An AI-based dual-channel representation-based knowledge base slice retrieval method

Abstract

The application discloses a kind of knowledge base slice retrieval methods based on AI dual-channel representation, comprising: receiving knowledge base original data, parsing structure attribute and content attribute, generate standard knowledge object set;Three-phase reversible split processing is executed, and knowledge unit is divided, three kinds of split are generated and split relationship atlas is constructed;Receive search request, execute main idea qualification boundary extraction, build query mirror body and generate target representation;Improved RetNet dual-channel response competition model is constructed, the unique response potential difference is calculated, and the target split group is determined;Around target split group, build source evidence lock spectrum, execute four kinds of lock judgment, form the minimum lock necessary chain;The minimum lock necessary chain corresponding lock answer chain is output, and the main answer and necessary boundary source split are returned.The application realizes slice main answer identification, boundary complement and source tracing by three-phase split, improved RetNet dual-channel competition and source evidence lock.

Description

Technical Field

[0001] This invention relates to the field of information retrieval technology, and in particular to a knowledge base slice retrieval method based on AI dual-channel representation. Background Technology

[0002] With the continuous advancement of enterprise digitalization and intelligent applications, a large amount of policy documents, technical specifications, product manuals, operation guides, Q&A records, and business knowledge data are being accumulated in knowledge base systems. To improve knowledge retrieval efficiency, existing systems typically use full-text search, keyword matching, vector search, or question-and-answer search to quickly locate knowledge base content and return relevant fragments to the user. To adapt to the needs of long document processing, existing technologies generally adopt text slicing, splitting the original document into multiple fragments, creating an index, and then combining semantic coding models or keyword ranking models to complete the retrieval process, thereby meeting the knowledge acquisition needs of basic queries and regular question-and-answer scenarios.

[0003] Existing technologies can achieve certain results in scenarios with relatively independent knowledge content and simple structures, but they still have significant limitations in complex knowledge base environments. Because the same knowledge content is often broken down into multiple slices, accompanied by additional information such as time conditions, applicable objects, version changes, reference continuations, and exception restrictions, existing methods typically only return candidate slices based on semantic similarity or keyword hit rate, making it difficult to identify the target slice that truly qualifies as the main answer among multiple candidate slices. When the core conclusion and its conditions are scattered across different slices, it is easy to return only the main conclusion slice while omitting constraints; when old and new versions of content coexist, it is easy to recall content that has been replaced; and when the document contains exception clauses or prohibited answer boundaries, it is also easy to output results with incorrect scope of application.

[0004] Most existing technologies treat knowledge fragments as independent text units, lacking the ability to jointly model dependencies, restrictions, citations, and version relationships between fragments. This results in outputs that, while seemingly relevant, are deficient in terms of source tracing, boundary integrity, and reliability. This is particularly problematic in scenarios such as enterprise policy queries, technical specification searches, and professional knowledge Q&A, where users not only focus on the answer content itself but also on whether the answer is applicable, whether it is subject to restrictions, and whether the source is clearly defined.

[0005] Therefore, how to provide a knowledge base slice retrieval method based on AI dual-channel representation is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a knowledge base slice retrieval method based on AI dual-channel representation. This invention fully utilizes the construction of three-phase reversible knowledge slices, improved RetNet dual-channel response competition modeling, and source evidence locking spectrum solidification processing technology. It describes in detail the implementation methods for identifying the main answer qualification, filling in the establishment boundary, excluding prohibited answers, and tracing the source output of the scattered slice content in the knowledge base. It has the advantages of high retrieval accuracy, strong applicability boundary integrity, high source credibility, and good adaptability to complex knowledge scenarios.

[0007] A knowledge base slice retrieval method based on AI dual-channel representation according to an embodiment of the present invention includes: Receive raw data from the knowledge base, perform parsing processing on the raw data, extract document structure attribute information and content attribute information, and generate a standard knowledge object set; Three-phase reversible fragmentation is performed on each standard knowledge object. Knowledge units are divided according to semantic boundaries, clause boundaries, condition boundaries, version boundaries, and reference boundaries. For each knowledge unit, main interpretation fragments, qualification supplementary evidence fragments, and forbidden answer counter-evidence fragments are generated, and a fragmentation relationship graph is generated. Receive a retrieval request, perform subject extraction, qualification extraction and boundary extraction on the retrieval request, generate query subject definition fragment, query qualification fragment and query boundary fragment, construct query mirror body, and generate query target representation based on query mirror body; A dual-channel response competition model based on an improved RetNet is constructed. The main interpretation fragment, the qualification supplementary evidence fragment, and the forbidden answer counter-evidence fragment are input into the dual-channel response competition model. The same question competition relationship and version substitution relationship are extracted by combining the fragment relationship graph. The unique response potential difference is calculated and the target fragment group is determined. A source evidence locking spectrum is constructed around the target fragment group. Interpretive locking, qualification locking, prohibition of answer locking and source locking are performed on the source evidence locking spectrum. When each locking judgment meets the preset conditions, progressive release processing is performed to identify the necessary locking nodes and form the minimum necessary locking chain. Output the chain of closure answers corresponding to the minimum closure necessary chain, and return the main answer fragment, the closure necessary qualification fragment, the closure necessary forbidden answer boundary fragment, and the closure necessary source fragment.

[0008] Optionally, the original data of the knowledge base includes policy document data, technical specification data, product manual data, operation guide data, question and answer record data, work order record data, database text record data, web page collection data, and version update record data.

[0009] Optionally, the document structure attribute information includes document title, chapter path, clause number, citation mark, version identifier, time identifier, source identifier, and permission identifier, and the content attribute information includes main text content and table fields.

[0010] Optionally, generating a standard knowledge object set includes performing character normalization, number normalization, version normalization, and reference normalization on document structure attribute information and content attribute information to generate a standard knowledge object set.

[0011] Optionally, the step of generating main definition fragments, qualification supplementary evidence fragments, and forbidden answer counter-evidence fragments for each knowledge unit, and generating a fragment relationship graph, includes: Perform core positioning processing on each standard knowledge object, divide candidate fragments into segments based on chapter level change positions, clause switch positions, condition switch positions, version switch positions, reference switch positions, and semantic transition positions, identify core statement groups that can independently express rules, definitions, conclusions, or operational matters in each candidate fragment segment, and generate corresponding knowledge cores; The inner circle convergence process is performed with each knowledge core as the center. The main conclusion content, main definition content, main rule content or main operation content directly corresponding to the knowledge core are extracted along the forward and backward content of the knowledge core. They are merged according to the continuous expression relationship to generate the corresponding main interpretation fragment. The outer ring of supplementary certification is processed around each main definition fragment. The time limit, version limit, applicable object, clause number, field definition and reference content that have a dependency relationship with the main definition fragment are retrieved and merged according to the same establishment condition relationship to generate the corresponding qualification supplementary certification fragment. Boundary sealing is performed around each main interpretation fragment. Exceptions, inapplicable scope, invalidation conditions, substitution relationships, conflicting restrictions, and abolished related content that have a restrictive relationship with the main interpretation fragment are retrieved and merged according to the same restriction boundary attribution relationship to generate forbidden answer counter-evidence fragments. Using the main interpretation fragment, qualification supplementary evidence fragment, and forbidden answer counter-evidence fragment as fragment nodes, qualification continuation edges are established based on the establishment dependency relationship, counter-evidence suppression edges are established based on the restriction relationship, boundary verification edges are established based on the boundary correspondence relationship, reference continuation edges are established based on the reference inheritance relationship, and version replacement edges are established based on the version inheritance relationship, thus generating a fragment relationship graph.

[0012] Optionally, the step of constructing a query mirror body and generating a query target representation based on the query mirror body includes: Upon receiving a retrieval request, the system performs a normalization and segmentation process on the retrieval request, dividing it into multiple query expression units according to pause marks, conjunctions, qualifiers, exclusion words, number markers, time markers, version markers, and source markers, and performs order preservation processing on each query expression unit to generate a sequence of query units arranged in the original expression order; Perform thematic anchoring processing on the query unit sequence to identify the query object content, query item content and query result content that represent the core retrieval intent. Merge query expression units that point to the same retrieval thematic intent to generate query thematic explanatory fragments. The eligibility linking process is performed around the main definition fragment of the query. It retrieves the time limit content, version limit content, applicable object content, clause number content, field definition content, reference content and source limit content that have an establishment dependency relationship with the main definition fragment of the query, and merges them according to the same establishment condition relationship to generate query eligibility fragments. Boundary merging is performed around the main definition fragment of the query, retrieving exceptions, inapplicable content, invalid content, alternative content, conflicting content, and excluded content that have a restrictive relationship with the main definition fragment of the query, and merging them according to the same restriction boundary attribution relationship to generate query boundary fragments; Using the main definition fragment of the query as the main mirror node, the query qualification fragment as the qualification attachment node, and the query boundary fragment as the boundary constraint node, a query mirror body is constructed according to the subject correspondence, qualification dependency relationship and boundary restriction relationship. The main retrieval thread, qualification screening conditions, boundary exclusion conditions, number positioning information, version positioning information, time positioning information, source positioning information and applicable scope positioning information are extracted from the query mirror body to generate the query target representation.

[0013] Optionally, the step of extracting competing relationships and version substitution relationships by combining the fragment relationship map, calculating the unique response potential difference, and determining the target fragment group includes: Read the query target representation, main definition fragments, qualification supplementary evidence fragments, prohibited answer counter-evidence fragments, and fragment relationship graph. Based on the main retrieval line, qualification screening conditions, boundary exclusion conditions, number positioning information, version positioning information, time positioning information, source positioning information, and applicable scope positioning information in the query target representation, extract the candidate fragment set corresponding to the retrieval request. Group the candidate fragment set according to the relationship of the same question, the relationship of qualification continuation, the relationship of version substitution, and the relationship of counter-evidence suppression to generate candidate fragment groups. A dual-channel response competition model based on an improved RetNet is constructed, with RetNet as the reserved coding backbone. The dual-channel response competition model sets up a response carrying channel and a substitute response encroachment channel. The response carrying channel sets up a main interpretation reservation unit and a qualification continuation reservation unit. The substitute response encroachment channel sets up a same-topic competition reservation unit, a version substitution reservation unit, and a counter-evidence suppression reservation unit. A relationship masking scheduling unit and a competition write-back unit are set up between the response carrying channel and the substitute response encroachment channel. Based on the main definition fragments, qualification supplementary fragments, and the main retrieval line and qualification screening conditions in the query target representation of the candidate fragment group, a response input sequence is constructed. The response input sequence is input into the response carrying channel. The core definition content in the main definition fragment is retained by the main definition retention unit, and the establishment condition content in the qualification supplementary fragment is retained by the qualification continuation retention unit. The carrying vector corresponding to each candidate fragment group is generated. Based on the forbidden answer counter-evidence fragments in the candidate fragment groups and the same-topic association, version substitution relationship and counter-evidence suppression relationship in the fragment relationship graph, an invasion input sequence is constructed. The invasion input sequence is input into the substitute answer invasion channel. The same-topic competition retention unit retains the same-topic competition content, the version substitution retention unit retains the version substitution content, and the counter-evidence suppression retention unit retains the forbidden answer suppression content, thereby generating the invasion vector corresponding to each candidate fragment group. The relational masking scheduling unit performs relational constraint alignment on the bearer vector and the encroachment vector. The competitive write-back unit writes the substitution suppression result in the encroachment vector back to the bearer vector and the qualification continuation result in the bearer vector back to the encroachment vector. The response bearer strength, substitution encroachment strength, counter-evidence suppression strength and unique response potential difference are generated for each candidate shard group. The target shard group is determined based on the unique response potential difference.

[0014] Optionally, forming the minimum necessary chain of locking includes: Source evidence aggregation is performed around the target fragment group. The main interpretation fragment, qualification supplementary evidence fragment, and forbidden answer counter-evidence fragment in the target fragment group are read. Based on the qualification continuation relationship, counter-evidence suppression relationship, reference continuation relationship, version substitution relationship and chapter affiliation relationship in the fragment relationship graph, the source-related fragments corresponding to the target fragment group are extracted to generate a source evidence candidate set. Using the main interpretive fragment in the target fragment group as the interpretive source node, the qualification supplementary evidence fragment as the qualification continuation node, the forbidden answer counter-evidence fragment as the forbidden answer boundary node, and the source-related fragment as the source return node, a continuation link is established between the interpretive source node and the qualification continuation node based on the qualification continuation relationship, a suppression link is established between the interpretive source node and the forbidden answer boundary node based on the counter-evidence suppression relationship, and a source return link is established between the interpretive source node, the qualification continuation node, the forbidden answer boundary node and the source return node based on the citation continuation relationship, the version substitution relationship and the chapter affiliation relationship, thereby generating the source evidence closure spectrum; Perform interpretation closure and qualification closure on the source certificate closure spectrum. The interpretation closure result is determined based on whether the main interpretation fragment corresponding to the interpretation source node completely covers the main retrieval line in the query target representation. The qualification closure result is determined based on whether the qualification supplementary certificate fragment corresponding to the qualification continuation node completely covers the qualification screening conditions in the query target representation. Perform forbidden answer blocking and source blocking on the source evidence blocking spectrum. Determine the forbidden answer blocking result based on whether the forbidden answer counter-evidence fragment corresponding to the forbidden answer boundary node forms a complete restriction on the boundary exclusion condition in the query boundary segment. Determine the source blocking result based on whether the source association fragment corresponding to the source node forms a continuous source path from the target fragment group to the original knowledge object. When the interpretation of the locking result, the qualification of the locking result, the prohibition of answer locking result, and the source locking result all meet the preset conditions, the source certificate locking spectrum is processed by progressive release. Non-core nodes are released layer by layer in the order of source node, qualification continuation node, and prohibition of answer boundary node. After each release, the interpretation of the locking result, qualification of the locking result, prohibition of answer locking result, and source locking result are re-executed. Nodes that still meet the preset conditions after release are identified as releaseable nodes. Nodes that cause any locking result to fail to meet the preset conditions after release are identified as locking necessary nodes. The minimum locking necessary chain is formed based on all locking necessary nodes.

[0015] Optionally, the chain of locking answers corresponding to the minimum necessary chain of locking output includes: Read all the necessary locking nodes in the minimum necessary locking chain, and divide the necessary locking nodes into main answer nodes, qualified nodes, forbidden answer boundary nodes and source nodes according to the fragment type corresponding to each necessary locking node; The main answer node is used as the main answer fragment, the qualification supplement fragment is used as the qualification necessary fragment, the forbidden answer boundary fragment is used as the forbidden answer boundary fragment, and the source association fragment is used as the source necessary fragment. According to the connection order and relationship order of the main answer fragment, the necessary qualification fragment for closure, the necessary forbidden answer boundary fragment for closure, and the necessary source fragment for closure in the minimum necessary chain for closure, the link orchestration process is performed on each fragment. Based on the original knowledge object identifier, chapter path identifier, clause number identifier, version identifier and source identifier corresponding to each necessary source fragment for closure, the source attachment process is performed on the orchestrated link to generate the closure answer chain and the corresponding source tracing result. Output the locked answer chain, and simultaneously return the main answer fragment, the locked necessary qualification fragment, the locked necessary forbidden answer boundary fragment, the locked necessary source fragment, and the source tracing result corresponding to the locked answer chain.

[0016] The beneficial effects of this invention are: This invention performs a three-phase reversible fragmentation process on the original content of the knowledge base, splitting the same knowledge unit into a main definition fragment, a qualification supplementary evidence fragment, and a prohibition answer counter-evidence fragment. Furthermore, it establishes qualification continuation relationships, counter-evidence suppression relationships, citation continuation relationships, and version substitution relationships, enabling the core conclusions, conditions for validity, and limiting boundaries originally scattered across different text fragments to form a structured association. Compared to existing methods that only slice text by length or paragraph, this invention effectively reduces the information fragmentation problem caused by the mechanical splitting of knowledge content, improving the recognizability and organizeability of complex knowledge content after slicing.

[0017] This invention constructs a dual-channel response competition model based on an improved RetNet, jointly analyzing the response carrying capacity and substitution invasion capacity of candidate fragment groups. It no longer relies solely on semantic similarity or keyword hit rate for ranking, but can identify the target fragment group truly qualified to be the primary answer among multiple similar candidate results. For knowledge scenarios involving competition for the same question, version substitution, exception restrictions, and multi-fragment combined answers, this invention can effectively reduce the problems of false recall of old version content, false output of partial answers, and output of results with missing conditions, thereby improving the accuracy and stability of knowledge base fragment retrieval results.

[0018] This invention constructs a source-evidence closure spectrum around a target fragment group and jointly determines the interpretation closure state, eligibility closure state, prohibited answer closure state, and source closure state. Upon satisfying the closure conditions, it outputs the closure answer chain corresponding to the minimum necessary closure chain. Compared to existing technologies that only return a few relevant fragments, this invention's output not only includes the core answer content but also simultaneously includes the validity boundary, prohibited answer boundary, and source tracing path. This improves the completeness, reliability, and auditability of the search results, making it particularly suitable for applications with high accuracy requirements, such as policy queries, technical specification searches, enterprise knowledge Q&A, and professional document searches. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a knowledge base slice retrieval method based on AI dual-channel representation proposed in this invention; Figure 2 This is a schematic diagram of the structure of a dual-channel response competition model based on an improved RetNet, which is the basis of the knowledge base slice retrieval method based on AI dual-channel representation proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 1 and Figure 2 A knowledge base slice retrieval method based on AI dual-channel representation includes: Receive raw data from the knowledge base, perform parsing processing on the raw data, extract document structure attribute information and content attribute information, and generate a standard knowledge object set; Three-phase reversible fragmentation is performed on each standard knowledge object. Knowledge units are divided according to semantic boundaries, clause boundaries, condition boundaries, version boundaries, and reference boundaries. For each knowledge unit, main interpretation fragments, qualification supplementary evidence fragments, and forbidden answer counter-evidence fragments are generated, and a fragmentation relationship graph is generated. Receive a retrieval request, perform subject extraction, qualification extraction and boundary extraction on the retrieval request, generate query subject definition fragment, query qualification fragment and query boundary fragment, construct query mirror body, and generate query target representation based on query mirror body; A dual-channel response competition model based on an improved RetNet is constructed. The main interpretation fragment, the qualification supplementary evidence fragment, and the forbidden answer counter-evidence fragment are input into the dual-channel response competition model. The same question competition relationship and version substitution relationship are extracted by combining the fragment relationship graph. The unique response potential difference is calculated and the target fragment group is determined. A source evidence locking spectrum is constructed around the target fragment group. Interpretive locking, qualification locking, prohibition of answer locking and source locking are performed on the source evidence locking spectrum. When each locking judgment meets the preset conditions, progressive release processing is performed to identify the necessary locking nodes and form the minimum necessary locking chain. Output the chain of closure answers corresponding to the minimum closure necessary chain, and return the main answer fragment, the closure necessary qualification fragment, the closure necessary forbidden answer boundary fragment, and the closure necessary source fragment.

[0022] In this embodiment, the original data of the knowledge base includes policy document data, technical specification data, product manual data, operation guide data, Q&A record data, work order record data, database text record data, web page collection data, and version update record data.

[0023] In this embodiment, the document structure attribute information includes document title, chapter path, clause number, citation mark, version identifier, time identifier, source identifier, and permission identifier, and the content attribute information includes text content and table fields.

[0024] In this embodiment, generating a standard knowledge object set includes performing character normalization, number normalization, version normalization, and reference normalization on document structure attribute information and content attribute information to generate a standard knowledge object set.

[0025] In this embodiment, the step of generating main definition fragments, qualification supplementary evidence fragments, and forbidden answer counter-evidence fragments for each knowledge unit, and generating a fragment relationship graph, includes: Perform core positioning processing on each standard knowledge object, divide candidate fragments into segments based on chapter level change positions, clause switch positions, condition switch positions, version switch positions, reference switch positions, and semantic transition positions, identify core statement groups that can independently express rules, definitions, conclusions, or operational matters in each candidate fragment segment, and generate corresponding knowledge cores; The inner circle convergence process is performed with each knowledge core as the center. The main conclusion content, main definition content, main rule content or main operation content directly corresponding to the knowledge core are extracted along the forward and backward content of the knowledge core. They are merged according to the continuous expression relationship to generate the corresponding main interpretation fragment. The outer ring of supplementary certification is processed around each main definition fragment. The time limit, version limit, applicable object, clause number, field definition and reference content that have a dependency relationship with the main definition fragment are retrieved and merged according to the same establishment condition relationship to generate the corresponding qualification supplementary certification fragment. Boundary sealing is performed around each main interpretation fragment. Exceptions, inapplicable scope, invalidation conditions, substitution relationships, conflicting restrictions, and abolished related content that have a restrictive relationship with the main interpretation fragment are retrieved and merged according to the same restriction boundary attribution relationship to generate forbidden answer counter-evidence fragments. Using the main interpretation fragment, qualification supplementary evidence fragment, and forbidden answer counter-evidence fragment as fragment nodes, qualification continuation edges are established based on the establishment dependency relationship, counter-evidence suppression edges are established based on the restriction relationship, boundary verification edges are established based on the boundary correspondence relationship, reference continuation edges are established based on the reference inheritance relationship, and version replacement edges are established based on the version inheritance relationship, thus generating a fragment relationship graph.

[0026] In this embodiment, the step of constructing a query mirror body and generating a query target representation based on the query mirror body includes: Upon receiving a search request, the system performs a normalization and segmentation process, dividing the request into multiple query expression units based on pause marks, conjunctions, qualifiers, exclusion words, number markers, time markers, version markers, and source markers. The system then performs order preservation processing on each query expression unit, generating a sequence of query units arranged in the original expression order. Specifically, the generated sequence of query units arranged in the original expression order is as follows: The search request is scanned from front to back according to the order of character appearance. The positions of each pause mark, conjunction, qualifier, exclusion word, number mark, time mark, version mark and source mark in the search request are recorded. The corresponding candidate query expression units are extracted according to the text interval between adjacent marks. For each candidate query expression unit, record the unit start position, unit end position, pre-mark type and post-mark type. Merge the candidate query expression units that only contain the marked content into the adjacent and semantically continuous candidate query expression units to generate multiple valid query expression units. Based on the starting position of each valid query expression unit in the retrieval request, they are arranged sequentially from front to back. Each valid query expression unit is assigned a corresponding sequence position identifier, generating a query unit sequence arranged in the original expression order. Perform thematic anchoring processing on the query unit sequence to identify the query object content, query item content and query result content that represent the core retrieval intent. Merge query expression units that point to the same retrieval thematic intent to generate query thematic explanatory fragments. The eligibility linking process is performed around the main definition fragment of the query. It retrieves the time limit content, version limit content, applicable object content, clause number content, field definition content, reference content and source limit content that have an establishment dependency relationship with the main definition fragment of the query, and merges them according to the same establishment condition relationship to generate query eligibility fragments. Boundary merging is performed around the main definition fragment of the query, retrieving exceptions, inapplicable content, invalid content, alternative content, conflicting content, and excluded content that have a restrictive relationship with the main definition fragment of the query, and merging them according to the same restriction boundary attribution relationship to generate query boundary fragments; Using the main definition fragment of the query as the main mirror node, the query qualification fragment as the qualification attachment node, and the query boundary fragment as the boundary constraint node, a query mirror body is constructed according to the subject correspondence, qualification dependency relationship, and boundary restriction relationship. From the query mirror body, the main retrieval thread, qualification filtering conditions, boundary exclusion conditions, number positioning information, version positioning information, time positioning information, source positioning information, and applicable scope positioning information are extracted to generate a query target representation. Specifically, the construction of the query mirror body according to the subject correspondence, qualification dependency relationship, and boundary restriction relationship is as follows: The main query definition fragment is identified as the main mirror node. The content of the search object, the content of the search item, and the content pointed to by the result corresponding to the main query definition fragment are recorded. The query qualification fragment and the query boundary fragment that point to the same search item as the main query definition fragment are respectively associated with the main mirror node. Based on the dependency relationships between each query qualification fragment and the main query definition fragment, qualification attachment edges are established. These dependency relationships include time-limited dependencies, version-limited dependencies, applicable object dependencies, clause number dependencies, field definition dependencies, reference basis dependencies, and source-limited dependencies. Based on the restrictive relationships between each query boundary fragment and the main query definition fragment, boundary constraint edges are established. These restrictive relationships include exception restriction relationships, inapplicability scope relationships, invalidity restriction relationships, substitution restriction relationships, conflict restriction relationships, and exclusion restriction relationships. The main mirror node, qualification attachment edge, boundary constraint edge, qualification attachment node, and boundary constraint node are then associated and combined to generate the query mirror body.

[0027] In this embodiment, the step of extracting competing relationships and version substitution relationships by combining the fragment relationship map, calculating the unique response potential difference, and determining the target fragment group includes: Read the query target representation, main definition fragments, qualification supplementary evidence fragments, prohibited answer counter-evidence fragments, and fragment relationship graph. Based on the main retrieval line, qualification screening conditions, boundary exclusion conditions, number positioning information, version positioning information, time positioning information, source positioning information, and applicable scope positioning information in the query target representation, extract the candidate fragment set corresponding to the retrieval request. Group the candidate fragment set according to the relationship of the same question, the relationship of qualification continuation, the relationship of version substitution, and the relationship of counter-evidence suppression to generate candidate fragment groups. A dual-channel response competition model based on an improved RetNet is constructed, with RetNet as the reserved coding backbone. The dual-channel response competition model sets up a response carrying channel and a substitute response encroachment channel. The response carrying channel sets up a main interpretation reservation unit and a qualification continuation reservation unit. The substitute response encroachment channel sets up a same-topic competition reservation unit, a version substitution reservation unit, and a counter-evidence suppression reservation unit. A relationship masking scheduling unit and a competition write-back unit are set up between the response carrying channel and the substitute response encroachment channel. Based on the main definition fragments, qualification supplementary evidence fragments, and the main retrieval thread and qualification screening conditions in the query target representation of the candidate fragment group, a response input sequence is constructed. This response input sequence is then input into the response carrying channel. The core definition content of the main definition fragment is preserved through the main definition retention unit, and the establishment condition content of the qualification supplementary evidence fragment is preserved through the qualification continuation retention unit. This generates the carrying vector corresponding to each candidate fragment group, where: The core interpretation content in the main interpretation segment is preserved through the main interpretation retention unit, specifically as follows: Arrange the main interpretation fragments in the candidate fragment group in order according to the reference continuation relationship and version substitution relationship in the fragment relationship graph to generate the main interpretation input fragment sequence; Based on the main retrieval line in the query target representation, the content of the retrieval object, the content of the retrieval item, and the content pointed to by the result in each main definition input segment are aligned and identified. The core definition content that is consistent with the main retrieval line is extracted, and the corresponding fragment position, continuation position, and substitution position are recorded. According to the order of the main definition input fragment sequence, the core definition content of each core definition is continuously retained. The core definition content already retained in the previous main definition input fragment is merged with the core definition content identified in the current main definition input fragment to generate the corresponding main definition retention result. The conditions for establishment in the qualification supplementary certificate are maintained through the qualification continuation retention unit, specifically as follows: Arrange the qualifying supplementary fragments in the candidate fragment group in order according to their qualifying succession relationship with the main interpretation fragment to generate a qualifying input fragment sequence; Based on the eligibility screening criteria in the query target representation, the time limit, version limit, applicable object, clause number, field definition, reference basis and source limit in each eligibility input segment are identified and matched. The content that meets the eligibility screening criteria is extracted and the corresponding attachment and continuation positions are recorded. According to the order of the qualification input segment sequence, the continuation and retention processing is performed on each qualification condition content. The qualification condition content already retained in the previous qualification input segment is merged with the qualification condition content identified in the current qualification input segment to generate the corresponding qualification continuation and retention result. Generate the bearing vector corresponding to each candidate fragment group, specifically as follows: The main interpretation retention results and qualification continuation retention results corresponding to each candidate segment group are aligned and combined according to the segment affiliation relationship to form a set of bearing features that correspond one-to-one with each candidate segment group. The main interpretation retention results and eligibility continuation retention results in each bearing feature set are processed by a unified dimension mapping, and then spliced ​​together in the order of main interpretation retention results first and eligibility continuation retention results last to generate the initial bearing representation corresponding to each candidate shard group. Perform intra-group aggregation processing on each initial bearer representation, retain the combined results that can characterize the core interpretation content of the main interpretation section and the conditions for the establishment of the qualification supplementary section, and generate the bearer vector corresponding to each candidate section group. Based on the forbidden answer counter-evidence fragments in the candidate fragment groups and the same-topic association, version substitution, and counter-evidence suppression relationships in the fragment relationship graph, an encroachment input sequence is constructed. This encroachment input sequence is then input into the answer substitution encroachment channel. Same-topic competition retention units maintain same-topic competition content, version substitution retention units maintain version substitution content, and counter-evidence suppression retention units maintain forbidden answer suppression content. This generates an encroachment vector corresponding to each candidate fragment group, where: The same topic competition is maintained by retaining the same topic competition unit, specifically as follows: Based on the related relationships of the same topic in the fragment relationship map, related fragments that point to the same search item as the main definition fragments corresponding to each candidate fragment group are extracted, and a sequence of competing fragments with the same topic is generated. The fragments are arranged according to the order of their association with the candidate fragment groups. The content of the search object, the content of the search item, and the content of the result in each fragment are identified and the content of the same topic that competes with the main definition fragment of the candidate fragment group is extracted. According to the order of the sequence of competing segments of the same topic, the continuous retention process is performed on each competing content of the same topic. The competing content already retained in the previous competing segment of the same topic is merged with the competing content identified in the current competing segment of the same topic to generate the corresponding competing retention result of the same topic. Version replacement content is preserved through version replacement retention units, specifically as follows: Based on the version substitution relationship in the fragment relationship map, related fragments with version succession relationship with the main interpretation fragments, qualification supplementary evidence fragments and forbidden answer counter-evidence fragments corresponding to each candidate fragment group are extracted, and a version substitution fragment sequence is generated. The replacement fragments of each version are arranged according to the order of version sequence and the order of replacement direction. The version identifier content, replacement direction content, successor content and invalid content in each version replacement fragment are identified and the version replacement content that forms a replacement relationship with the candidate fragment group is extracted. According to the order of the version replacement fragment sequence, the replacement content of each version is continuously retained. The replacement content already retained in the previous version replacement fragment is merged with the replacement content identified in the current version replacement fragment to generate the corresponding version replacement retention result. The suppressed content is preserved by using the proof by contradiction suppression unit, specifically as follows: Based on the counter-evidence suppression relationship in the fragment relationship map, forbidden counter-evidence fragments that have a restrictive relationship with the main interpretation fragments and qualifying supplementary evidence fragments corresponding to each candidate fragment group are extracted, and a forbidden answer suppression fragment sequence is generated. The suppression fragments are arranged according to the suppression association order between each forbidden response suppression fragment and the candidate fragment group. The exception condition content, inapplicable scope content, invalid condition content, alternative restriction content and conflict restriction content in each forbidden response suppression fragment are identified and the forbidden response suppression content that suppresses the candidate fragment group is extracted. According to the order of the forbidden answer suppression fragment sequence, the forbidden answer suppression content is continuously retained. The suppressed content already retained in the previous forbidden answer suppression fragment is merged with the suppressed content identified in the current forbidden answer suppression fragment to generate the corresponding counter-evidence suppression retention result. The relational masking scheduling unit performs relational constraint alignment on the bearer vector and the encroachment vector. The contention write-back unit writes the substitution suppression result from the encroachment vector back to the bearer vector, and the qualification continuation result from the bearer vector back to the encroachment vector. This generates the response bearer strength, substitution encroachment strength, counter-evidence suppression strength, and unique response potential difference for each candidate fragment group. The target fragment group is then determined based on the unique response potential difference, where: The relational masking scheduling unit performs relational constraint alignment on the carrying vector and the encroaching vector, specifically as follows: Based on the related relationships, qualification continuation relationships, version substitution relationships, and counter-evidence suppression relationships in the fragment relationship graph, the relationship constraint information corresponding to each candidate fragment group is extracted, and the corresponding relationship positions are marked for the carrying vector and encroachment vector in each candidate fragment group respectively. According to the fragment affiliation relationship of each candidate fragment group, the main interpretation and qualification continuation parts in the bearing vector are aligned with the same question competition part, version replacement part and forbidden answer suppression part in the encroachment vector to generate the corresponding alignment vector group; Based on the relational constraint information in each pair of vector groups, masking is performed on vector parts that do not have a direct relationship, and retention is performed on vector parts that have a direct relationship, generating carrying pairing results and encroaching pairing results that satisfy the relational constraints; The response bearing capacity, substitution encroachment capacity, counter-evidence suppression capacity, and unique response potential difference are generated for each candidate fragment group, specifically as follows: The load alignment results corresponding to each candidate segment group are aggregated within the group. The effective retention results of the main interpretation load part and the qualification continuation part are extracted. Based on the degree of continuity of the effective retention results in the candidate segment group and the degree of correspondence with the query target representation, the response load strength corresponding to each candidate segment group is generated. The intrusion alignment results corresponding to each candidate fragment group are aggregated within the group to extract the effective intrusion results of the same-topic competition part and the version substitution part. Based on the effective intrusion results, the substitution degree of the main interpretation fragment and the qualification supplement fragment of the candidate fragment group is evaluated to generate the substitution intrusion intensity corresponding to each candidate fragment group. The forbidden response suppression part in the occupancy and alignment results corresponding to each candidate fragment group is aggregated separately, and the effective suppression results of the exception condition content, inapplicable scope content, invalid condition content, alternative restriction content and conflict restriction content are extracted. Based on the effective suppression results, the degree of restriction of the main interpretation fragment and the qualification supplementary evidence fragment of the candidate fragment group is used to generate the counter-evidence suppression strength corresponding to each candidate fragment group. The response carrying capacity, substitution encroachment capacity and counter-evidence suppression capacity of each candidate fragment group are jointly compared and processed. The effective main response portion of the response carrying capacity is retained, and the substitution portion corresponding to the substitution encroachment capacity and the restriction portion corresponding to the counter-evidence suppression capacity are deducted to generate the unique response potential difference for each candidate fragment group. The candidate cleavage groups are sorted from largest to smallest according to their unique response potential differences. The candidate cleavage group whose unique response potential difference meets the preset condition and is ranked first is selected as the target cleavage group.

[0028] In this embodiment, forming the minimum necessary chain of locking includes: Source evidence aggregation is performed around the target fragment group. The main interpretation fragment, qualification supplementary evidence fragment, and forbidden answer counter-evidence fragment in the target fragment group are read. Based on the qualification continuation relationship, counter-evidence suppression relationship, reference continuation relationship, version substitution relationship and chapter affiliation relationship in the fragment relationship graph, the source-related fragments corresponding to the target fragment group are extracted to generate a source evidence candidate set. Using the main interpretive fragment in the target fragment group as the interpretive source node, the qualification supplementary evidence fragment as the qualification continuation node, the forbidden answer counter-evidence fragment as the forbidden answer boundary node, and the source-related fragment as the source-backward node, a continuation link is established between the interpretive source node and the qualification continuation node based on the qualification continuation relationship. A suppression link is established between the interpretive source node and the forbidden answer boundary node based on the counter-evidence suppression relationship. A source-backward link is established between the interpretive source node, the qualification continuation node, the forbidden answer boundary node, and the source-backward node based on the citation continuation relationship, the version substitution relationship, and the chapter affiliation relationship. This generates a source-evidence closure spectrum, where: Based on the qualification continuation relationship, a continuation link is established between the source node and the qualification continuation node, specifically as follows: Extract the qualification continuation relationship between the main interpretation segment and each qualification supplementary segment in the target segment group, and record the starting interpretation node, target qualification node and continuation order information corresponding to each qualification continuation relationship; According to the order in which each qualification supplementary certificate fragment depends on the establishment of the main definition fragment, the qualification continuation relationship is arranged sequentially, and multiple qualification supplementary certificate fragments corresponding to the same main definition fragment are sequentially attached to the corresponding qualification continuation node. Using the source node corresponding to the main interpreting section as the starting point of the link, and the qualification continuation node corresponding to each qualification supplementary section as the link extension node, the continuation links between the source node and each qualification continuation node are established sequentially according to the order of the arrangement results, and the qualification continuation links corresponding to the target section group are generated. Based on citation continuation relationships, version replacement relationships, and chapter affiliation relationships, a back-to-source link is established between the definition source node, qualification continuation node, forbidden answer boundary node, and back-to-source node, specifically as follows: Extract the reference continuation relationship, version replacement relationship and chapter affiliation relationship associated with the fragments corresponding to the definition source node, qualification continuation node and forbidden answer boundary node, and determine the source associated fragments corresponding to each node; Based on the association direction between the source-related fragments and the source nodes of the definition, the qualification continuation nodes and the forbidden answer boundary nodes, the citation continuation positions, the version continuation positions and the chapter belonging positions are arranged in order to determine the connection order of each source node; Using the source node of interpretation, the qualification continuation node and the forbidden answer boundary node as the starting node of the back-to-source, and the back-to-source node corresponding to the source associated fragment as the ending node of the back-to-source, the back-to-source links between the source node of interpretation, the qualification continuation node, the forbidden answer boundary node and each back-to-source node are established in the order of connection, and the back-to-source links corresponding to the target fragment group are generated. The source certificate closure spectrum is subjected to interpretation closure and eligibility closure determination. The interpretation closure result is determined based on whether the main interpretation fragment corresponding to the interpretation source node completely covers the main retrieval line in the query target representation. The eligibility closure result is determined based on whether the eligibility supplementary certificate fragment corresponding to the eligibility continuation node completely covers the eligibility screening conditions in the query target representation. Specifically, the interpretation closure and eligibility closure determinations for the source certificate closure spectrum are performed as follows: Extract the main definition fragment content corresponding to the definition source node, match the main definition fragment content with the main retrieval line in the query target representation item by item, and identify the coverage of the retrieval object content, retrieval item content and result pointing content in the main definition fragment in the main definition fragment; Based on the completeness of the coverage of the main definition fragments of the content of the search object, the content of the search item, and the content pointed to by the result, the definition closure status corresponding to the definition source node is determined. The definition source node that can continuously cover all the content of the main search line is determined as the definition closure node, and the definition source node with missing object, missing item, or missing result is determined as the definition closure node, and the definition closure result is generated. Extract the qualification supplementary certificate fragment content corresponding to the qualification continuation node, match the qualification supplementary certificate fragment content with the qualification filtering conditions in the query target representation item by item, and identify the coverage of time-limited content, version-limited content, applicable object content, clause number content, field definition content, reference basis content and source-limited content in the qualification supplementary certificate fragment; Based on the complete coverage of each qualification screening condition by the qualification supplementary certificate fragment, the qualification locking status corresponding to the qualification continuation node is determined. The qualification continuation node that can continuously cover all qualification screening conditions is determined as the qualification locking established node, and the qualification continuation node with any missing qualification screening condition is determined as the qualification locking not established node, and the qualification locking result is generated. The source-proof locking spectrum is subjected to forbidden-answer locking and source-locking determination. The forbidden-answer locking result is determined based on whether the forbidden-answer counter-evidence fragment corresponding to the forbidden-answer boundary node forms a complete constraint on the boundary exclusion condition in the query boundary segment. The source-locking result is determined based on whether the source-related fragment corresponding to the source-backlink node forms a continuous back-source path from the target fragment group to the original knowledge object. Specifically, the forbidden-answer locking and source-locking determination for the source-proof locking spectrum are performed as follows: Extract the content of the forbidden answer rebuttal fragment corresponding to the forbidden answer boundary node, match the content of the forbidden answer rebuttal fragment with the boundary exclusion conditions in the query boundary segment, and identify the coverage of exception restriction content, inapplicable scope content, invalid restriction content, alternative restriction content, conflict restriction content and exclusion restriction content in the forbidden answer rebuttal fragment; Based on the complete coverage of each boundary exclusion condition by the forbidden answer proof split piece, the forbidden answer closure state corresponding to the forbidden answer boundary node is determined. The forbidden answer boundary node that can continuously cover all boundary exclusion conditions is determined as the forbidden answer closure established node, and the forbidden answer boundary node with any missing boundary exclusion condition is determined as the forbidden answer closure not established node, and the forbidden answer closure result is generated. Extract the source-related fragment content corresponding to the source node, and identify the source connection between the target fragment group and the original knowledge object based on the reference continuation relationship, version replacement relationship and chapter affiliation relationship. Record the reference continuation position, version continuation position and chapter affiliation position corresponding to each source node. Based on the continuity of connections and termination attribution between each source node, the source closure status is determined. Source nodes that can form a continuous source path from the target fragment group to the original knowledge object are identified as source closure established nodes, while source nodes with connection interruptions, missing inheritance, or missing termination attribution are identified as source closure not established nodes, and source closure results are generated. When the interpretation blockade result, qualification blockade result, prohibition blockade result, and source blockade result all meet the preset conditions, a progressive release process is performed on the source certificate blockade spectrum. Non-core nodes are released layer by layer in the order of source node, qualification continuation node, and prohibition boundary node. After each release, the interpretation blockade judgment, qualification blockade judgment, prohibition blockade judgment, and source blockade judgment are re-executed. Nodes that still maintain the preset conditions for all blockade results after release are identified as releaseable nodes. Nodes that cause any blockade result to no longer meet the preset conditions after release are identified as necessary blockade nodes. A minimum necessary blockade chain is formed based on all necessary blockade nodes, where: The preset conditions are as follows: in the definition closure results, the main definition fragment covers the content of the search object, the content of the search item, and the content of the result pointing to the query target representation; in the qualification closure results, the qualification supplementary certificate fragment covers the time limit content, version limit content, applicable object content, clause number content, field definition content, reference basis content, and source limit content in the query target representation; in the prohibited answer closure results, the prohibited answer counter-evidence fragment covers the exception limit content, inapplicable scope content, invalid limit content, alternative limit content, conflict limit content, and exclusion limit content in the query boundary fragment; and in the source closure results, the source association fragment forms a continuous back-source path from the target fragment group to the original knowledge object. The source evidence locking spectrum is subjected to progressive release processing, specifically as follows: Extract all source nodes, qualified continuation nodes, and forbidden response boundary nodes from the source certificate blocking spectrum. Based on the node type, node connection position, and node hierarchical position in the source certificate blocking spectrum, identify the non-core nodes in each node. Following the order of source node first, qualification continuation node in the middle, and forbidden answer boundary node last, the identified non-core nodes are released layer by layer. In each layer, the corresponding non-core nodes are released in order of connection distance between the node and the interpretation source node from farthest to closest. After each non-core node is released, the source certificate blocking spectrum after the release is reread, and the interpretation blocking judgment, qualification blocking judgment, prohibition blocking judgment and source blocking judgment are re-executed, and the preservation status of each blocking result is recorded. Nodes that, after release, still satisfy the preset conditions for the interpretation lock result, qualification lock result, prohibition lock result, and source lock result are determined as releaseable nodes. Nodes that, after release, cause any lock result to no longer satisfy the preset conditions are determined as lock-necessary nodes, and the release process of these lock-necessary nodes is stopped. After all non-core nodes have been released layer by layer, all necessary locking nodes and interpretation source nodes are retained, and the minimum necessary locking chain is generated according to the connection order between each necessary locking node and interpretation source node.

[0029] In this embodiment, the locking answer chain corresponding to the output minimum locking necessary chain includes: Read all the necessary locking nodes in the minimum necessary locking chain, and divide the necessary locking nodes into main answer nodes, qualified nodes, forbidden answer boundary nodes and source nodes according to the fragment type corresponding to each necessary locking node; The main answer node is used as the main answer fragment, the qualification supplement fragment is used as the qualification necessary fragment, the forbidden answer boundary fragment is used as the forbidden answer boundary fragment, and the source association fragment is used as the source necessary fragment. According to the connection order and relationship order of the main answer fragment, the necessary qualification fragment for closure, the necessary forbidden answer boundary fragment for closure, and the necessary source fragment for closure in the minimum necessary chain for closure, the link orchestration process is performed on each fragment. Based on the original knowledge object identifier, chapter path identifier, clause number identifier, version identifier and source identifier corresponding to each necessary source fragment for closure, the source attachment process is performed on the orchestrated link to generate the closure answer chain and the corresponding source tracing result. Output the locked answer chain, and simultaneously return the main answer fragment, the locked necessary qualification fragment, the locked necessary forbidden answer boundary fragment, the locked necessary source fragment, and the source tracing result corresponding to the locked answer chain.

[0030] Example 1: In a device operation and maintenance knowledge base retrieval scenario, the system receives a batch of data to be added to the database, including 412 policy documents, 286 operation guidelines, 173 maintenance manuals, 5240 Q&A records, and 318 sets of version revision records, with a total original text length of approximately 18.6 million words. After the data enters the processing flow, the system performs normalization processing, unifying title levels, clause numbers, version tags, and citation tags, removing 2317 duplicate paragraphs, 6841 invalid header and footer tags, and 11326 abnormal spaces and garbled characters, generating 6429 standard knowledge objects. After normalization, each knowledge object contains an average of 6.8 chapter levels, 14.3 clause fragments, and 2.1 explicit citation tags. To train the method of this invention, the system constructs training samples from historical retrieval logs and manually labeled results, forming a sample set of 23,800 sets of query-main definition fragments-qualification supplementary evidence fragments-prohibited answer counter-evidence fragments, including a training set of 19,040 sets, a validation set of 2,380 sets, and a test set of 2,380 sets; in addition, 23,800 sets of query-answer fragment pairs samples used in traditional methods are constructed to ensure consistent comparison conditions.

[0031] After the standard knowledge object enters the three-phase reversible fragmentation processing stage, the system first locates candidate fragmentation segments based on chapter switching, condition switching, version switching, and reference switching positions, identifying a total of 42,876 candidate segments. Then, it performs core localization, identifying 12,652 knowledge cores that can independently express rules, conclusions, or operational matters. Inner-circle convergence processing is performed around each knowledge core, yielding 12,652 main interpretation fragments with an average length of 83 characters. Next, outer-circle supplementary verification processing is performed, extracting time limits, version limits, applicable objects, numbering conditions, and reference bases, yielding 18,437 qualification supplementary verification fragments with an average length of 41 characters. Finally, boundary sealing processing is performed, extracting exceptional conditions, inapplicable scope, substitution relationships, and invalidation conditions, yielding 9,736 prohibited answer counter-evidence fragments with an average length of 37 characters. The system further established a fragment relationship graph, forming 21,984 qualification continuation edges, 11,208 counter-evidence suppression edges, 10,516 boundary verification edges, 14,873 reference continuation edges, and 4,261 version substitution edges. Sampling inspection revealed that, on average, a single knowledge unit in the fragment graph connects to 2.7 qualification supplementary fragments and 1.3 prohibited counter-evidence fragments, indicating that complex knowledge no longer exists as isolated fragments.

[0032] In a simulated search, the operator input the following question: Can the system be directly reset after an abnormal shutdown? What conditions must be met under the current rules, and in what situations is execution prohibited? The system first breaks down the request, resulting in 11 query units, retaining 3 subject terms, 4 qualifying terms, 2 exclusion terms, and 2 version-related expressions. After anchoring the main idea, a 26-character main definition fragment is generated; after qualification attachment, 3 qualification fragments are generated, corresponding to "current rule," "pre-execution conditions," and "applicable objects," respectively; after boundary convergence, 2 query boundary fragments are generated, corresponding to "cannot be executed" and "exceptional situations," respectively. Subsequently, a query mirror body is constructed, containing 1 main mirror node, 3 qualification attachment nodes, and 2 boundary constraint nodes. From this, 1 main search line, 3 qualification screening conditions, 2 boundary exclusion conditions, 1 version location information, and 1 applicable scope location information are extracted to form the query target representation. Compared with traditional methods, which only extract 7 keywords and establish a single vector representation without explicitly distinguishing between qualification conditions and boundary restrictions, this approach yields superior results.

[0033] The query target indicates entry into a dual-channel response competition model based on an improved RetNet. The system first recalls 47 candidate fragment groups from the fragment relationship graph, including 18 groups involving same-topic competition, 9 groups involving version substitution, and 12 groups involving explicit prohibition of response. The response carrying channel reads 96 main interpretation fragments and qualification supplementary evidence fragments, forming a response input sequence in relational order, with an average sequence length of 5.1 fragments. The improved RetNet maintains a strong response to executable reset-related main cores in the main interpretation retention unit, with a mean carrying value of 0.84. In the qualification continuation retention unit, it maintains responses to "self-check completed," "protection not triggered," and "read confirmation field," with a mean qualification continuation value of 0.79. The substitution encroachment channel simultaneously reads 71 ​​prohibition of response counter-evidence fragments, version substitution fragments, and same-topic competition fragments, competing to maintain the old version rules, exceptions, and restrictions. The mean old version encroachment value is 0.63, and the mean prohibition of response suppression value is 0.74. After relational masking scheduling and competitive write-back, the response carrying strength of the target fragment group is 0.88, the substitution encroachment strength is 0.29, and the counter-evidence suppression strength is reduced from 0.74 to 0.21 after boundary completion. The final unique response potential difference is 0.59. The highest similarity fragment score of the traditional method is 0.91, but it corresponds to the old version fragment and does not include the forbidden response boundary.

[0034] Around the target fragment group, the system continues to construct the source evidence blocking spectrum. The source evidence candidate set includes 1 main definition fragment, 4 qualification supplementary evidence fragments, 2 prohibited response counter-evidence fragments, and 3 source association fragments. During the first blocking determination, the definition blocking completeness rate was 100%, and the qualification blocking completeness rate was 83.3%, due to a lack of applicable control unit conditions. After the system added one qualification supplementary evidence fragment along the reference continuation edge, the qualification blocking completeness rate increased to 100%. Subsequently, a prohibited response blocking determination was performed, finding that the "unreset is not allowed if the chain is not released" rule was already covered, but the "prohibition of reset for thermal protection alarms" rule was not included. After adding this rule, the prohibited response blocking completeness rate increased from 71.5% to 100%. In the source blocking determination, the initial back-to-source path length was 2 layers and not connected to the version revision source node. After adding one version source fragment, a 4-layer continuous back-to-source path was formed. After completing the four types of locking determination, the system executes a progressive release mechanism. Starting from the outer source node, it releases three non-core nodes, two of which retain the locking after release, while the source locking is interrupted after release of one. Next, it releases two weakly related nodes from the qualifying node, one of which is releasable and the other is a locking-necessary node. This ultimately forms the minimum locking necessity chain, consisting of one main answer fragment, three locking necessity qualifying fragments, two locking necessity forbidden answer boundary fragments, and two locking necessity source fragments. After the system outputs the locking answer chain, manual verification takes 2.8 seconds, while traditional methods require manual cross-checking of the first five results, averaging 11.6 seconds.

[0035] On a test set of 2380 questions, the method of this invention was compared with the traditional keyword plus vector fusion method. The main answer fragment hit rate increased from 81.7% to 95.2%, the completeness rate of qualification conditions increased from 62.9% to 91.8%, the forbidden answer boundary recognition rate increased from 57.4% to 89.6%, the source backtracking completeness rate increased from 60.8% to 96.1%, and the false hit rate of the old version decreased from 19.3% to 2.7%. For 812 questions involving multi-slice joint answers, the traditional method returned an average of 5.8 candidate fragments before a usable answer could be manually constructed, while the method of this invention output an average of one closed answer chain with 3.4 necessary fragments within the chain. After training, the overall answer validity rate on the test set reached 93.7%, compared to 76.5% for the traditional method.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A knowledge base slice retrieval method based on AI dual-channel representation, characterized in that, include: Receive raw data from the knowledge base, perform parsing processing on the raw data, extract document structure attribute information and content attribute information, and generate a standard knowledge object set; Three-phase reversible fragmentation is performed on each standard knowledge object. Knowledge units are divided according to semantic boundaries, clause boundaries, condition boundaries, version boundaries, and reference boundaries. For each knowledge unit, main interpretation fragments, qualification supplementary evidence fragments, and forbidden answer counter-evidence fragments are generated, and a fragmentation relationship graph is generated. Receive a retrieval request, perform subject extraction, qualification extraction and boundary extraction on the retrieval request, generate query subject definition fragment, query qualification fragment and query boundary fragment, construct query mirror body, and generate query target representation based on query mirror body; A dual-channel response competition model based on an improved RetNet is constructed. The main interpretation fragment, the qualification supplementary evidence fragment, and the forbidden answer counter-evidence fragment are input into the dual-channel response competition model. The same question competition relationship and version substitution relationship are extracted by combining the fragment relationship graph. The unique response potential difference is calculated and the target fragment group is determined. A source evidence locking spectrum is constructed around the target fragment group. Interpretive locking, qualification locking, prohibition of answer locking and source locking are performed on the source evidence locking spectrum. When each locking judgment meets the preset conditions, progressive release processing is performed to identify the necessary locking nodes and form the minimum necessary locking chain. Output the chain of closure answers corresponding to the minimum closure necessary chain, and return the main answer fragment, the closure necessary qualification fragment, the closure necessary forbidden answer boundary fragment, and the closure necessary source fragment.

2. The knowledge base slice retrieval method based on AI dual-channel representation according to claim 1, characterized in that, The original data in the knowledge base includes policy documents, technical specifications, product manuals, operation guidelines, Q&A records, work order records, database text records, web page data, and version update records.

3. The knowledge base slice retrieval method based on AI dual-channel representation according to claim 1, characterized in that, The document structure attribute information includes document title, chapter path, clause number, citation mark, version identifier, time identifier, source identifier, and permission identifier. The content attribute information includes the main text content and table fields.

4. The knowledge base slice retrieval method based on AI dual-channel representation according to claim 1, characterized in that, The process of generating a standard knowledge object set involves performing character normalization, number normalization, version normalization, and reference normalization on document structure attribute information and content attribute information to generate a standard knowledge object set.

5. The knowledge base slice retrieval method based on AI dual-channel representation according to claim 1, characterized in that, The process involves generating main definition fragments, qualification supplementary evidence fragments, and forbidden answer counter-evidence fragments for each knowledge unit, and generating a fragment relationship graph, including: Perform core positioning processing on each standard knowledge object, divide candidate fragments into segments based on chapter level change positions, clause switch positions, condition switch positions, version switch positions, reference switch positions, and semantic transition positions, identify core statement groups that can independently express rules, definitions, conclusions, or operational matters in each candidate fragment segment, and generate corresponding knowledge cores; The inner circle convergence process is performed with each knowledge core as the center. The main conclusion content, main definition content, main rule content or main operation content directly corresponding to the knowledge core are extracted along the forward and backward content of the knowledge core. They are merged according to the continuous expression relationship to generate the corresponding main interpretation fragment. The outer ring of supplementary certification is processed around each main definition fragment. The time limit, version limit, applicable object, clause number, field definition and reference content that have a dependency relationship with the main definition fragment are retrieved and merged according to the same establishment condition relationship to generate the corresponding qualification supplementary certification fragment. Boundary sealing is performed around each main interpretation fragment. Exceptions, inapplicable scope, invalidation conditions, substitution relationships, conflicting restrictions, and abolished related content that have a restrictive relationship with the main interpretation fragment are retrieved and merged according to the same restriction boundary attribution relationship to generate forbidden answer counter-evidence fragments. Using the main interpretation fragment, qualification supplementary evidence fragment, and forbidden answer counter-evidence fragment as fragment nodes, qualification continuation edges are established based on the establishment dependency relationship, counter-evidence suppression edges are established based on the restriction relationship, boundary verification edges are established based on the boundary correspondence relationship, reference continuation edges are established based on the reference inheritance relationship, and version replacement edges are established based on the version inheritance relationship, thus generating a fragment relationship graph.

6. The knowledge base slice retrieval method based on AI dual-channel representation according to claim 1, characterized in that, The construction of the query mirror body and the generation of the query target representation based on the query mirror body include: Upon receiving a retrieval request, the system performs a normalization and segmentation process on the retrieval request, dividing it into multiple query expression units according to pause marks, conjunctions, qualifiers, exclusion words, number markers, time markers, version markers, and source markers, and performs order preservation processing on each query expression unit to generate a sequence of query units arranged in the original expression order; Perform thematic anchoring processing on the query unit sequence to identify the query object content, query item content and query result content that represent the core retrieval intent. Merge query expression units that point to the same retrieval thematic intent to generate query thematic explanatory fragments. The eligibility linking process is performed around the main definition fragment of the query. It retrieves the time limit content, version limit content, applicable object content, clause number content, field definition content, reference content and source limit content that have an establishment dependency relationship with the main definition fragment of the query, and merges them according to the same establishment condition relationship to generate query eligibility fragments. Boundary merging is performed around the main definition fragment of the query, retrieving exceptions, inapplicable content, invalid content, alternative content, conflicting content, and excluded content that have a restrictive relationship with the main definition fragment of the query, and merging them according to the same restriction boundary attribution relationship to generate query boundary fragments; Using the main definition fragment of the query as the main mirror node, the query qualification fragment as the qualification attachment node, and the query boundary fragment as the boundary constraint node, a query mirror body is constructed according to the subject correspondence, qualification dependency relationship and boundary restriction relationship. The main retrieval thread, qualification screening conditions, boundary exclusion conditions, number positioning information, version positioning information, time positioning information, source positioning information and applicable scope positioning information are extracted from the query mirror body to generate the query target representation.

7. The knowledge base slice retrieval method based on AI dual-channel representation according to claim 1, characterized in that, The process of extracting competing relationships and version substitution relationships by combining the fragment relationship map, calculating the unique response potential difference, and determining the target fragment group includes: Read the query target representation, main definition fragments, qualification supplementary evidence fragments, prohibited answer counter-evidence fragments, and fragment relationship graph. Based on the main retrieval line, qualification screening conditions, boundary exclusion conditions, number positioning information, version positioning information, time positioning information, source positioning information, and applicable scope positioning information in the query target representation, extract the candidate fragment set corresponding to the retrieval request. Group the candidate fragment set according to the relationship of the same question, the relationship of qualification continuation, the relationship of version substitution, and the relationship of counter-evidence suppression to generate candidate fragment groups. A dual-channel response competition model based on an improved RetNet is constructed, with RetNet as the reserved coding backbone. The dual-channel response competition model sets up a response carrying channel and a substitute response encroachment channel. The response carrying channel sets up a main interpretation reservation unit and a qualification continuation reservation unit. The substitute response encroachment channel sets up a same-topic competition reservation unit, a version substitution reservation unit, and a counter-evidence suppression reservation unit. A relationship masking scheduling unit and a competition write-back unit are set up between the response carrying channel and the substitute response encroachment channel. Based on the main definition fragments, qualification supplementary fragments, and the main retrieval line and qualification screening conditions in the query target representation of the candidate fragment group, a response input sequence is constructed. The response input sequence is input into the response carrying channel. The core definition content in the main definition fragment is retained by the main definition retention unit, and the establishment condition content in the qualification supplementary fragment is retained by the qualification continuation retention unit. The carrying vector corresponding to each candidate fragment group is generated. Based on the forbidden answer counter-evidence fragments in the candidate fragment groups and the same-topic association, version substitution relationship and counter-evidence suppression relationship in the fragment relationship graph, an invasion input sequence is constructed. The invasion input sequence is input into the substitute answer invasion channel. The same-topic competition retention unit retains the same-topic competition content, the version substitution retention unit retains the version substitution content, and the counter-evidence suppression retention unit retains the forbidden answer suppression content, thereby generating the invasion vector corresponding to each candidate fragment group. The relational masking scheduling unit performs relational constraint alignment on the bearer vector and the encroachment vector. The competitive write-back unit writes the substitution suppression result in the encroachment vector back to the bearer vector and the qualification continuation result in the bearer vector back to the encroachment vector. The response bearer strength, substitution encroachment strength, counter-evidence suppression strength and unique response potential difference are generated for each candidate shard group. The target shard group is determined based on the unique response potential difference.

8. The knowledge base slice retrieval method based on AI dual-channel representation according to claim 1, characterized in that, The formation of the minimum necessary chain of locking includes: Source evidence aggregation is performed around the target fragment group. The main interpretation fragment, qualification supplementary evidence fragment, and forbidden answer counter-evidence fragment in the target fragment group are read. Based on the qualification continuation relationship, counter-evidence suppression relationship, reference continuation relationship, version substitution relationship and chapter affiliation relationship in the fragment relationship graph, the source-related fragments corresponding to the target fragment group are extracted to generate a source evidence candidate set. Using the main interpretive fragment in the target fragment group as the interpretive source node, the qualification supplementary evidence fragment as the qualification continuation node, the forbidden answer counter-evidence fragment as the forbidden answer boundary node, and the source-related fragment as the source return node, a continuation link is established between the interpretive source node and the qualification continuation node based on the qualification continuation relationship, a suppression link is established between the interpretive source node and the forbidden answer boundary node based on the counter-evidence suppression relationship, and a source return link is established between the interpretive source node, the qualification continuation node, the forbidden answer boundary node and the source return node based on the citation continuation relationship, the version substitution relationship and the chapter affiliation relationship, thereby generating the source evidence closure spectrum; Perform interpretation closure and qualification closure on the source certificate closure spectrum. The interpretation closure result is determined based on whether the main interpretation fragment corresponding to the interpretation source node completely covers the main retrieval line in the query target representation. The qualification closure result is determined based on whether the qualification supplementary certificate fragment corresponding to the qualification continuation node completely covers the qualification screening conditions in the query target representation. Perform forbidden answer blocking and source blocking on the source evidence blocking spectrum. Determine the forbidden answer blocking result based on whether the forbidden answer counter-evidence fragment corresponding to the forbidden answer boundary node forms a complete restriction on the boundary exclusion condition in the query boundary segment. Determine the source blocking result based on whether the source association fragment corresponding to the source node forms a continuous source path from the target fragment group to the original knowledge object. When the interpretation of the locking result, the qualification of the locking result, the prohibition of answer locking result, and the source locking result all meet the preset conditions, the source certificate locking spectrum is processed by progressive release. Non-core nodes are released layer by layer in the order of source node, qualification continuation node, and prohibition of answer boundary node. After each release, the interpretation of the locking result, qualification of the locking result, prohibition of answer locking result, and source locking result are re-executed. Nodes that still meet the preset conditions after release are identified as releaseable nodes. Nodes that cause any locking result to fail to meet the preset conditions after release are identified as locking necessary nodes. The minimum locking necessary chain is formed based on all locking necessary nodes.

9. The knowledge base slice retrieval method based on AI dual-channel representation according to claim 1, characterized in that, The chain of locking answers corresponding to the minimum necessary chain of locking output includes: Read all the necessary locking nodes in the minimum necessary locking chain, and divide the necessary locking nodes into main answer nodes, qualified nodes, forbidden answer boundary nodes and source nodes according to the fragment type corresponding to each necessary locking node; The main answer node is used as the main answer fragment, the qualification supplement fragment is used as the qualification necessary fragment, the forbidden answer boundary fragment is used as the forbidden answer boundary fragment, and the source association fragment is used as the source necessary fragment. According to the connection order and relationship order of the main answer fragment, the necessary qualification fragment for closure, the necessary forbidden answer boundary fragment for closure, and the necessary source fragment for closure in the minimum necessary chain for closure, the link orchestration process is performed on each fragment. Based on the original knowledge object identifier, chapter path identifier, clause number identifier, version identifier and source identifier corresponding to each necessary source fragment for closure, the source attachment process is performed on the orchestrated link to generate the closure answer chain and the corresponding source tracing result. Output the locked answer chain, and simultaneously return the main answer fragment, the locked necessary qualification fragment, the locked necessary forbidden answer boundary fragment, the locked necessary source fragment, and the source tracing result corresponding to the locked answer chain.

Images