to the repeated segmentation and vectorization of unchanged knowledge, increasing computational load. Furthermore, the update process introduces duplicate knowledge and conflicts between old and new knowledge, affecting the accuracy and stability of knowledge retrieval.
By obtaining the location anchor points of the document to be integrated and historical documents, stable anchor points and minimum reconstruction intervals are determined. Only knowledge units within the range of change are re-blocked and vectorized. Duplicate knowledge is cleaned up and conflicts are handled in combination with block-level lineage, and candidate incremental indexes are constructed for replay verification.
This reduces the amount of data involved in repetitive vector calculations and index updates, improves the efficiency of insurance knowledge base updates and the stability of retrieval, avoids merging semantically similar but different business conditions, and ensures the accuracy of knowledge updates.
Smart Images

Figure CN123115169A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of knowledge base data processing technology, specifically to an incremental self-healing update method and system for an insurance knowledge base. Background Technology
[0002] With the application of large language models in scenarios such as insurance intelligent question answering, business assistance, and internal knowledge retrieval, an insurance knowledge base is usually built by using retrieval-enhanced generation methods. This involves segmenting and vectorizing documents such as insurance terms, product descriptions, claims rules, and business systems, and then providing knowledge support for the large language model based on the vector retrieval results.
[0003] When existing insurance knowledge bases undergo version updates or partial content modifications to insurance documents, they typically re-segment and vectorize the modified documents. Since changes in partial document content can alter the boundaries of subsequent knowledge blocks, this can lead to the repeated processing of numerous knowledge blocks whose actual content remains unchanged, increasing the amount of data required for vector calculations and index updates. Furthermore, the inclusion of insurance knowledge from multiple versions and sources in the knowledge base can result in duplicated knowledge and conflicts between old and new information. Directly updating the existing knowledge base could negatively impact the accuracy and stability of subsequent knowledge retrieval results. Summary of the Invention
[0004] To address the challenges of accurately defining the scope of knowledge reconstruction in existing insurance knowledge bases when insurance documents undergo partial modifications, leading to repetitive segmentation and vectorization of unchanged knowledge, and the potential impact of duplicate knowledge, conflicts between old and new knowledge, and unverified knowledge updates on the accuracy and stability of knowledge retrieval during the update process, this invention provides an incremental self-healing update method for insurance knowledge bases. The method includes: Obtain the insurance document to be accessed, the historical insurance document corresponding to the insurance document to be accessed, and the current knowledge base; parse the insurance document to be accessed and the historical insurance document into knowledge unit sequences respectively; and obtain the location anchor point based on the knowledge unit sequence. The location anchor points are matched to obtain several stable anchor points, and the minimum reconstruction interval is determined based on the stable anchor points; The knowledge units within the minimum reconstruction interval are re-blocked and quantized. Knowledge blocks outside the minimum reconstruction interval that meet the inheritance conditions inherit the knowledge block identifier and vector index relationship of the original knowledge block to obtain the target knowledge block set and block-level lineage relationship. Based on the target knowledge block set and the existing knowledge base, perform duplicate knowledge cleanup and conflict knowledge handling to obtain a candidate update set; Based on the candidate update set, a candidate incremental index is constructed, and historical question and answer records associated with the candidate update set are retrieved for replay verification. When the replay verification meets the release conditions, the current version corresponding to the current knowledge base is updated to the target version containing the candidate incremental index.
[0005] By matching the location of knowledge units in the insurance documents to be accessed with those in historical insurance documents, the impact of document updates can be limited to the localized areas where changes have actually occurred. Only the knowledge within the smallest reconstructed interval is re-blocked and vectorized, while the unchanged knowledge blocks can continue to use the existing knowledge block identifiers and vector index relationships, thereby reducing large-scale redundant calculations caused by local document modifications. At the same time, before the official release, duplicate knowledge cleanup, conflict knowledge handling, and historical question and answer playback verification are further carried out, so that knowledge updates, knowledge governance, and online verification form a continuous processing chain, improving the efficiency of insurance knowledge base updates and the stability of knowledge retrieval after updates.
[0006] Furthermore, the specific steps for obtaining a stable anchor point include: Each knowledge unit in the knowledge unit sequence is normalized to obtain normalized content. A content fingerprint is obtained based on the normalized content, and a hierarchical path fingerprint is obtained based on the title hierarchical path corresponding to the knowledge unit. For any current knowledge unit in the insurance document to be accessed, obtain the historical knowledge unit in the historical insurance document that has the same content fingerprint as the current knowledge unit to obtain a first candidate knowledge unit; when there are at least two first candidate knowledge units, obtain the hierarchical path fingerprint, the content fingerprint of the preceding adjacent knowledge unit and the content fingerprint of the following adjacent knowledge unit for each first candidate knowledge unit, and match them with the corresponding information in the insurance document to be accessed to obtain a second candidate knowledge unit that uniquely corresponds to the current knowledge unit. Anchor pairs are obtained based on the current knowledge unit and the second candidate knowledge unit. The anchor pairs are sorted according to the order of the current knowledge unit in the insurance document to be accessed. Anchor pairs that have reverse order in the historical insurance document are deleted. Stable anchors and stable anchor sequences are obtained based on the remaining anchor pairs.
[0007] Based on content fingerprint matching, a secondary constraint is introduced by incorporating title-level paths and relationships between adjacent knowledge units. This can reduce the incorrect correspondence of knowledge units caused by repetitive expressions in insurance clauses. By deleting abnormal anchor points where the arrangement relationship is reversed in the old and new documents, stable anchor points with continuous positional correspondence can be obtained, providing a reliable positional benchmark for accurately determining the scope of document changes in the future.
[0008] Furthermore, the specific steps to obtain the minimum reconstruction interval include: Obtain the current knowledge unit that does not form a corresponding relationship between two adjacent stable anchor points, and obtain the initial change section; The initial change segment is extended to both sides to the block boundaries of existing knowledge blocks in the current knowledge base; when the block boundary is located inside a complete knowledge unit, it is extended to the outer boundary of the complete knowledge unit, and at least one unchanged knowledge unit located on both sides of the initial change segment is retained as the reconstruction context to obtain the minimum reconstruction interval.
[0009] Instead of directly using the location where text changes occur between old and new documents as the scope of re-blocking, the scope of change is modified by combining the existing knowledge block boundaries and the boundaries of complete knowledge units. This covers the actual changed content while retaining the necessary context, so that the re-formed knowledge blocks have more complete business semantics and can limit the continuous drift of knowledge block boundaries caused by local modifications, thereby reducing the scope of re-vectorized data.
[0010] Furthermore, the specific steps for obtaining the target knowledge block set and the block-level lineage relationship include: Obtain several original knowledge blocks in the existing knowledge base that correspond to the minimum reconstruction interval. Based on the positional relationship between each original knowledge block and the minimum reconstruction interval, divide the original knowledge blocks into knowledge blocks outside the interval, knowledge blocks with intersecting boundaries, and knowledge blocks inside the interval. For the knowledge block outside the interval, compare the content fingerprints of the current knowledge unit and the historical knowledge unit corresponding to the knowledge block outside the interval. When the content fingerprints of the two are consistent, mark the knowledge block outside the interval as an inherited knowledge block, keep the knowledge block identifier and vector index relationship of the inherited knowledge block unchanged, and obtain the inherited knowledge block set. For the knowledge blocks with intersecting boundaries, the knowledge blocks with intersecting boundaries are split into inheritance fragments and reconstruction fragments based on the knowledge unit boundaries; The reconstructed fragments and knowledge units within the interval are recombined and divided into blocks according to the original text arrangement to obtain several reconstructed knowledge blocks. Vectorization processing is performed on the reconstructed knowledge blocks to obtain a set of reconstructed knowledge blocks. The target knowledge block set is obtained based on the inherited knowledge block set and the reconstructed knowledge block set; Record the current insurance document identifier, current knowledge unit range, parent knowledge block identifier, and update batch identifier for each reconstructed knowledge block to obtain the block-level lineage relationship.
[0011] Instead of discarding and rebuilding the entire original knowledge block, knowledge blocks that cross the boundary of the minimum reconstruction interval are further broken down, allowing unchanged content outside the changed area to continue inheriting the original data. Only the parts actually affected by document changes are recombined and vectorized, further reducing the amount of unnecessary vector computation. Simultaneously, using block-level lineage records to document the correspondence between reconstructed knowledge blocks and insurance documents, knowledge units, and parent knowledge blocks allows for reverse identification of knowledge sources during subsequent knowledge conflict verification, version updates, and anomaly tracing.
[0012] Furthermore, the specific steps for performing deduplication cleanup include: Each target knowledge block in the target knowledge block set is subjected to text standardization processing, and an accurate content fingerprint is obtained based on the standardized text. Based on the precise content fingerprint, knowledge blocks with the same precise content fingerprint as any target knowledge block are obtained from the target knowledge block set and the existing knowledge base to obtain a precise duplicate knowledge block group. From the precisely repeated knowledge block group, determine the retained knowledge blocks, associate the current document identifier, current knowledge unit range and block-level lineage of the non-retained knowledge blocks in the precisely repeated knowledge block group with the retained knowledge blocks, and delete the repeated vector index relationship of the non-retained knowledge blocks; For candidate knowledge blocks that do not form the exact repeating knowledge block group, several candidate similar knowledge blocks are obtained based on the vector of the candidate knowledge block. Constraint data is extracted from the candidate knowledge block and the candidate similar knowledge block respectively to obtain a first constraint element set and a second constraint element set. The constraint data includes insurance object, insurance liability, applicable conditions, exclusion conditions and numerical constraints. When the constraint elements in the first constraint element set and the second constraint element set are consistent, and the semantic relevance between the candidate knowledge block and the candidate similar knowledge block meets the preset merging condition, the candidate knowledge block and the candidate similar knowledge block are merged, and a deduplicated knowledge block set is obtained based on the target knowledge block after merging. When the semantic relevance satisfies the preset merging condition, and the first constraint element set and the second constraint element set have at least one inconsistent corresponding constraint element, a conflicting candidate knowledge block pair is obtained based on the candidate knowledge block and the candidate similar knowledge block.
[0013] First, precise content fingerprints are used to clean up duplicate knowledge with completely identical text, avoiding duplicate vector storage while retaining mapping information from different knowledge sources. For knowledge with different textual expressions but similar semantics, the judgment is further combined with constraints such as objects, responsibilities, conditions and values in insurance business. This can avoid erroneously merging insurance knowledge with different waiting periods, insurance amounts or liability conditions based solely on semantic similarity, and can proactively identify potential differences between new and old knowledge as conflicting knowledge that needs to be dealt with later.
[0014] Furthermore, the specific steps for performing the conflict knowledge resolution include: For any of the conflict candidate knowledge block pairs, the current document identifier and current knowledge unit range corresponding to the two conflict candidate knowledge blocks are obtained based on the block-level lineage relationship, and the first source evidence fragment and the second source evidence fragment are obtained from the corresponding insurance document. Obtain the insurance product identifier, clause hierarchy path, document version identifier, effective time and expiration time corresponding to the first source evidence fragment and the second source evidence fragment, and obtain the first evidence information and the second evidence information respectively; Input the first source evidence fragment, the second source evidence fragment, the first evidence information, and the second evidence information into the large language model to obtain candidate conflict elements; Based on the candidate conflict elements, the original text is located in the first source evidence segment and the second source evidence segment respectively, the actual constraint content in the original text is obtained, and the consistency of the actual constraint content with the candidate conflict elements is verified to obtain the valid conflict elements. When two conflict candidate knowledge blocks have the same insurance product identifier and clause level path, and the effective time of the latter source evidence fragment of the effective conflict element is later than that of the former source evidence fragment, a replacement relationship between the predecessor knowledge block and the successor knowledge block is established based on the document version relationship of the two source evidence fragments. Based on the replacement relationship, the successor knowledge block is added to the candidate effective knowledge block set, and the predecessor knowledge block is added to the candidate ineffective knowledge block set. The candidate update set is obtained based on the deduplicated knowledge block set, the candidate effective knowledge block set, and the candidate invalid knowledge block set.
[0015] The large language model is only used to discover candidate conflicting elements that may have differences, rather than directly determining the effective status of knowledge based on the model output. After obtaining candidate conflicting elements, the original content of the insurance document is returned for verification, thereby reducing the risk that incorrect content generated by the model will directly affect the knowledge base. By combining the insurance product, clause location, version and effective time to establish predecessor and successor relationships, new and old knowledge of the same insurance matter can be replaced with evidence, avoiding conflicting new and old insurance knowledge from participating in the retrieval at the same time.
[0016] Furthermore, the method also includes: For predecessor and successor knowledge blocks that have the aforementioned replacement relationship, a knowledge lifecycle record is established. The knowledge lifecycle record includes at least the predecessor knowledge block identifier, the successor knowledge block identifier, the successor knowledge block effective time, and the predecessor knowledge block expiration time. A knowledge switching task is generated based on the effective time of the subsequent knowledge block. When the effective time of the successor knowledge block is reached, the current insurance document identifier and the current knowledge unit range are obtained based on the block-level lineage of the successor knowledge block, and the current verification content corresponding to the current knowledge unit range is obtained from the current insurance document. Obtain the knowledge block content corresponding to the subsequent knowledge block, and perform a consistency check between the knowledge block content and the current verification content; When the consistency check passes, the successor knowledge block is added to the set of effective knowledge blocks, and the predecessor knowledge block that has the replacement relationship with the successor knowledge block is added to the set of invalid knowledge blocks. When the consistency check fails, or the current check content is not obtained based on the block-level lineage, the latest insurance document corresponding to the current insurance document is obtained, the latest insurance document is updated to the insurance document to be accessed, and the stable anchor matching and minimum reconstruction interval determination are re-executed.
[0017] After insurance knowledge is pre-entered into the knowledge base, the system not only mechanically switches between old and new knowledge according to the scheduled effective time, but also verifies the corresponding insurance document content again before it officially takes effect. If the pre-generated subsequent knowledge is inconsistent with the current insurance document content, the system re-enters the partial update process, so that the knowledge lifecycle can be adaptively corrected as the insurance document changes, reducing the risk of incorrect deployment caused by changes in pre-generated knowledge before it officially takes effect.
[0018] Furthermore, the specific steps for constructing candidate incremental indexes include: Obtain the basic vector index and current version list of the existing knowledge base; Obtain knowledge blocks that have been added or whose content has changed relative to the current knowledge base from the candidate update set, obtain the knowledge blocks to be written, and establish an incremental vector index based on the vector of the knowledge blocks to be written. Obtain the predecessor knowledge blocks that have been replaced in the candidate update set, the original knowledge blocks that have been replaced by the reconstructed knowledge blocks due to re-segmentation, and other knowledge blocks that have stopped participating in knowledge retrieval, and generate an invalid knowledge block identifier table based on the corresponding knowledge block identifier; Obtain the vector index relationship of the inherited knowledge blocks in the target knowledge block set, so that the inherited knowledge blocks continue to be associated with the basic vector index; Establish an index mapping relationship between the knowledge block to be written in the incremental vector index and the corresponding current knowledge unit range, and establish a version association relationship between the knowledge block to be written and the corresponding parent knowledge block based on the block-level lineage relationship; Based on the basic vector index, the incremental vector index, the invalid knowledge block identifier table, and the version association relationship, a candidate version list is generated, and the candidate incremental index is obtained.
[0019] Candidate versions do not need to copy all vector data in the current knowledge base. Instead, they continue to use the basic vector indexes corresponding to unchanged knowledge, and only create incremental vector indexes for newly added or changed knowledge. They also use an invalid knowledge block identifier table to shield old knowledge that should no longer be involved in retrieval, thereby further reducing the amount of index reconstruction and storage usage during the knowledge base version update process. At the same time, the version source relationship of incremental knowledge is maintained through the knowledge unit scope and parent knowledge block association.
[0020] Furthermore, the specific steps of the playback verification include: Obtain the knowledge block identifiers corresponding to newly added knowledge blocks, reconstructed knowledge blocks, and knowledge blocks that have stopped participating in knowledge retrieval in the candidate update set, and obtain the set of knowledge blocks affected by the update. Retrieve historical question and answer records from the historical question and answer records that have called any knowledge block in the set of knowledge blocks whose updates have been affected, and obtain the directly affected question and answer records. Based on the block-level lineage of each knowledge block in the knowledge block set affected by the update, obtain the corresponding insurance document identifier and knowledge unit scope, and obtain the insurance product identifier and clause level path corresponding to the knowledge unit scope; From the historical question and answer records, obtain historical question and answer records that did not call the updated impact knowledge block set, but called other knowledge blocks with the same insurance product identifier or the same clause level path, and obtain related impact question and answer records; A replay question and answer set is obtained based on the directly impacting question and answer records and the related impacting question and answer records; Knowledge retrieval is performed again on the same historical questions in the replay question-and-answer set based on the current version and the candidate version, respectively, to obtain the current replay results and the candidate replay results; For the directly affected question and answer records, determine whether the candidate replay results exclude the corresponding invalid knowledge blocks, and determine whether the candidate replay results recall newly added knowledge blocks or reconstructed knowledge blocks associated with the corresponding historical questions; For the associated impact question and answer records, compare the current playback results with the first preset number of search knowledge blocks in the candidate playback results to obtain the change in search results; When the directly affecting question and answer records meet the exclusion requirements of the invalid knowledge block and the recall requirements of the newly added or reconstructed knowledge block, and the change in the retrieval results of the associated affecting question and answer records does not exceed the preset change range, it is determined that the replay verification meets the release conditions, and the version list corresponding to the knowledge base retrieval service is switched from the current version list to the candidate version list. When the newly added question and answer records after the version list switch is completed meet the preset rollback conditions, the version list corresponding to the knowledge base retrieval service will be switched from the candidate version list back to the current version list.
[0021] Replay testing does not randomly select test data from all historical Q&A. Instead, it targets and filters Q&A that directly or indirectly affect the knowledge blocks that have actually changed and their business relevance. This can improve the coverage of the actual impact of the knowledge update in replay testing. By simultaneously verifying the exclusion of invalid knowledge, the recall of new knowledge, and the changes in the search results of related questions, abnormal disturbances caused by partial updates to existing search results can be discovered before the official release of candidate knowledge versions. This allows for rapid release and rollback through version list switching.
[0022] 10. An incremental self-healing update system for an insurance knowledge base, characterized in that the system comprises: Document parsing unit: used to obtain the insurance document to be accessed, the historical insurance documents corresponding to the insurance document to be accessed, and the current knowledge base, and to parse the insurance document to be accessed and the historical insurance documents into knowledge unit sequences respectively, and to obtain the location anchor point based on the knowledge unit sequence; Reconstruction unit: used to match the location anchor points to obtain several stable anchor points, determine the minimum reconstruction interval based on the stable anchor points; re-block and quantize the knowledge units within the minimum reconstruction interval, and inherit the knowledge block identifier and vector index relationship of the original knowledge block for knowledge blocks outside the minimum reconstruction interval that meet the inheritance conditions, to obtain the target knowledge block set and block-level lineage relationship; Knowledge governance unit: used to perform duplicate knowledge cleanup and conflict knowledge handling based on the target knowledge block set and the existing knowledge base, and obtain a candidate update set; Feedback self-healing unit: used to construct a candidate incremental index based on the candidate update set, obtain historical question and answer records associated with the candidate update set for replay verification; when the replay verification meets the release conditions, update the current version corresponding to the current knowledge base to the target version containing the candidate incremental index. One or more technical solutions provided by this invention have at least the following technical effects or advantages: The principle and beneficial effects of this system are similar to those of this method, and will not be elaborated further on this system.
[0023] One or more technical solutions provided by this invention have at least the following technical effects or advantages: 1. By identifying reliable positional correspondences between old and new insurance documents through stable anchor points, and combining them with the existing knowledge block boundaries to determine the minimum reconstruction interval, the entire document does not need to be reprocessed when the insurance document is partially modified. For unchanged knowledge blocks, the original knowledge blocks and vector index relationships are inherited. Only the knowledge content that is actually affected by the modification is re-blocked and vectorized, thereby reducing the amount of repeated vector calculations and index update data.
[0024] 2. Two-level duplicate knowledge identification is performed through precise content fingerprints and semantic similarity relationships. Furthermore, the substantive differences between similar knowledge are verified by using insurance objects, insurance liabilities, applicable conditions, exclusion conditions, and numerical constraints, so that insurance knowledge with similar semantics but different business conditions will not be directly merged. For knowledge with differences, predecessor and successor relationships are established through source evidence and version information, thereby reducing the situation where conflicting new and old knowledge participate in the retrieval at the same time. Attached Figure Description
[0025] The accompanying drawings, which are provided to further illustrate embodiments of the invention and constitute a part of this invention, are not intended to limit the scope of the invention. Figure 1 This is the overall flowchart of the incremental self-healing update method for the insurance knowledge base; Figure 2 This is a flowchart of the method for stable anchor point matching and minimum reconstruction interval determination; Figure 3 This is a flowchart of methods for cleaning up duplicate knowledge and handling conflicting knowledge; Figure 4 This is a flowchart of the candidate incremental index construction, replay verification, and version rollback methods. Detailed Implementation
[0026] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, where there is no conflict, the embodiments of the present invention and the features thereof can be combined with each other.
[0027] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.
[0028] Example 1 refer to Figures 1-4This embodiment provides an incremental self-healing update method for an insurance knowledge base. Taking a critical illness insurance policy of an insurance company as an example, the current knowledge base is based on version V3, and the insurance company has obtained the insurance document V4 to be integrated. V4 does not rewrite all insurance terms, but only modifies the waiting period, partial exclusions, and one payment ratio.
[0029] Retrieve the insurance document V4 to be integrated, the historical insurance document V3 corresponding to the insurance document V4 to be integrated, and the current knowledge base.
[0030] The correspondence between the insurance document to be accessed and the historical insurance document can be determined by the insurance product identifier, document type, document number, or file metadata. For example, if both V3 and V4 have the product identifier P10086 and the document type "insurance clauses", and the version numbers are V3 and V4 respectively, then V3 is determined to be the historical insurance document corresponding to V4.
[0031] The knowledge unit sequence is obtained by parsing V3 and V4 respectively.
[0032] For example, V3 is formed in the following order: H1: Article 1 Contract Structure; H2: This insurance contract is composed of the insurance terms and conditions...; H3: Article 2 Scope of Coverage; ... H26: Article 5 Waiting Period; H27: The waiting period is 90 days from the effective date of this contract; H28: If an insured event occurs during the waiting period...; H29: Article 6 Insurance Liability.
[0033] V4 corresponds to the following positions: C1, C2...C26, C27, C28, C29.
[0034] The content of C27 was changed to: "A waiting period of 180 days shall be imposed from the effective date of this contract." Apart from this change, most of the content of the preceding and following knowledge units remains unchanged.
[0035] A location anchor is generated for each knowledge unit. In this embodiment, the location anchor may include at least the knowledge unit content fingerprint, the title level path fingerprint, and the knowledge unit's position in the document.
[0036] The title hierarchy path indicates the position of the knowledge unit in the insurance document directory structure, for example: A. Critical Illness Insurance Terms / Chapter 2 Insurance Liability / Article 5 Waiting Period.
[0037] The anchor points of the current knowledge unit and the historical knowledge unit are matched to obtain stable anchor points.
[0038] Specifically, the knowledge unit text is normalized to obtain normalized content.
[0039] For example: The waiting period is 90 days. The waiting period is 90 days. After space normalization, both can form the same text expression.
[0040] Content fingerprints are obtained based on normalized content, and hierarchical path fingerprints are obtained based on title-level paths.
[0041] For the current knowledge unit C26, if its content fingerprint is the same as that of the historical knowledge unit H26, then H26 can be used as a candidate knowledge unit.
[0042] If multiple knowledge units with the same content fingerprint exist in historical insurance documents, such as multiple chapters containing other unresolved matters to be handled in accordance with relevant regulations, then relying solely on the content fingerprint is insufficient to provide a unique match. Therefore, it is necessary to further compare the title hierarchy path, the preceding adjacent knowledge unit, and the following adjacent knowledge unit.
[0043] In this embodiment, if the content of the current knowledge unit C50 is the same as that of historical knowledge units H30 and H50, then the title-level path fingerprints of H30 and H50 are obtained respectively, and the content fingerprints of the preceding and following knowledge units of both are also obtained. If the title-level path, preceding adjacent knowledge unit, and following adjacent knowledge unit corresponding to C50 are all consistent with the information corresponding to H50, then H50 is determined as the second candidate knowledge unit uniquely corresponding to C50.
[0044] An anchor pair is formed by the current knowledge unit and its unique corresponding historical knowledge unit.
[0045] For example: C25-H25; C26-H26; C28-H28; C29-H29.
[0046] Sort the anchor pairs according to the current order of knowledge units in V4.
[0047] If the sequence C20-H20; C21-H35; C22-H22 appears, then H35 follows H22, but appears before H22 in the corresponding relationship, thus forming a reverse order. This situation may be caused by duplicate paragraphs in the document, so the anchor pair corresponding to C21-H35 should be deleted.
[0048] After reverse cleaning, a stable anchor point sequence is obtained.
[0049] Obtain the current knowledge unit that does not form a corresponding relationship between two adjacent stable anchor points, and obtain the initial change section; For example, stable anchor point matching revealed that C26 corresponds to H26; C28 corresponds to H28; while C27, located between C26 and C28, did not form a consistent correspondence with historical knowledge units. Therefore, C27 was identified as the initial change segment.
[0050] The initial change section represents the range of text changes obtained by directly comparing the old and new insurance documents, but this range may not necessarily be directly used for knowledge block updates.
[0051] This is because the boundaries of knowledge blocks and knowledge units in the current knowledge base may be different.
[0052] The initial change segment is extended to both sides to the block boundaries of existing knowledge blocks in the current knowledge base; when the block boundary is located inside a complete knowledge unit, it is extended to the outer boundary of the complete knowledge unit, and at least one unchanged knowledge unit located on both sides of the initial change segment is retained as the reconstruction context to obtain the minimum reconstruction interval.
[0053] For example, the current knowledge base contains: knowledge block K105: H25+H26+H27; knowledge block K106: H28+H29. Although only the position corresponding to H27 is actually changed, H27, H25, and H26 are all located in K105. Therefore, the initial change segment is extended to both sides to the block boundaries of the existing knowledge blocks in the current knowledge base.
[0054] Furthermore, if the boundary of a knowledge block happens to be located inside a complete knowledge unit, then the boundary is extended to the complete knowledge unit to avoid truncating a clause with complete insurance implications.
[0055] In this embodiment, at least one unchanged knowledge unit on each side of the initial changed section is also retained as a reconstruction context.
[0056] For example, if the actual knowledge unit of change is C27, then C26, C27, and C28 can be used together as the minimum reconstruction interval.
[0057] The reason for retaining C26 and C28 is that when regenerating knowledge blocks, it is necessary to ensure that the changed text and its context have complete semantics. Otherwise, if only C27 is vectorized independently, the title of the fifth waiting period and the context of subsequent clauses may be lost.
[0058] Therefore, the minimum reconstruction interval can be understood as: the smallest continuous range that covers the actual changed content and contains enough to re-form the boundary of a stable knowledge block.
[0059] The knowledge units within the minimum reconstruction interval are re-blocked and quantized. Knowledge blocks outside the minimum reconstruction interval that meet the inheritance conditions inherit the knowledge block identifier and vector index relationship of the original knowledge block to obtain the target knowledge block set and block-level lineage relationship. Specifically: Retrieve the original knowledge block in the current knowledge base that corresponds to the minimum reconstruction interval.
[0060] In actual implementation, the original knowledge blocks within the minimum reconstruction interval and the preset range on both sides can also be obtained to identify knowledge blocks outside the interval, knowledge blocks with intersecting boundaries, and knowledge blocks within the interval.
[0061] Among them: knowledge blocks outside the interval refer to knowledge blocks whose entire content is outside the minimum reconstruction interval; knowledge blocks inside the interval refer to knowledge blocks whose entire content is within the minimum reconstruction interval; and knowledge blocks with intersecting boundaries refer to knowledge blocks whose only part is within the minimum reconstruction interval, while the other part is outside the minimum reconstruction interval.
[0062] For example: K104 corresponds to H23-H24; K105 corresponds to H25-H27; K106 corresponds to H28-H30. Assume the minimum reconstruction interval corresponds to C26-C28. Then K104 can be considered as a knowledge block outside the interval; K105 and K106 intersect with the boundary of the minimum reconstruction interval.
[0063] For knowledge blocks outside the interval, find the current knowledge unit in the insurance document to be accessed based on its corresponding historical knowledge unit.
[0064] For example, K104 corresponds to H23 and H24, and the stable anchor point relationship indicates that H23 corresponds to C23 and H24 corresponds to C24.
[0065] Comparison: Content fingerprints of H23 and C23; Content fingerprints of H24 and C24.
[0066] If both are consistent, it means that the text corresponding to K104 has not changed in the new version. Therefore, there is no need to regenerate K104 or recalculate the corresponding vector. Mark K104 as an inherited knowledge block and continue to use the original knowledge block identifier and vector index relationship.
[0067] Inheritance refers to the new version of the knowledge base continuing to reference the original knowledge blocks and their vectors.
[0068] For knowledge blocks with intersecting boundaries, they are split into inherited fragments and reconstructed fragments based on the knowledge unit boundaries.
[0069] For example, K105 is composed of H25, H26, and H27, where H25 corresponds to C25 and remains unchanged, and H26 and H27 fall within the smallest reconstruction interval. Therefore: the portion corresponding to H25 is treated as an inherited fragment; the portions corresponding to H26 and H27 are treated as reconstruction fragments. The reconstruction fragments are then recombined with the current knowledge units within the interval according to their original textual arrangement in V4.
[0070] For example: C26: Article 5 Waiting period; C27: 180 days from the effective date of this contract is the waiting period; C28: If an insured event occurs during the waiting period... After recombining C26-C28, the knowledge blocks are re-divided according to the preset knowledge block length and the integrity of the knowledge unit.
[0071] For example, generating: NK201: C26+C27; NK202: C28+subsequent necessary context.
[0072] Among these measures, the re-segmentation prioritizes ensuring that a complete knowledge unit is not truncated; when the length of a single knowledge unit exceeds the preset knowledge block length, it can be further segmented according to sentence boundaries.
[0073] Vectorize NK201 and NK202 to obtain vectors corresponding to the two reconstructed knowledge blocks.
[0074] The inherited knowledge blocks and the reconstructed knowledge blocks are combined to obtain the target knowledge block set.
[0075] For example, in this V4 update: The original knowledge base contained 10,000 knowledge blocks; after comparison, only K105 and K106 were affected by this modification; among them, 9,998 knowledge blocks can be directly inherited, and only three knowledge blocks, NK201, NK202 and NK203, need to be regenerated.
[0076] Therefore, it is not necessary to recalculate the vectors for all 10,000 knowledge blocks. Instead, a block-level lineage relationship is established for each reconstructed knowledge block.
[0077] For example, the block-level lineage record of NK201 can be represented as: Current insurance document identifier: DOC-P10086-V4; Current knowledge unit range: C26-C27; Parent knowledge block identifier: K105; Update batch identifier: UPD-20260929-001.
[0078] The parent knowledge block represents the original knowledge block that has a direct content inheritance relationship with the refactored knowledge block before the update.
[0079] If a reconstructed knowledge block is formed by combining parts of two old knowledge blocks, then multiple parent knowledge block identifiers can be recorded.
[0080] The update batch identifier is used to distinguish knowledge base updates performed at different times. For example, when two regulatory documents are accessed on the same day, different update batches can be formed.
[0081] By using block-level lineage relationships, when a question-and-answer error occurs or a version rollback is required, the corresponding insurance document and its location can be located in reverse based on the refactored knowledge block.
[0082] Based on the target knowledge block set and the existing knowledge base, perform duplicate knowledge cleanup and conflict knowledge handling to obtain a candidate update set; Specifically, the steps for repetitive knowledge cleanup are as follows: Obtain the target knowledge blocks from the target knowledge block set and perform standardization processing on the text.
[0083] The text standardization process may include: removing consecutive spaces; standardizing Chinese and English punctuation; standardizing number formats; standardizing the expression formats of amounts and percentages; and removing content such as headers, footers, and page numbers that do not contain business semantics.
[0084] Generate accurate content fingerprints based on standardized text.
[0085] If the target knowledge block NK301 has the same exact content fingerprint as the existing knowledge base knowledge block K800, then the two are considered as an exact duplicate knowledge block group.
[0086] Among them, the same knowledge content may come from: Insurance terms and conditions; sales brochure; internal claims manual. Deleting just two of these documents would result in the loss of knowledge source relationships.
[0087] Therefore, select one knowledge block from the exact repeating knowledge block group as the retained knowledge block, and associate the document mapping relationship and block-level lineage relationship of other repeating knowledge blocks with the retained knowledge block.
[0088] For example: K800 is retained; NK301 no longer has a separate vector index; However, the product manual V4 corresponding to NK301, page 5, paragraph 3, continues to be preserved as a source of knowledge for K800.
[0089] For knowledge blocks with different content fingerprints, semantic similarity knowledge filtering is then performed.
[0090] For example, NK410 states: The waiting period for this product is 180 days. During the waiting period, insurance claims due to illness will not be covered.
[0091] K411 states: During the 180-day waiting period for illness after the contract takes effect, the company will not assume corresponding insurance liability.
[0092] The two texts are different, but their semantics are highly similar. Constraint data is extracted from the two knowledge blocks respectively.
[0093] In this embodiment, the constraint data may include at least the insured object, insurance liability, applicable conditions, exclusion conditions, and numerical constraints.
[0094] For example, from NK410: Insured: The insured; Insurance liability: Illness insurance benefit; Applicable conditions: After the contract takes effect; Exclusions: Illness during the waiting period; Numerical constraint: 180 days.
[0095] If the constraint elements corresponding to NK411 are completely identical and the semantic relevance meets the preset merging conditions, then the two knowledge blocks are considered to be different textual expressions of the same insurance knowledge and can be merged.
[0096] Conversely, if the numerical constraint in another knowledge block NK412 is 90 days, then no merging will be performed even if the rest of the text is highly similar.
[0097] At this point, a conflict candidate knowledge block pair is formed: NK410-NK412.
[0098] Specifically, the specific steps for conflict knowledge resolution are as follows: For conflict candidate knowledge blocks NK410 and NK412, their corresponding insurance documents and knowledge unit scopes are obtained according to their block-level lineage.
[0099] For example: NK410 comes from C26-C28 of DOC-V4; NK412 comes from H26-H28 of DOC-V3. Obtain the corresponding original document fragments to obtain the first source evidence fragment and the second source evidence fragment respectively.
[0100] The source evidence fragment is the original text extracted from the corresponding insurance document according to the knowledge unit scope, and is used to verify the content of the knowledge block in the knowledge base.
[0101] Further information to obtain: Insurance product identifier P10086; Clause hierarchy path Article 5 / Waiting period; Document version V3, V4; Effective date; Expiration date.
[0102] Input the above source evidence fragments and evidence information into the large language model.
[0103] In this step, the large language model does not directly determine which knowledge block takes effect, but is only used to identify where the two texts express different constraints.
[0104] For example, the large language model outputs candidate conflict elements: conflict type: waiting period value; first segment: 180 days; second segment: 90 days.
[0105] Then, we returned to the first and second source evidence fragments to locate the original positions corresponding to 180 days and 90 days.
[0106] If the candidate conflicting elements extracted by the large language model can be found in the original evidence, then the valid conflicting elements are obtained through consistency verification.
[0107] When two knowledge blocks belong to the same insurance product and the same clause level, and V4's effective date is later than V3, a replacement relationship is established based on the version relationship: NK412—NK410. Wherein: NK412 is the predecessor knowledge block, and NK410 is the successor knowledge block.
[0108] Add NK410 to the candidate effective knowledge block set, and add NK412 to the candidate invalid knowledge block set.
[0109] Finally, the deduplicated knowledge blocks, candidate effective knowledge blocks, and candidate invalid knowledge blocks are combined to form a candidate update set.
[0110] The candidate update set not only indicates which knowledge needs to be added this time, but also includes update information on which old knowledge needs to be stopped from participating in the retrieval.
[0111] Based on the candidate update set, a candidate incremental index is constructed, and historical question and answer records associated with the candidate update set are retrieved for replay verification. When the replay verification meets the release conditions, the current version corresponding to the current knowledge base is updated to the target version containing the candidate incremental index.
[0112] Specifically, the steps for constructing a candidate incremental index are as follows: Retrieve the base vector index and current version list of the current knowledge base. The base vector index represents the knowledge block vector indexes that are already in use in the current version.
[0113] Obtain knowledge blocks that have been added or whose content has changed relative to the current knowledge base from the candidate update set, obtain the knowledge blocks to be written, and establish an incremental vector index based on the vector of the knowledge blocks to be written. Assuming the current knowledge base contains 10,000 valid knowledge blocks, this update will ultimately result in: 3 newly added or restructured knowledge blocks; 2 replaced knowledge blocks; and 9,998 existing knowledge blocks that can still be used. For the 3 newly added or restructured knowledge blocks, only incremental vector indices will be created for their corresponding vectors.
[0114] An incremental vector index represents an index structure that stores only knowledge block vectors that have been added or changed relative to the base vector index.
[0115] Obtain the predecessor knowledge blocks that have been replaced in the candidate update set, the original knowledge blocks that have been replaced by the reconstructed knowledge blocks due to re-segmentation, and other knowledge blocks that have stopped participating in knowledge retrieval, and generate an invalid knowledge block identifier table based on the corresponding knowledge block identifier; For example: Failed knowledge block identifier table: K105; K106; K512.
[0116] The original vector can be temporarily retained in the basic vector index, and physical deletion is not required immediately.
[0117] For inherited knowledge blocks, the original vector index relationship is retained.
[0118] At the same time, establish the association between the newly added knowledge block and the current knowledge unit scope as well as the parent knowledge block.
[0119] For example: NK201—DOC-V4 / C26-C27—K105.
[0120] A candidate version list is generated based on the basic vector index, incremental vector index, invalid knowledge block identifier table, and version association relationship.
[0121] A version list can be understood as a data record that describes which indexes, valid knowledge blocks, and invalid knowledge blocks constitute a knowledge base version.
[0122] For example, the candidate version list MV4 includes: Basic index: BASE-V3; Incremental index: DELTA-20260929-001; Expired knowledge block table: INVALID-20260929-001; Version number: KB-V4.
[0123] This embodiment also includes: when performing knowledge retrieval based on the candidate version list, obtaining a first candidate retrieval result and a second candidate retrieval result from the basic vector index and the incremental vector index respectively, merging the first candidate retrieval result and the second candidate retrieval result, and deleting the candidate retrieval result whose knowledge block identifier is located in the invalid knowledge block identifier table to obtain the target retrieval result.
[0124] For example, the base index returns: K100, K105, K512, K600; Incremental index returned: NK201, NK410.
[0125] The merged series consisted of: K100, K105, K512, K600, NK201, and NK410.
[0126] After checking the invalid knowledge block identifier table, it was found that K105 and K512 have become invalid.
[0127] Therefore, K105 and K512 are removed, resulting in: K100, K600, NK201, NK410.
[0128] The above results are provided as target retrieval results to the subsequent question-and-answer generation module.
[0129] In this way, the new version does not need to copy the vector index of the entire 10,000 knowledge blocks, but only needs to save the incremental parts that have changed and the failure indicators.
[0130] Specifically, the specific steps for playback verification are as follows: Obtain the knowledge block identifiers corresponding to newly added knowledge blocks, reconstructed knowledge blocks, and knowledge blocks that have stopped participating in knowledge retrieval in the candidate update set, and obtain the set of knowledge blocks affected by the update. For example: NK201, NK202, NK410, K105, K106, K512.
[0131] Retrieve historical question and answer records from the historical question and answer records that have called any of the knowledge blocks in the set of updated and affected knowledge blocks, and obtain the directly affected question and answer records; that is, query historical question and answer records from the historical question and answer records that have called any of the above knowledge blocks, and obtain the directly affected question and answer records.
[0132] Among them, the historical Q&A records can save at least: user question, time of question occurrence, actual recalled knowledge block identifier, knowledge block arrangement order, final answer and cited knowledge block identifier.
[0133] For example, historical question Q001: What is the waiting period for this critical illness insurance? The old version actually recalled K512. Since K512 belongs to the knowledge block affected by the update, Q001 directly affects the question and answer record.
[0134] Based on the block-level lineage of each knowledge block in the knowledge block set affected by the update, obtain the corresponding insurance document identifier and knowledge unit scope, and obtain the insurance product identifier and clause level path corresponding to the knowledge unit scope; For example, NK410 corresponds to: P10086; Article 5 / Waiting Period.
[0135] From the historical question and answer records, obtain historical question and answer records that did not call the updated impact knowledge block set, but called other knowledge blocks with the same insurance product identifier or the same clause level path, and obtain related impact question and answer records; Then, continue to retrieve questions and answers from the historical question and answer records that, although K512 was not directly called, have called other knowledge blocks of the same product or the same clause level.
[0136] For example: Q002: Can I get compensation if I get sick during the waiting period? Q003: How long after the contract takes effect does the disease liability become effective? The above questions are recorded as related impact Q&A records.
[0137] The question and answer records that directly affect the question and answer, and the question and answer records that have a related impact, together constitute the replay question and answer set.
[0138] The same historical questions in the replay question-and-answer set are re-executed based on the current version and the candidate version to obtain the current replay results and the candidate replay results; that is, the same historical questions are re-executed in the current version KB-V3 and the candidate version KB-V4 respectively.
[0139] For the directly affected question and answer records, determine whether the candidate replay results exclude the corresponding invalid knowledge blocks, and determine whether the candidate replay results recall newly added knowledge blocks or reconstructed knowledge blocks associated with the corresponding historical questions; For example, for questions and answers that directly affect the record, at least the following should be checked: whether the candidate version no longer recalls invalid knowledge blocks; and whether the candidate version can recall new knowledge blocks or reconstruct knowledge blocks related to the question.
[0140] For example, Q001 was recalled in KB-V3 for: K512, K510, and K700.
[0141] Recall of NK410, K510, and K700 in KB-V4.
[0142] Since K512 has been excluded and NK410 has been successfully recalled, the expected direct impact issue has been met.
[0143] For the associated impact question and answer records, compare the current playback results with the first preset number of search knowledge blocks in the candidate playback results to obtain the change in search results; For example, for questions and answers related to the impact of the relationship, compare the top N search knowledge blocks of the current version and the candidate version.
[0144] In this embodiment, N can be 5.
[0145] Assume the Top 5 before Q003 update were: A, B, C, D, E; The updated Top 5 are: A, B, C, D, and F.
[0146] Then only one result will change.
[0147] The change in search results can be obtained by calculating the proportion of different knowledge blocks to N. For example, in this case, it is 1 / 5, or 20%.
[0148] If the changes in the related impact factors do not exceed the preset range, it indicates that this partial update has not caused abnormal disturbances to the retrieval of a large amount of irrelevant knowledge.
[0149] When the direct impact of the questions and answers meets the requirements for excluding expired knowledge and recalling new knowledge, and the change in the search results of the related impact of the questions and answers meets the preset requirements, the replay verification is determined to meet the release conditions.
[0150] The version list pointer used by the knowledge base retrieval service was then switched from MV3 to MV4.
[0151] The version list pointer switching means that the knowledge retrieval service no longer parses the basic index, incremental index and invalid knowledge blocks according to the old version list, but instead performs the retrieval according to the candidate version list. Therefore, there is no need to recopy all knowledge base data.
[0152] After the candidate version is launched, new questions and answers will continue to be recorded.
[0153] When newly added questions and answers meet the preset rollback conditions, the version list will be switched back to MV3.
[0154] The preset rollback conditions can be set according to actual business needs, for example: The following are examples of situations where more than a preset number of knowledge gap feedbacks appear within a certain time window; newly added knowledge fails to retrieve the target knowledge block for consecutive issues; invalid knowledge reappears in the final search results; or a candidate version is confirmed to be abnormal through manual review.
[0155] Version rollback only changes the version list pointed to by the retrieval service and does not require recalculation of the original vectors, thus enabling a relatively quick recovery of the previous stable knowledge base version.
[0156] Example 2 Based on Example 1, in this example, the method further includes: For predecessor and successor knowledge blocks that have the aforementioned replacement relationship, a knowledge lifecycle record is established. The knowledge lifecycle record includes at least the predecessor knowledge block identifier, the successor knowledge block identifier, the successor knowledge block effective time, and the predecessor knowledge block expiration time. For example: Predecessor knowledge block: K512; Successor knowledge block: NK410; Successor knowledge block effective time: October 1, 2026, 00:00; Predecessor knowledge block expiration time: October 1, 2026, 00:00.
[0157] The knowledge lifecycle refers to the effective time interval during which a knowledge block is allowed to participate in actual knowledge retrieval, and is not equivalent to the legal validity period of the insurance contract itself.
[0158] For example, new insurance terms may have been imported into the system on September 20, but if the terms stipulate that they will apply to newly signed contracts from October 1, then the new knowledge can be included in the candidate version in advance, but will not be officially activated until October 1.
[0159] A knowledge switching task is generated based on the effective time of the subsequent knowledge block. When the effective time of the successor knowledge block is reached, the current insurance document identifier and the current knowledge unit range are obtained based on the block-level lineage of the successor knowledge block, and the current verification content corresponding to the current knowledge unit range is obtained from the current insurance document. For example, upon reaching 00:00 on October 1, 2026, instead of immediately switching mechanically, the current insurance document and the current knowledge unit scope are first obtained based on the block-level lineage of subsequent knowledge blocks. For example, the following is obtained: DOC-P10086-V4; C26-C28.
[0160] Obtain the knowledge block content corresponding to the subsequent knowledge block, and perform a consistency check between the knowledge block content and the current verification content; re-obtain the actual text of C26-C28 from the corresponding current insurance document, and perform a consistency check between it and the content of the subsequent knowledge block NK410.
[0161] When the consistency check passes, the subsequent knowledge block is added to the set of active knowledge blocks, and the predecessor knowledge block that has the replacement relationship with the subsequent knowledge block is added to the set of inactive knowledge blocks; this indicates that the corresponding position in the insurance document has not been modified since NK410 was generated. Therefore, NK410 is activated, and the predecessor knowledge block K512 is added to the set of inactive knowledge blocks.
[0162] When the consistency check fails, or the current check content is not obtained based on the block-level lineage, the latest insurance document corresponding to the current insurance document is obtained, the latest insurance document is updated to the insurance document to be accessed, and the stable anchor matching and minimum reconstruction interval determination are re-executed.
[0163] If the actual text has changed again, for example, the waiting period has been changed from 180 days to 120 days, then NK410 will not be activated directly.
[0164] The system will use the latest insurance document as the new insurance document to be accessed, and re-execute the stable anchor point matching, minimum reconstruction interval determination, and subsequent local update steps.
[0165] Example 3 Based on the above embodiments, this embodiment provides a specific update example to further explain the relationship between the intermediate data: The current knowledge base corresponds to the insurance terms V3 of insurance product P10086, in which Article 5 / Waiting Period stipulates a waiting period of 90 days.
[0166] In the current knowledge base: K105 corresponds to H25-H27; K106 corresponds to H28-H30.
[0167] The insurance company has switched to V4, and the waiting period has been changed from 90 days to 180 days. After parsing V4, C1-C120 are obtained.
[0168] Determined after matching content fingerprint, title path, and adjacency relationship: C26-H26 are stable anchor points; C28-H28 are stable anchor points; C27 and H27 have different contents.
[0169] Therefore, C27 is first determined as the initial change section, and then C26-C28 is determined as the minimum reconstruction interval based on the original knowledge block boundary and context.
[0170] The recombined result is: NK201:C26-C27; NK202:C28, and the necessary subsequent context.
[0171] Vectorization is performed on NK201 and NK202. The remaining unchanged knowledge blocks in the existing knowledge base continue to be inherited.
[0172] NK201 records lineage: DOC-V4; C26-C27; Parent knowledge block K105; Update batch UPD001.
[0173] The system detected that K512 still exists in V3, and its content indicates a waiting period of 90 days.
[0174] NK201 and K512 are highly similar in semantics, but their numerical constraints are 180 days and 90 days, respectively. Therefore, they are not merged repeatedly, but instead form a pair of conflict candidate knowledge blocks.
[0175] After reviewing the documents, it was determined that NK201 originated from V4; K512 originated from V3; both belong to the same insurance product and the same waiting period clause; V4 took effect later than V3.
[0176] Therefore, the substitution relationship between K512 and NK201 is established.
[0177] K512 is added to the candidate invalidation knowledge block set, and NK201 is added to the candidate effective knowledge block set.
[0178] Subsequently, an incremental vector index was created for NK201, and K512 was written into the failure knowledge block identifier table.
[0179] Retrieve and replay issues from historical records that previously led to the K512 recall.
[0180] If the candidate version can recall NK201 and no longer recall K512, then the verification will be performed by direct impact.
[0181] Then, review and revisit related historical issues concerning the same product and related terms.
[0182] If the search results for related questions do not change abnormally, then perform a version list switch.
[0183] Example 4 Based on the above embodiments, this embodiment also provides an incremental self-healing update system for an insurance knowledge base, the system comprising: Document parsing unit: used to obtain the insurance document to be accessed, the historical insurance documents corresponding to the insurance document to be accessed, and the current knowledge base, and to parse the insurance document to be accessed and the historical insurance documents into knowledge unit sequences respectively, and to obtain the location anchor point based on the knowledge unit sequence; Reconstruction unit: used to match the location anchor points to obtain several stable anchor points, determine the minimum reconstruction interval based on the stable anchor points; re-block and quantize the knowledge units within the minimum reconstruction interval, and inherit the knowledge block identifier and vector index relationship of the original knowledge block for knowledge blocks outside the minimum reconstruction interval that meet the inheritance conditions, to obtain the target knowledge block set and block-level lineage relationship; Knowledge governance unit: used to perform duplicate knowledge cleanup and conflict knowledge handling based on the target knowledge block set and the existing knowledge base, and obtain a candidate update set; Feedback self-healing unit: used to construct a candidate incremental index based on the candidate update set, obtain historical question and answer records associated with the candidate update set for replay verification; when the replay verification meets the release conditions, update the current version corresponding to the current knowledge base to the target version containing the candidate incremental index.
[0184] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.
[0185] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
Claims
1. A method for incremental self-healing updating of an insurance knowledge base, characterized in that, The method includes: Obtain the insurance document to be accessed, the historical insurance document corresponding to the insurance document to be accessed, and the current knowledge base; parse the insurance document to be accessed and the historical insurance document into knowledge unit sequences respectively; and obtain the location anchor point based on the knowledge unit sequence. The location anchor points are matched to obtain several stable anchor points, and the minimum reconstruction interval is determined based on the stable anchor points; The knowledge units within the minimum reconstruction interval are re-blocked and quantized. Knowledge blocks outside the minimum reconstruction interval that meet the inheritance conditions inherit the knowledge block identifier and vector index relationship of the original knowledge block to obtain the target knowledge block set and block-level lineage relationship. Based on the target knowledge block set and the existing knowledge base, perform duplicate knowledge cleanup and conflict knowledge handling to obtain a candidate update set; Based on the candidate update set, a candidate incremental index is constructed, and historical question and answer records associated with the candidate update set are retrieved for replay verification. When the replay verification meets the release conditions, the current version corresponding to the current knowledge base is updated to the target version containing the candidate incremental index.
2. The incremental self-healing update method for an insurance knowledge base according to claim 1, characterized in that, The specific steps to obtain a stable anchor point include: Each knowledge unit in the knowledge unit sequence is normalized to obtain normalized content. A content fingerprint is obtained based on the normalized content, and a hierarchical path fingerprint is obtained based on the title hierarchical path corresponding to the knowledge unit. For any current knowledge unit in the insurance document to be accessed, obtain the historical knowledge unit in the historical insurance document that has the same content fingerprint as the current knowledge unit to obtain a first candidate knowledge unit; when there are at least two first candidate knowledge units, obtain the hierarchical path fingerprint, the content fingerprint of the preceding adjacent knowledge unit and the content fingerprint of the following adjacent knowledge unit for each first candidate knowledge unit, and match them with the corresponding information in the insurance document to be accessed to obtain a second candidate knowledge unit that uniquely corresponds to the current knowledge unit. Anchor pairs are obtained based on the current knowledge unit and the second candidate knowledge unit. The anchor pairs are sorted according to the order of the current knowledge unit in the insurance document to be accessed. Anchor pairs that have reverse order in the historical insurance document are deleted. Stable anchors and stable anchor sequences are obtained based on the remaining anchor pairs.
3. The incremental self-healing update method for an insurance knowledge base according to claim 2, characterized in that, The specific steps to obtain the minimum reconstruction interval include: Obtain the current knowledge unit that does not form a corresponding relationship between two adjacent stable anchor points, and obtain the initial change section; The initial change segment is extended to both sides to the block boundaries of existing knowledge blocks in the current knowledge base; when the block boundary is located inside a complete knowledge unit, it is extended to the outer boundary of the complete knowledge unit, and at least one unchanged knowledge unit located on both sides of the initial change segment is retained as the reconstruction context to obtain the minimum reconstruction interval.
4. The incremental self-healing update method for an insurance knowledge base according to claim 3, characterized in that, The specific steps for obtaining the target knowledge block set and the block-level lineage relationship include: Obtain several original knowledge blocks in the existing knowledge base that correspond to the minimum reconstruction interval. Based on the positional relationship between each original knowledge block and the minimum reconstruction interval, divide the original knowledge blocks into knowledge blocks outside the interval, knowledge blocks with intersecting boundaries, and knowledge blocks inside the interval. For the knowledge block outside the interval, compare the content fingerprints of the current knowledge unit and the historical knowledge unit corresponding to the knowledge block outside the interval. When the content fingerprints of the two are consistent, mark the knowledge block outside the interval as an inherited knowledge block, keep the knowledge block identifier and vector index relationship of the inherited knowledge block unchanged, and obtain the inherited knowledge block set. For the knowledge blocks with intersecting boundaries, the knowledge blocks with intersecting boundaries are split into inheritance fragments and reconstruction fragments based on the knowledge unit boundaries; The reconstructed fragments and knowledge units within the interval are recombined and divided into blocks according to the original text arrangement to obtain several reconstructed knowledge blocks. Vectorization processing is performed on the reconstructed knowledge blocks to obtain a set of reconstructed knowledge blocks. The target knowledge block set is obtained based on the inherited knowledge block set and the reconstructed knowledge block set; Record the current insurance document identifier, current knowledge unit range, parent knowledge block identifier, and update batch identifier for each reconstructed knowledge block to obtain the block-level lineage relationship.
5. The incremental self-healing update method for an insurance knowledge base according to claim 4, characterized in that, The specific steps for performing deduplication cleanup include: Each target knowledge block in the target knowledge block set is subjected to text standardization processing, and an accurate content fingerprint is obtained based on the standardized text. Based on the precise content fingerprint, knowledge blocks with the same precise content fingerprint as any target knowledge block are obtained from the target knowledge block set and the existing knowledge base to obtain a precise duplicate knowledge block group. From the precisely repeated knowledge block group, determine the retained knowledge blocks, associate the current document identifier, current knowledge unit range and block-level lineage of the non-retained knowledge blocks in the precisely repeated knowledge block group with the retained knowledge blocks, and delete the repeated vector index relationship of the non-retained knowledge blocks; For candidate knowledge blocks that do not form the exact repeating knowledge block group, several candidate similar knowledge blocks are obtained based on the vector of the candidate knowledge block. Constraint data is extracted from the candidate knowledge block and the candidate similar knowledge block respectively to obtain a first constraint element set and a second constraint element set. The constraint data includes insurance object, insurance liability, applicable conditions, exclusion conditions and numerical constraints. When the constraint elements in the first constraint element set and the second constraint element set are consistent, and the semantic relevance between the candidate knowledge block and the candidate similar knowledge block meets the preset merging condition, the candidate knowledge block and the candidate similar knowledge block are merged, and a deduplicated knowledge block set is obtained based on the target knowledge block after merging. When the semantic relevance satisfies the preset merging condition, and the first constraint element set and the second constraint element set have at least one inconsistent corresponding constraint element, a conflicting candidate knowledge block pair is obtained based on the candidate knowledge block and the candidate similar knowledge block.
6. The incremental self-healing update method for an insurance knowledge base according to claim 5, characterized in that, The specific steps for performing the aforementioned conflict knowledge resolution include: For any of the conflict candidate knowledge block pairs, the current document identifier and current knowledge unit range corresponding to the two conflict candidate knowledge blocks are obtained based on the block-level lineage relationship, and the first source evidence fragment and the second source evidence fragment are obtained from the corresponding insurance document. Obtain the insurance product identifier, clause hierarchy path, document version identifier, effective time and expiration time corresponding to the first source evidence fragment and the second source evidence fragment, and obtain the first evidence information and the second evidence information respectively; Input the first source evidence fragment, the second source evidence fragment, the first evidence information, and the second evidence information into the large language model to obtain candidate conflict elements; Based on the candidate conflict elements, the original text is located in the first source evidence segment and the second source evidence segment respectively, the actual constraint content in the original text is obtained, and the consistency of the actual constraint content with the candidate conflict elements is verified to obtain the valid conflict elements. When two conflict candidate knowledge blocks have the same insurance product identifier and clause level path, and the effective time of the latter source evidence fragment of the effective conflict element is later than that of the former source evidence fragment, a replacement relationship between the predecessor knowledge block and the successor knowledge block is established based on the document version relationship of the two source evidence fragments. Based on the replacement relationship, the successor knowledge block is added to the candidate effective knowledge block set, and the predecessor knowledge block is added to the candidate ineffective knowledge block set. The candidate update set is obtained based on the deduplicated knowledge block set, the candidate effective knowledge block set, and the candidate invalid knowledge block set.
7. The incremental self-healing update method for an insurance knowledge base according to claim 6, characterized in that, The method further includes: For predecessor and successor knowledge blocks that have the aforementioned replacement relationship, a knowledge lifecycle record is established. The knowledge lifecycle record includes at least the predecessor knowledge block identifier, the successor knowledge block identifier, the successor knowledge block effective time, and the predecessor knowledge block expiration time. A knowledge switching task is generated based on the effective time of the subsequent knowledge block. When the effective time of the successor knowledge block is reached, the current insurance document identifier and the current knowledge unit range are obtained based on the block-level lineage of the successor knowledge block, and the current verification content corresponding to the current knowledge unit range is obtained from the current insurance document. Obtain the knowledge block content corresponding to the subsequent knowledge block, and perform a consistency check between the knowledge block content and the current verification content; When the consistency check passes, the successor knowledge block is added to the set of effective knowledge blocks, and the predecessor knowledge block that has the replacement relationship with the successor knowledge block is added to the set of invalid knowledge blocks. When the consistency check fails, or the current check content is not obtained based on the block-level lineage, the latest insurance document corresponding to the current insurance document is obtained, the latest insurance document is updated to the insurance document to be accessed, and the stable anchor matching and minimum reconstruction interval determination are re-executed.
8. The incremental self-healing update method for an insurance knowledge base according to claim 7, characterized in that, The specific steps for constructing a candidate incremental index include: Obtain the basic vector index and current version list of the existing knowledge base; Obtain knowledge blocks that have been added or whose content has changed relative to the current knowledge base from the candidate update set, obtain the knowledge blocks to be written, and establish an incremental vector index based on the vector of the knowledge blocks to be written. Obtain the predecessor knowledge blocks that have been replaced in the candidate update set, the original knowledge blocks that have been replaced by the reconstructed knowledge blocks due to re-segmentation, and other knowledge blocks that have stopped participating in knowledge retrieval, and generate an invalid knowledge block identifier table based on the corresponding knowledge block identifier; Obtain the vector index relationship of the inherited knowledge blocks in the target knowledge block set, so that the inherited knowledge blocks continue to be associated with the basic vector index; Establish an index mapping relationship between the knowledge block to be written in the incremental vector index and the corresponding current knowledge unit range, and establish a version association relationship between the knowledge block to be written and the corresponding parent knowledge block based on the block-level lineage relationship; Based on the basic vector index, the incremental vector index, the invalid knowledge block identifier table, and the version association relationship, a candidate version list is generated, and the candidate incremental index is obtained.
9. The incremental self-healing update method for an insurance knowledge base according to claim 8, characterized in that, The specific steps of the playback verification include: Obtain the knowledge block identifiers corresponding to newly added knowledge blocks, reconstructed knowledge blocks, and knowledge blocks that have stopped participating in knowledge retrieval in the candidate update set, and obtain the set of knowledge blocks affected by the update. Retrieve historical question and answer records from the historical question and answer records that have called any knowledge block in the set of knowledge blocks whose updates have been affected, and obtain the directly affected question and answer records. Based on the block-level lineage of each knowledge block in the knowledge block set affected by the update, obtain the corresponding insurance document identifier and knowledge unit scope, and obtain the insurance product identifier and clause level path corresponding to the knowledge unit scope; From the historical question and answer records, obtain historical question and answer records that did not call the updated impact knowledge block set, but called other knowledge blocks with the same insurance product identifier or the same clause level path, and obtain related impact question and answer records; A replay question and answer set is obtained based on the directly impacting question and answer records and the related impacting question and answer records; Knowledge retrieval is performed again on the same historical questions in the replay question-and-answer set based on the current version and the candidate version, respectively, to obtain the current replay results and the candidate replay results; For the directly affected question and answer records, determine whether the candidate replay results exclude the corresponding invalid knowledge blocks, and determine whether the candidate replay results recall newly added knowledge blocks or reconstructed knowledge blocks associated with the corresponding historical questions; For the associated impact question and answer records, compare the current playback results with the first preset number of search knowledge blocks in the candidate playback results to obtain the change in search results; When the directly affecting question and answer records meet the exclusion requirements of the invalid knowledge block and the recall requirements of the newly added or reconstructed knowledge block, and the change in the retrieval results of the associated affecting question and answer records does not exceed the preset change range, it is determined that the replay verification meets the release conditions, and the version list corresponding to the knowledge base retrieval service is switched from the current version list to the candidate version list. When the newly added question and answer records after the version list switch is completed meet the preset rollback conditions, the version list corresponding to the knowledge base retrieval service will be switched from the candidate version list back to the current version list.
10. An incremental self-healing update system for an insurance knowledge base, characterized in that, The system includes: Document parsing unit: used to obtain the insurance document to be accessed, the historical insurance documents corresponding to the insurance document to be accessed, and the current knowledge base, and to parse the insurance document to be accessed and the historical insurance documents into knowledge unit sequences respectively, and to obtain the location anchor point based on the knowledge unit sequence; Reconstruction unit: used to match the location anchor points to obtain several stable anchor points, determine the minimum reconstruction interval based on the stable anchor points; re-block and quantize the knowledge units within the minimum reconstruction interval, and inherit the knowledge block identifier and vector index relationship of the original knowledge block for knowledge blocks outside the minimum reconstruction interval that meet the inheritance conditions, to obtain the target knowledge block set and block-level lineage relationship; Knowledge governance unit: used to perform duplicate knowledge cleanup and conflict knowledge handling based on the target knowledge block set and the existing knowledge base, and obtain a candidate update set; Feedback self-healing unit: used to construct a candidate incremental index based on the candidate update set, obtain historical question and answer records associated with the candidate update set for replay verification; when the replay verification meets the release conditions, update the current version corresponding to the current knowledge base to the target version containing the candidate incremental index.
