Conversational AI Query Method Based on Enterprise Knowledge Base

Abstract

This invention discloses a conversational AI query method based on an enterprise knowledge base, relating to the field of artificial intelligence technology. The method includes S1: receiving query text input by a user, reading the current session history and user permission domain information, and standardizing the query text to form a standardized query representation; performing initial recall of knowledge fragments within the user permission domain to obtain a set of seed fragments related to the current query and with consistent permissions. The input to this step is the user query, session history, and permission domain information, and the output is the set of seed fragments and the set of entity candidates involved in the seed fragments. This invention transforms shared entities from expandable associated nodes into broken-bridge projection nodes that can only return structured constraint results, enabling the system to introduce only cross-domain relationship constraints without introducing cross-domain original text content in shared entity scenarios.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a conversational AI query method based on an enterprise knowledge base. Background Technology

[0002] As enterprise knowledge bases expand, the relationships between knowledge objects increase. The same supplier, project code, equipment identifier, interface name, or compliance entry may be repeatedly referenced across different business domains. To adapt to this knowledge organization, some enterprise knowledge base query systems adopt a hybrid retrieval architecture combining vector recall and relationship expansion. This involves first recalling initial knowledge fragments from the user's visible scope, and then further expanding the retrieval context around the entities and relationships involved, thereby improving the ability to handle complex questions, follow-up questions, and cross-document questions.

[0003] When the system adopts a hybrid retrieval method of initial knowledge fragment recall + entity association expansion, and the initial knowledge fragment contains cross-departmental or cross-tenant shared entities, the existing query process cannot retain the relational facts and constraint information corresponding to the shared entity without introducing the original content of the external permission domain. This would prevent the shared entity from becoming a springboard for cross-domain retrieval and ensure that the system output results meet the permission boundary requirements while maintaining usable question-and-answer capabilities and a clear processing chain. Summary of the Invention

[0004] The purpose of this invention is to provide a conversational AI query method based on an enterprise knowledge base to solve the problems mentioned in the background art.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a conversational AI query method based on an enterprise knowledge base, comprising: S1: receiving query text input by a user, reading the current session history and user permission domain information, standardizing the query text to form a standardized query representation; performing initial recall of knowledge fragments within the user permission domain to obtain a set of seed fragments related to the current query and with consistent permissions, wherein the input of this step is the user query, session history, and permission domain information, and the output is the set of seed fragments and the set of entity candidates involved in the seed fragments; S2: Perform sharing analysis and bridging risk analysis on the entity candidates in the seed fragment set to identify whether there are shared entities that are simultaneously distributed in multiple permission domains and may trigger the expansion of external permission domains; when a shared entity meets the bridging risk condition, the shared entity is identified as a bridging entity. The input of this step is the seed fragment set and the entity candidate set, and the output is the bridging entity set. S3: Establish a broken bridge projection node for the bridging entity, so that the bridging entity only participates in subsequent processing as a restricted intermediate object in the current query process; the broken bridge projection node initiates a constraint request to the external permission domain and receives a constraint capsule without original text content. The input of this step is the bridging entity set and the evidence index of this permission domain, and the output is the broken bridge projection node set and the constraint capsule set. The broken bridge projection node is a restricted intermediate data object used to replace the shared entity to continue cross-domain original text expansion. The constraint capsule is a data object returned by the external permission domain that only contains structured constraint information and does not contain original text content. S4: Merge the evidence fragments of this permission domain with the constraint capsules to form a restricted response context, and determine whether the current query meets the respondability conditions based on the sufficiency of evidence, the consistency of constraints, and the degree of bridging risk; if the respondability conditions are met, proceed to the response generation stage; if the respondability conditions are not met, proceed to the approval transfer stage. The input of this step is the seed fragment set and the constraint capsule set, and the output is the restricted response context and the respondability determination result. S5: Output the corresponding text based on the answerability determination result. In the case of direct response, output the complete answer based on the evidence of this authority domain. In the case of restricted response, output the restricted answer containing relational conclusions and binding statements. In the case of approval transfer, output the approval prompt information. At the same time, record the complete processing link. The input of this step is the restricted response context and the answerability determination result. The output is the final response result and audit record.

[0006] According to the above technical solution, step S1 specifically includes the following sub-steps: S1-1: User query text Current session history summary and user permission domain identifier set Perform unified encoding to obtain a standardized query vector. ,in The query normalization encoding function is used to encode text semantic information, session continuity information and permission boundary information into the same query representation. This formula incorporates query semantics and permission boundaries into subsequent retrieval input in a synchronized manner, avoiding subsequent processing that ignores permission constraints based solely on text semantics. S1-2: For each candidate knowledge fragment Calculate the initial recall score ,in Represents the standardized query vector With candidate knowledge fragment vectors The semantic similarity is used to characterize the degree of semantic fit between candidate knowledge fragments and the current query; This indicates the degree of explicit matching between query keywords and candidate knowledge fragment text, used to characterize the consistency between technical terms and fixed appellations. Represents the set of user permission domain identifiers The degree of consistency between the candidate knowledge fragment and its permission tag is used to characterize whether the candidate knowledge fragment is within the current accessible boundary. , , This represents the corresponding weight, and the sum of the three is 1. This formula is used to simultaneously consider semantic relevance, term matching, and permission consistency to uniformly rank candidate knowledge fragments. S1-3: Select those that satisfy... The candidate knowledge fragments are used as seed fragments, among which This represents the initial recall threshold, used to control the minimum relevance requirement for seed fragments to enter subsequent bridging analysis; and extracts entity name, entity type, and entity location from the seed fragments to form an entity candidate set. The output of this sub-step is the seed fragment set and the entity candidate set corresponding to the seed fragments.

[0007] According to the above technical solution, step S2 specifically includes the following sub-steps: S2-1: For each entity candidate Calculate cross-domain distribution degree Unauthorized Adjacency Ratio and query dependency ,in , , ,in Indicates entity candidate The actual number of permission domains distributed across the knowledge graph. This indicates the maximum number of preset permission domains in the system, used to normalize the cross-domain distribution range. Indicates entity candidate Adjacent to each other and not belonging to the user permission domain identifier set The number of adjacent nodes, Indicates entity candidate The total number of all adjacent nodes. Indicates entity candidate entity description vector, The cosine similarity function is represented by the three formulas above, which are used to quantify the cross-domain diffusion range of entity candidates, the tendency of unauthorized expansion, and the degree of dependence of the current query on the entity, respectively. S2-2: Based on the cross-domain distribution degree Unauthorized Adjacency Ratio and query dependency Calculate bridging risk score ,in , , The formula represents the bridging risk weight, and the sum of the three factors is 1. The significance of this formula is to uniformly quantify three factors: whether the shared entity is easy to distribute across domains, whether it is easy to continue to connect to unauthorized neighboring domains, and whether it is strongly related to the current query, so as to identify the degree of bridging risk of the shared entity in the current query. S2-3: When At that time, entity candidates It is identified as a bridging entity and is prohibited from triggering external permission domain plaintext expansion during the current query process. Retain its normal related search capabilities within the current permission domain, among which The output of this sub-step is the set of bridging entities, which serves as the threshold for triggering bridging risks.

[0008] According to the above technical solution, step S3 specifically includes the following sub-steps: S3-1: For each bridging entity Constructing the projection node of the broken bridge The bridge break projection node It only includes the evidence index of this permission domain, the cross-domain constraint request identifier, and the location where the constraint result is received, and does not include the original text fragments of external permission domains; and it is based on the seed fragment vector set corresponding to the bridging entity within this permission domain. Calculate the evidence center vector in this domain ,in This indicates the connection between the current user permission domain and the bridging entity. The number of associated seed fragments: This formula uses evidence fragments directly related to the bridging entity within the current authority domain to form a local evidence center, providing a stable semantic benchmark for subsequent constraint fusion. S3-2: Projection node of broken bridge Send a cross-domain constraint request to the local policy engine of the external permission domain and receive the constraint capsule. The constraint capsule At least include cross-domain dependency presence markers Risk level value Consistency with time ,in This indicates that the external permission domain returns the bridging entity. The effective time interval of the relevant constraints This represents the target time window corresponding to the current query. The significance of this formula lies in quantifying the consistency between the external permission domain's returned constraints and the current query's time range, thereby preventing the system from introducing relational conclusions that do not match the current query in the time dimension. The risk level value represents the length of the corresponding time interval. The external permission domain's local policy engine outputs a risk classification rule based on preset risk classification rules. This preset risk classification rule maps the risk level of cross-domain dependencies to ordered numerical values. When a cross-domain dependency exists... The value is 1 if it is not 0 otherwise. The external permission domain local policy engine is a software processing unit deployed within the external permission domain, used to calculate the structured constraint results based on the local permission policy and risk rules.

[0009] S3-3: A marker exists based on the cross-domain dependency. Risk level value Consistency with time Calculate the constraint capsule score ,in , , The formula represents the weight of the constraint capsule, and the sum of the three is 1. The significance of the formula is to uniformly quantify three factors: whether there is a cross-domain dependency, what risk level the cross-domain dependency is at, and whether the dependency is valid within the current query time window, so as to form a structured constraint result that can be directly called in the subsequent restricted response stage.

[0010] According to the above technical solution, step S4 specifically includes the following sub-steps: S4-1: The evidence fragments of this authority domain in the seed fragment set that meet the initial recall requirements are fused with the constraint capsules corresponding to each bridging entity to form a restricted response context, wherein the restricted response context only contains the evidence content of this authority domain and the structured constraint results, and does not contain the original text fragments of the external authority domain. S4-2: Calculate the resolvability score for the current query. ,in This indicates the initial recall score of evidence fragments within the current authority domain that are incorporated into the context of the restricted response. The average value is used to characterize the overall relevance of evidence within this authority domain. The constrained capsule score represents the score that incorporates the context of the limited response. The average value is used to characterize the overall availability of cross-domain relationship constraints. This represents the bridging risk score for all bridging entities in the current query. The maximum value is used to characterize the highest cross-domain expansion risk faced by the current query. , , The formula indicates that the system should only proceed to the response generation stage when the evidence in the current authority domain is sufficiently sufficient, the constraint capsule is sufficiently consistent, and the highest bridging risk is effectively suppressed. S4-3: When When the current query meets the conditions for being respondable, it is determined that the query is ready to be answered. If the current query does not meet the criteria for being answered, it will be transferred to the approval process. This indicates the respondability threshold, and the output of this sub-step is the respondability determination result.

[0011] According to the above technical solution, step S5 specifically includes the following sub-steps: S5-1: Count the number of bridging entities corresponding to the current query. There is a labeled mean for cross-domain dependencies. ,in Indicates the first Each bridging entity corresponds to a cross-domain dependency in the constraint capsule. The significance of the formula is to quantify the overall occurrence of cross-domain dependencies in the current query, thereby providing a basis for the final output type selection. S5-2: Select the output type according to the following rules: When and When, a direct response result is output, which is generated solely based on evidence fragments from this permission domain; when and When the time comes, a restricted response result is output. This restricted response result includes verifiable conclusions within this permission domain, an explanation of the existence of cross-domain dependencies, and a risk level description. When the approval transfer result is output, the approval transfer result shall at least include the transfer reason, the object to be approved and the suggested verification scope. Without introducing the original text fragments of external permission domains, the respondability determination and the generation of restricted responses shall be jointly supported by the evidence of this permission domain and the constraint capsule. S5-3: Query text User permission domain identifier set Seed fragment set, bridging entity set, bridging risk score Constraint capsules Responsibility score The output type and final response result are written into the audit log to form a complete and traceable processing chain. The purpose of this sub-step is to ensure that the input source, boundary control process and output basis can be traced back for each question and answer.

[0012] According to the above technical solution, the system used in this method includes: The query boundary construction module is used to receive user query content, combine the current session context and user permission domain information, perform normalization processing on the query, and form an initial set of candidate knowledge fragments within the user permission domain; The broken bridge projection retrieval module is used to identify shared entities from the initial candidate knowledge fragment set, establish broken bridge projection nodes for shared entities that meet the bridging risk conditions, so that the shared entities no longer trigger the original text expansion of the external permission domain, and obtain the structured constraint results without original text content from the external permission domain according to the broken bridge projection nodes. The restricted output module is used to fuse evidence fragments within the user's permission domain with the structured constraint results, determine whether the current query meets the conditions for being respondable, and output a direct response result, a restricted response result, or an approval transfer result, while recording the complete processing chain.

[0013] According to the above technical solution, the query boundary construction module includes: a query normalization unit, used to uniformly encode the user's current input, current session history, and user permission domain identifier to form a standardized query representation; a permission consistency recall unit, used to sort candidate knowledge fragments and output an initial seed fragment set within the range where the knowledge fragment permission tags are consistent with the user permission domain; and a seed entity extraction unit, used to extract entity names, entity types, and the associated positions of entities in seed fragments from the initial seed fragment set to form a shared entity candidate set. The broken bridge projection retrieval module includes: a bridging risk determination unit, used to identify whether to determine the corresponding shared entity as a bridging entity based on the cross-domain distribution of the shared entity, the unauthorized adjacency situation, and the query-related situation; a projection node construction unit, used to generate broken bridge projection nodes for bridging entities, wherein the broken bridge projection nodes include the evidence index of the current permission domain, the cross-domain constraint request identifier, and the constraint result receiving location, and do not include the original text fragment of the external permission domain; and a constraint capsule acquisition unit, used to send a cross-domain constraint request to the local policy engine of the external permission domain and receive the structured constraint result, wherein the structured constraint result includes at least a cross-domain dependency existence marker, a risk level value, and a time validity consistency, and does not include the original text of the external permission domain, the original entity list, and the full-text summary; The restricted output module includes: an evidence fusion unit, used to uniformly organize evidence fragments and constraint capsules of the current permission domain to form a restricted response context; a respondability determination unit, used to determine the output type corresponding to the current query based on the sufficiency of evidence in the current permission domain, the consistency of constraint capsules, and the intensity of bridging risk; a result generation unit, used to generate direct response text, restricted response text, or approval transfer text according to the output type; and a link recording unit, used to record query input, permission domain identifier, seed fragment, shared entity, bridging risk determination result, constraint capsule, output type, and final result.

[0014] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: In the conversational AI query scenario of enterprise knowledge base, the present invention transforms the shared entity from a traversable node in the traditional hybrid retrieval chain into a restricted intermediate object. Under the premise that the initial knowledge fragment still comes from the current permission domain, a broken bridge projection node is constructed for the entity that triggers the shared boundary, and the external permission domain only returns a constraint capsule without the original text. Therefore, it can block the direct extension chain of the fragment in this domain - the shared entity - the original text fragment in the external domain from the system structure.

[0015] With this structure, the system can retain relational information such as whether cross-domain dependencies exist, the level of dependency risk, and whether they are valid in time, so that users can still obtain executable and reviewable answers in shared entity scenarios. At the same time, the system only uses evidence and constraint capsules of the current permission domain to generate limited answers during the answer generation stage, and combines answerability determination to achieve three types of output: direct answer, limited answer, and approval transfer. Therefore, it can fully cover the requirements of permission boundary control, relational fact preservation, question and answer availability maintenance, and clear and reproducible processing links in the above-mentioned technical issues. Attached Figure Description

[0016] 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 schematic diagram of the overall modular structure of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 The present invention provides a technical solution: a conversational AI query method based on an enterprise knowledge base, comprising: S1: receiving query text input by a user, reading the current session history and user permission domain information, performing normalization processing on the query text to form a standardized query representation; performing initial recall of knowledge fragments within the user permission domain to obtain a set of seed fragments related to the current query and with consistent permissions, the input of this step being the user query, session history and permission domain information, and the output being the set of seed fragments and the set of entity candidates involved in the seed fragments; S2: Perform sharing analysis and bridging risk analysis on the entity candidates in the seed fragment set to identify whether there are shared entities that are simultaneously distributed in multiple permission domains and may trigger the expansion of external permission domains; when a shared entity meets the bridging risk condition, the shared entity is identified as a bridging entity. The input of this step is the seed fragment set and the entity candidate set, and the output is the bridging entity set. S3: Establish a broken bridge projection node for the bridging entity, so that the bridging entity only participates in subsequent processing as a restricted intermediate object in the current query process; the broken bridge projection node initiates a constraint request to the external permission domain and receives a constraint capsule without original text content. The input of this step is the set of bridging entities and the evidence index of this permission domain, and the output is the set of broken bridge projection nodes and the set of constraint capsules. The broken bridge projection node is a restricted intermediate data object used to replace the shared entity to continue cross-domain original text expansion. The constraint capsule is a data object returned by the external permission domain that only contains structured constraint information and does not contain original text content. S4: Merge the evidence fragments of this permission domain with the constraint capsules to form a restricted response context, and determine whether the current query meets the respondability conditions based on the sufficiency of evidence, the consistency of constraints, and the degree of bridging risk; if the respondability conditions are met, proceed to the response generation stage; if the respondability conditions are not met, proceed to the approval transfer stage. The input of this step is the seed fragment set and the constraint capsule set, and the output is the restricted response context and the respondability determination result. S5: Output the corresponding text based on the answerability determination result. In the case of direct response, output the complete answer based on the evidence of this authority domain. In the case of restricted response, output the restricted answer containing relational conclusions and binding statements. In the case of approval transfer, output the approval prompt information. At the same time, record the complete processing link. The input of this step is the restricted response context and the answerability determination result. The output is the final response result and audit record. Step S1 specifically includes the following sub-steps: S1-1: User query text Current session history summary and user permission domain identifier set Perform unified encoding to obtain a standardized query vector. ,in The query normalization encoding function is used to encode text semantic information, session continuity information and permission boundary information into the same query representation. This formula incorporates query semantics and permission boundaries into subsequent retrieval input in a synchronized manner, avoiding subsequent processing that ignores permission constraints based solely on text semantics. S1-2: For each candidate knowledge fragment Calculate the initial recall score ,in Represents the standardized query vector With candidate knowledge fragment vectors The semantic similarity is used to characterize the degree of semantic fit between candidate knowledge fragments and the current query; This indicates the degree of explicit matching between query keywords and candidate knowledge fragment text, used to characterize the consistency between technical terms and fixed appellations. Represents the set of user permission domain identifiers The degree of consistency between the candidate knowledge fragment and its permission tag is used to characterize whether the candidate knowledge fragment is within the current accessible boundary. , , This represents the corresponding weight, and the sum of the three is 1. This formula is used to simultaneously consider semantic relevance, term matching, and permission consistency to uniformly rank candidate knowledge fragments. S1-3: Select those that satisfy... The candidate knowledge fragments are used as seed fragments, among which This represents the initial recall threshold, used to control the minimum relevance requirement for seed fragments to enter subsequent bridging analyses; and it extracts entity names, entity types, and entity locations from seed fragments to form an entity candidate set. The output of this sub-step is the seed fragment set and the entity candidate set corresponding to the seed fragments. Step S2 specifically includes the following sub-steps: S2-1: For each entity candidate Calculate cross-domain distribution degree Unauthorized Adjacency Ratio and query dependency ,in , , ,in Indicates entity candidate The actual number of permission domains distributed across the knowledge graph. This indicates the maximum number of preset permission domains in the system, used to normalize the cross-domain distribution range. Indicates entity candidate Adjacent to each other and not belonging to the user permission domain identifier set The number of adjacent nodes, Indicates entity candidate The total number of all adjacent nodes. Indicates entity candidate entity description vector, The cosine similarity function is represented by the three formulas above, which are used to quantify the cross-domain diffusion range of entity candidates, the tendency of unauthorized expansion, and the degree of dependence of the current query on the entity, respectively. S2-2: Based on cross-domain distribution degree Unauthorized Adjacency Ratio and query dependency Calculate bridging risk score ,in , , The formula represents the bridging risk weight, and the sum of the three factors is 1. The significance of this formula is to uniformly quantify three factors: whether the shared entity is easy to distribute across domains, whether it is easy to continue to connect to unauthorized neighboring domains, and whether it is strongly related to the current query, so as to identify the degree of bridging risk of the shared entity in the current query. S2-3: When At that time, entity candidates It is identified as a bridging entity and is prohibited from triggering external permission domain plaintext expansion during the current query process. Retain its normal related search capabilities within the current permission domain, among which The output of this sub-step is a set of bridging entities, which serves as the threshold for triggering bridging risks. In this implementation, by jointly analyzing the cross-domain distribution, unauthorized adjacency, and query dependencies of shared entities, bridging entities that are truly likely to trigger cross-domain expansion are first identified, and then a decision is made on whether to initiate bridge breaking. Compared to the approach of expanding first and then filtering, this approach moves boundary control forward to before expansion occurs, reducing the probability of original content from external permission domains entering the answer context. At the same time, for shared entities that do not meet the bridging risk conditions, they are still allowed to participate in ordinary association searches within the current permission domain, thus balancing security boundaries and search availability.

[0019] Step S3 specifically includes the following sub-steps: S3-1: For each bridging entity Constructing the projection node of the broken bridge Among them, the projection nodes of the broken bridge It only includes the evidence index of this permission domain, the cross-domain constraint request identifier, and the location where the constraint result is received, and does not include the original text fragments of external permission domains; and it is based on the seed fragment vector set corresponding to the bridging entity within this permission domain. Calculate the evidence center vector in this domain ,in This indicates the connection between the current user permission domain and the bridging entity. The number of associated seed fragments: This formula uses evidence fragments directly related to the bridging entity within the current authority domain to form a local evidence center, providing a stable semantic benchmark for subsequent constraint fusion. In this embodiment, the bridging entity no longer acts as a regular entity node to trigger the original text expansion of the external permission domain. Instead, it is replaced by a broken-bridge projection node. The broken-bridge projection node only retains the evidence index and cross-domain constraint interaction capabilities of its own permission domain and does not have the ability to directly read the original text content of the external permission domain. Thus, the shared entity changes from an expansion bridge point in the original hybrid retrieval chain to a restricted valve point, structurally blocking the direct expansion path from the local domain fragment to the shared entity to the external domain original text fragment. This is the key difference between this invention and conventional permission filtering and conventional graph expansion retrieval.

[0020] S3-2: Projection node of broken bridge Send a cross-domain constraint request to the local policy engine of the external permission domain and receive the constraint capsule. Constraint capsule At least include cross-domain dependency presence markers Risk level value Consistency with time ,in This indicates that the external permission domain returns the bridging entity. The effective time interval of the relevant constraints This represents the target time window corresponding to the current query. The significance of this formula lies in quantifying the consistency between the external permission domain's returned constraints and the current query's time range, thereby preventing the system from introducing relational conclusions that do not match the current query in the time dimension. Indicates the length of the corresponding time interval and the risk level value. The risk level of cross-domain dependencies is output by the local policy engine of the external permission domain based on preset risk classification rules. These rules map the risk level of cross-domain dependencies to ordered numerical values. When a cross-domain dependency exists... The value is 1 if it is not 0 otherwise; the external permission domain local policy engine is a software processing unit deployed inside the external permission domain, used to calculate the structured constraint results based on the local permission policy and risk rules.

[0021] S3-3: Based on the existence of cross-domain dependencies, a marker exists. Risk level value Consistency with time Calculate the constraint capsule score ,in , , The formula represents the weight of the constraint capsule, and the sum of the three is 1. The significance of the formula is to uniformly quantify three factors: whether there is a cross-domain dependency, what risk level the cross-domain dependency is at, and whether the dependency is valid within the current query time window, so as to form a structured constraint result that can be directly called in the subsequent restricted response stage. In this implementation, the external permission domain does not return the original text fragment, nor is it a summary information from which details of the external permission domain can be deduced. Instead, it returns structured constraint results such as the existence of cross-domain dependencies, risk levels, and time validity consistency. In this way, the system can determine whether a shared entity is associated with a cross-domain dependency, the risk level of that dependency, and whether the dependency is valid within the current query time window, without introducing the original text content of the external permission domain. This transforms cross-domain information from textual content into computable and determinate relational facts, improving the system's usability in boundary control scenarios.

[0022] Step S4 specifically includes the following sub-steps: S4-1: Merge the evidence fragments of this authority domain in the seed fragment set that meet the initial recall requirements with the constraint capsules corresponding to each bridging entity to form a restricted response context. The restricted response context only contains the evidence content of this authority domain and the structured constraint results, and does not contain the original text fragments of the external authority domain. S4-2: Calculate the resolvability score for the current query. ,in This indicates the initial recall score of evidence fragments within the current authority domain that are incorporated into the context of the restricted response. The average value is used to characterize the overall relevance of evidence within this authority domain. The constrained capsule score represents the score that incorporates the context of the limited response. The average value is used to characterize the overall availability of cross-domain relationship constraints. This represents the bridging risk score for all bridging entities in the current query. The maximum value is used to characterize the highest cross-domain expansion risk faced by the current query. , , The formula indicates that the system should only proceed to the response generation stage when the evidence in the current authority domain is sufficiently sufficient, the constraint capsule is sufficiently consistent, and the highest bridging risk is effectively suppressed. S4-3: When When the current query meets the conditions for being respondable, it is determined that the query is ready to be answered. If the current query does not meet the criteria for being answered, it will be transferred to the approval process. This indicates the respondability threshold, and the output of this sub-step is the respondability determination result. In this implementation, the evidence fragments and constraint capsules of this permission domain are not directly used as the final answer input. Instead, they first form a restricted answer context, and then the answerability score is used to determine whether the current query meets the answering conditions. Through this processing method, the system can distinguish between two scenarios: those with sufficient evidence in this domain and clear constraint results, and those with only partial relational constraints that are insufficient to support a clear answer. This avoids hasty output due to insufficient information and also avoids rejecting answers indiscriminately when there is a basis for an answer, thereby improving the reliability and executability of enterprise knowledge base query results.

[0023] Step S5 specifically includes the following sub-steps: S5-1: Count the number of bridging entities corresponding to the current query. There is a labeled mean for cross-domain dependencies. ,in Indicates the first Each bridging entity corresponds to a cross-domain dependency in the constraint capsule. The significance of the formula is to quantify the overall occurrence of cross-domain dependencies in the current query, thereby providing a basis for the final output type selection. S5-2: Select the output type according to the following rules: When and When, output a direct response result; the direct response result is generated only based on evidence fragments within this permission domain; when and When this happens, a restricted response result is output. This restricted response result includes verifiable conclusions within this permission domain, an explanation of the existence of cross-domain dependencies, and a risk level description. When the approval transfer result is output, the approval transfer result shall at least include the reason for the transfer, the object to be approved, and the suggested scope of verification. Without introducing the original text fragments of external permission domains, the respondability determination and the generation of restricted responses shall be jointly supported by the evidence of this permission domain and the constraint capsule. S5-3: Query text User permission domain identifier set Seed fragment set, bridging entity set, bridging risk score Constraint capsules Responsibility score The output type and final response result are written into the audit log to form a complete and traceable processing chain. The purpose of this sub-step is to ensure that the input source, boundary control process and output basis can be traced back for each question and answer. In this embodiment, the system outputs a direct response, a restricted response, or an approval transfer result based on the respondability determination result and cross-domain dependency situation, and simultaneously records information such as query input, bridging entity, constraint capsule, and output type. Therefore, this invention not only controls the risk of boundary penetration in the intermediate retrieval chain, but also implements this boundary control result at the final output level and in the post-event traceability chain, enabling the system to possess verifiable and auditable enterprise-level application capabilities.

[0024] The system used in this method includes: The query boundary construction module is used to receive user query content, combine the current session context and user permission domain information, perform normalization processing on the query, and form an initial set of candidate knowledge fragments within the user permission domain; The broken bridge projection retrieval module is used to identify shared entities from the initial candidate knowledge fragment set, establish broken bridge projection nodes for shared entities that meet the bridging risk conditions, so that the shared entities no longer trigger the original text expansion of the external permission domain, and obtain the structured constraint results without the original text content from the external permission domain based on the broken bridge projection nodes. The restricted output module is used to merge evidence fragments within the user's permission domain with the structured constraint results, determine whether the current query meets the conditions for being respondable, and output the direct response result, restricted response result, or approval transfer result, while recording the complete processing chain. The query boundary construction module includes: a query normalization unit, used to uniformly encode the user's current input, current session history, and user permission domain identifier to form a standardized query representation; a permission consistency recall unit, used to sort candidate knowledge fragments and output an initial seed fragment set within the range where the knowledge fragment permission tags are consistent with the user permission domain; and a seed entity extraction unit, used to extract entity names, entity types, and the associated positions of entities in seed fragments from the initial seed fragment set to form a shared entity candidate set. The broken bridge projection retrieval module includes: a bridging risk determination unit, used to identify whether to determine the corresponding shared entity as a bridging entity based on the cross-domain distribution of shared entities, unauthorized adjacency, and query-related information; a projection node construction unit, used to generate broken bridge projection nodes for bridging entities, wherein the broken bridge projection nodes contain the evidence index of the current permission domain, the cross-domain constraint request identifier, and the constraint result receiving location, and do not contain the original text fragment of the external permission domain; and a constraint capsule acquisition unit, used to send cross-domain constraint requests to the local policy engine of the external permission domain and receive structured constraint results, wherein the structured constraint results include at least a cross-domain dependency existence marker, a risk level value, and a time validity consistency, and do not contain the original text of the external permission domain, the original entity list, and the full-text summary; The restricted output module includes: an evidence fusion unit, used to uniformly organize evidence fragments and constraint capsules within the current permission domain to form a restricted response context; a respondability determination unit, used to determine the output type corresponding to the current query based on the sufficiency of evidence within the current permission domain, the consistency of constraint capsules, and the intensity of bridging risk; a result generation unit, used to generate direct response text, restricted response text, or approval transfer text according to the output type; and a link recording unit, used to record query input, permission domain identifier, seed fragment, shared entity, bridging risk determination result, constraint capsule, output type, and final result.

[0025] The unique feature of this invention is not that it proposes separate access control, separate entity recognition, or separate restricted output, but that it forms a closed-loop link around shared entities, which is the most likely bridge point to trigger cross-domain expansion. This link includes bridging risk assessment, broken bridge projection node construction, constraint capsule acquisition, respondability assessment, and hierarchical output, so that the system retains the ability to judge cross-domain relationships without introducing the original content of external access domains.

[0026] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0027] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A conversational AI query method based on an enterprise knowledge base, characterized by: include: S1: Receive the query text input by the user, read the current session history and user permission domain information, and perform normalization processing on the query text to form a standardized query representation; Within the user permission domain, an initial recall of knowledge fragments is performed to obtain a set of seed fragments that are relevant to the current query and have the same permissions. S2: Perform sharing analysis and bridging risk analysis on the entity candidates in the seed fragment set to identify whether there are shared entities that are simultaneously distributed in multiple permission domains and may trigger the expansion of external permission domains; when a shared entity meets the bridging risk condition, the shared entity is identified as a bridging entity; S3: Establish a broken bridge projection node for the bridging entity, so that the bridging entity only participates in subsequent processing as a restricted intermediate object in the current query process; the broken bridge projection node initiates a constraint request to the external permission domain and receives a constraint capsule that does not contain the original text content. The broken bridge projection node is used to replace the shared entity to continue the cross-domain original text expansion of the restricted intermediate data object. The constraint capsule is a data object returned by the external permission domain that only contains structured constraint information and does not contain the original text content. S4: Integrate the evidence fragments of this permission domain with the constraint capsule to form a restricted response context, and determine whether the current query meets the respondable conditions based on the sufficiency of evidence, the consistency of constraints, and the degree of bridging risk; if the respondable conditions are met, proceed to the response generation stage; if the respondable conditions are not met, proceed to the approval and transfer stage. S5: Output the corresponding text based on the answerability determination result. In the case of direct reply, output the complete answer based on the evidence of this permission domain. In the case of restricted reply, output the restricted answer containing relational conclusions and binding explanations. In the case of approval transfer, output the approval prompt information. At the same time, record the complete processing link.

2. The conversational AI query method based on enterprise knowledge base according to claim 1, characterized in that: Step S1 specifically includes the following sub-steps: S1-1: User query text Current session history summary and user permission domain identifier set Perform unified encoding to obtain a standardized query vector. ,in This is a query normalization encoding function used to encode textual semantic information, session continuity information, and permission boundary information into the same query representation; S1-2: For each candidate knowledge fragment Calculate the initial recall score ,in Represents the standardized query vector With candidate knowledge fragment vectors The semantic similarity is used to characterize the degree of semantic fit between candidate knowledge fragments and the current query; This indicates the degree of explicit matching between query keywords and candidate knowledge fragment text, used to characterize the consistency between technical terms and fixed appellations. Represents the set of user permission domain identifiers The degree of consistency between the candidate knowledge fragment and its permission tag is used to characterize whether the candidate knowledge fragment is within the current accessible boundary. , , This represents the corresponding weight, and the sum of the three is 1; S1-3: Select the option that satisfies The candidate knowledge fragments are used as seed fragments, among which This represents the initial recall threshold, used to control the minimum relevance required for seed fragments to enter subsequent bridging analysis; and entity names, entity types, and entity locations are extracted from the seed fragments to form an entity candidate set.

3. The conversational AI query method based on an enterprise knowledge base according to claim 2, characterized in that: Step S2 specifically includes the following sub-steps: S2-1: For each entity candidate Calculate cross-domain distribution degree Unauthorized Adjacency Ratio and query dependency ,in , , ,in Indicates entity candidate The actual number of permission domains distributed across the knowledge graph. This indicates the maximum number of preset permission domains in the system, used to normalize the cross-domain distribution range. Indicates entity candidate Adjacent to each other and not belonging to the user permission domain identifier set The number of adjacent nodes, Indicates entity candidate The total number of all adjacent nodes. Indicates entity candidate entity description vector, Represents the cosine similarity function; S2-2: Based on the cross-domain distribution degree Unauthorized Adjacency Ratio and query dependency Calculate bridging risk score ,in , , This indicates the bridging risk weight, and the sum of the three values ​​is 1; S2-3: When At that time, entity candidates It is identified as a bridging entity and is prohibited from triggering external permission domain plaintext expansion during the current query process. Retain its normal related search capabilities within the current permission domain, among which This is the threshold for triggering bridging risks.

4. The conversational AI query method based on an enterprise knowledge base according to claim 3, characterized in that: Step S3 specifically includes the following sub-steps: S3-1: For each bridging entity Constructing the projection node of the broken bridge The bridge break projection node It only includes the evidence index of this permission domain, the cross-domain constraint request identifier, and the location where the constraint result is received, and does not include the original text fragments of external permission domains; and it is based on the seed fragment vector set corresponding to the bridging entity within this permission domain. Calculate the evidence center vector in this domain ,in This indicates the connection between the current user permission domain and the bridging entity. The number of associated seed fragments; S3-2: Projection node of broken bridge Send a cross-domain constraint request to the local policy engine of the external permission domain and receive the constraint capsule. The constraint capsule At least include cross-domain dependency presence markers Risk level value Consistency with time ,in This indicates that the external permission domain returns the bridging entity. The effective time interval of the relevant constraints This indicates the target time window corresponding to the current query. The risk level value represents the length of the corresponding time interval. The external permission domain's local policy engine outputs a risk classification rule based on preset risk classification rules. This preset risk classification rule maps the risk level of cross-domain dependencies to ordered numerical values. When a cross-domain dependency exists... The value is 1 if it is set to 1, and 0 otherwise. S3-3: A marker exists based on the cross-domain dependency. Risk level value Consistency with time Calculate the constraint capsule score ,in , , This represents the weights of the constraint capsule, and the sum of the three is 1.

5. The conversational AI query method based on an enterprise knowledge base according to claim 4, characterized in that: Step S4 specifically includes the following sub-steps: S4-1: The evidence fragments of this authority domain in the seed fragment set that meet the initial recall requirements are fused with the constraint capsules corresponding to each bridging entity to form a restricted response context, wherein the restricted response context only contains the evidence content of this authority domain and the structured constraint results, and does not contain the original text fragments of the external authority domain. S4-2: Calculate the resolvability score for the current query. ,in This indicates the initial recall score of evidence fragments within the current authority domain that are incorporated into the context of the restricted response. The average value is used to characterize the overall relevance of evidence within this authority domain. The constrained capsule score represents the score that incorporates the context of the limited response. The average value is used to characterize the overall availability of cross-domain relationship constraints. This represents the bridging risk score for all bridging entities in the current query. The maximum value is used to characterize the highest cross-domain expansion risk faced by the current query. , , Indicates the reproducibility weight; S4-3: When When the current query meets the conditions for being respondable, it is determined that the query is ready to be answered. If the current query does not meet the criteria for being answered, it will be transferred to the approval and transfer process. This indicates the threshold at which a response is possible.

6. The conversational AI query method based on an enterprise knowledge base according to claim 5, characterized in that: Step S5 specifically includes the following sub-steps: S5-1: Count the number of bridging entities corresponding to the current query. There is a labeled mean for cross-domain dependencies. ,in Indicates the first Each bridging entity corresponds to a cross-domain dependency in the constraint capsule that is marked; S5-2: Select the output type according to the following rules: When and When, a direct response result is output, which is generated solely based on evidence fragments from this permission domain; when and When the time comes, a restricted response result is output. This restricted response result includes verifiable conclusions within this permission domain, an explanation of the existence of cross-domain dependencies, and a risk level description. When the approval transfer result is output, the approval transfer result shall at least include the transfer reason, the object to be approved and the suggested verification scope. Without introducing the original text fragments of external permission domains, the respondability determination and the generation of restricted responses shall be jointly supported by the evidence of this permission domain and the constraint capsule. S5-3: Query text User permission domain identifier set Seed fragment set, bridging entity set, bridging risk score Constraint capsules Responsibility score The output type and final response result are written into the audit log to form a complete and traceable processing chain.

7. The conversational AI query method based on an enterprise knowledge base according to claim 6, characterized in that: The system used in this method includes: The query boundary construction module is used to receive user query content, combine the current session context and user permission domain information, perform normalization processing on the query, and form an initial set of candidate knowledge fragments within the user permission domain; The broken bridge projection retrieval module is used to identify shared entities from the initial candidate knowledge fragment set, establish broken bridge projection nodes for shared entities that meet the bridging risk conditions, so that the shared entities no longer trigger the original text expansion of the external permission domain, and obtain the structured constraint results without original text content from the external permission domain according to the broken bridge projection nodes. The restricted output module is used to fuse evidence fragments within the user's permission domain with the structured constraint results, determine whether the current query meets the conditions for being respondable, and output a direct response result, a restricted response result, or an approval transfer result, while recording the complete processing chain.

8. The conversational AI query method based on an enterprise knowledge base according to claim 7, characterized in that: The query boundary construction module includes: a query normalization unit, used to uniformly encode the user's current input, current session history, and user permission domain identifier to form a standardized query representation; a permission consistency recall unit, used to sort candidate knowledge fragments and output an initial seed fragment set within the range where the knowledge fragment permission tags are consistent with the user permission domain; and a seed entity extraction unit, used to extract entity names, entity types, and the associated positions of entities in seed fragments from the initial seed fragment set to form a shared entity candidate set. The broken bridge projection retrieval module includes: a bridging risk determination unit, used to identify whether to determine the corresponding shared entity as a bridging entity based on the cross-domain distribution of the shared entity, the unauthorized adjacency situation, and the query-related situation; a projection node construction unit, used to generate broken bridge projection nodes for bridging entities, wherein the broken bridge projection nodes include the evidence index of the current permission domain, the cross-domain constraint request identifier, and the constraint result receiving location, and do not include the original text fragment of the external permission domain; and a constraint capsule acquisition unit, used to send a cross-domain constraint request to the local policy engine of the external permission domain and receive the structured constraint result, wherein the structured constraint result includes at least a cross-domain dependency existence marker, a risk level value, and a time validity consistency, and does not include the original text of the external permission domain, the original entity list, and the full-text summary; The restricted output module includes: an evidence fusion unit, used to uniformly organize evidence fragments and constraint capsules of the current permission domain to form a restricted response context; a respondability determination unit, used to determine the output type corresponding to the current query based on the sufficiency of evidence in the current permission domain, the consistency of constraint capsules, and the intensity of bridging risk; a result generation unit, used to generate direct response text, restricted response text, or approval transfer text according to the output type; and a link recording unit, used to record query input, permission domain identifier, seed fragment, shared entity, bridging risk determination result, constraint capsule, output type, and final result.

Images