A multi-model unified interface and dynamic configuration hot updating method and system

Abstract

The application discloses a multi-model unified interface and dynamic configuration hot updating method and system, relates to the technical field of large language model application, and comprises the following steps: collecting the capability parameters of candidate reasoning services, performing semantic normalization and field rewriting, and constructing a capability semantic graph; according to the capability semantic graph, the calling requirements submitted by the business side are constrained and disassembled and semantic reorganization is performed to generate a unified request semantic carrier; the unified request semantic carrier is mapped to the capability semantic graph to generate a unified interface mapping scheme; the current effective active intergenerational configuration is read from the unified interface mapping scheme, and when the active intergenerational configuration is detected to be changed, a target intergenerational configuration snapshot and an intergenerational difference relationship are generated. The application realizes accurate transmission of configuration changes to running object changes and smooth handover of new and old generations.

Description

Technical Field

[0001] This invention relates to the field of large language model application technology, and in particular to a unified interface for multiple models and a dynamic configuration hot update method and system. Background Technology

[0002] With the development of large language models, enhanced retrieval generation, intelligent agent collaborative invocation, and multi-cloud inference services, unified access technology for heterogeneous inference services has evolved from static interface encapsulation to a software architecture integrating semantic orchestration, protocol adaptation, field rewriting, configuration management, and online updates. Especially in microservice architecture environments, different inference capabilities are typically deployed as independent service units and provided externally through service orchestration links. Existing technologies generally use adaptation layers, gateway layers, or orchestration layers to convert the request fields, authentication methods, return structures, streaming protocols, and exception code systems of different inference services, and introduce mechanisms such as configuration centers, hot reloading, canary releases, and shadow verification into the inference service access link.

[0003] Existing technologies still have two shortcomings: First, existing multi-model access solutions mostly remain at the field-level adaptation or rule-level routing level, lacking a unified semantic expression for the boundaries of different inference service capabilities. This makes it difficult to establish a stable mapping relationship between call requirements and underlying service capabilities. In microservice architectures, this is especially prone to semantic mismatches and expansion difficulties in scenarios where structured output, streaming interaction, tool calls, and exception normalization coexist. Second, existing configuration hot update technologies mostly adopt full reload or instance replacement methods. In microservice architectures, it is difficult to identify the scope of impact of configuration changes around the actual call path, and it is also difficult to achieve differential rewriting, shadow carrying, smooth request switching, and exception rollback. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a unified interface for multiple models and a dynamic configuration hot update method to solve the problems of inconsistent semantic mapping and coarse configuration hot updates between heterogeneous inference services.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for hot updating of a unified interface and dynamic configuration for multiple models, comprising: collecting capability parameters of candidate inference services and performing semantic normalization and field rewriting to construct a capability semantic graph; based on the capability semantic graph, performing constraint decomposition and semantic reorganization on the call requests submitted by the business side to generate a unified request semantic carrier; mapping the unified request semantic carrier to the capability semantic graph to generate a unified interface mapping scheme; reading the currently effective active generation configuration from the unified interface mapping scheme, and generating a target generation configuration snapshot and a generation difference relationship table when a change in the active generation configuration is detected; using the target generation configuration snapshot, the generation difference relationship table, and the unified interface mapping scheme as a basis, performing differential rewriting to construct a set of shadow call objects bound to the target generation, and allocating new requests to the set of shadow call objects; based on the set of shadow call objects, initiating a call according to the unified interface mapping scheme, and performing unified response rewriting on the returned content to generate generation confirmation information and generation rollback information.

[0008] As a preferred embodiment of the multi-model unified interface and dynamic configuration hot update method described in this invention, the specific steps for constructing the capability semantic graph are as follows:

[0009] Collect the raw capability parameters of candidate inference services, and aggregate and organize them according to service identifiers to form a raw capability record set;

[0010] The original capability record set is used to organize different records of the same capability item, and semantic normalization and field rewriting are performed according to the capability semantic anchor point to form a standard capability item set;

[0011] Establish the correspondence between candidate reasoning services and capability semantic anchors using a standard capability entry set, and generate a capability semantic graph.

[0012] As a preferred embodiment of the multi-model unified interface and dynamic configuration hot update method of the present invention, the specific steps for generating the unified request semantic carrier are as follows:

[0013] Based on the capability semantic graph, the call request messages submitted by the business side are read, and the task information, request content and constraint information in the call requirements are aggregated and organized to form the original call requirement table;

[0014] Based on the original call request table, each request item is decomposed and semantically reorganized according to the capability semantic anchor points in the capability semantic graph to form a reorganized constraint chain;

[0015] The reorganized constraint chain is encapsulated according to a unified carrier structure to generate a unified request semantic carrier.

[0016] As a preferred embodiment of the multi-model unified interface and dynamic configuration hot update method described in this invention, the specific steps of mapping the unified request semantic carrier to the capability semantic graph are as follows:

[0017] The task information, request content and constraint information in the unified request semantic carrier are matched with the corresponding capability semantic anchors in the capability semantic graph, and the mapping association value between each requirement item and each capability semantic anchor is calculated to form a request anchor correspondence table.

[0018] Using the request anchor point mapping table as a clue, candidate associations connected to each capability semantic anchor point are extracted from the capability semantic graph to obtain a candidate mapping relationship set.

[0019] As a preferred embodiment of the multi-model unified interface and dynamic configuration hot update method of the present invention, the specific steps for generating the unified interface mapping scheme are as follows:

[0020] The candidate mapping relationship set is subjected to conflict resolution, position correction and order reorganization to form a mapping relationship chain;

[0021] The mapping relationship chain is combined and arranged with unified request fields, service fields, interaction rules, and exception normalization rules to generate a unified interface mapping scheme.

[0022] As a preferred embodiment of the multi-model unified interface and dynamic configuration hot update method of the present invention, the specific steps of reading the currently effective active generation configuration from the unified interface mapping scheme and detecting a change in the active generation configuration are as follows:

[0023] Extract the configuration scope path from the unified interface mapping scheme, and read the currently effective active generation configuration based on the configuration scope path to form an active generation configuration projection set;

[0024] The active generation configuration projection set is associated and solidified to generate an active generation configuration fingerprint, and configuration changes are continuously captured to form a generation change candidate set.

[0025] As a preferred embodiment of the multi-model unified interface and dynamic configuration hot update method described in this invention, the specific steps for generating the target generation configuration snapshot and generational difference relationship table are as follows:

[0026] The generational change candidate set and the active generational configuration projection set are merged in a targeted manner according to the configuration action path, and generational reorganization is performed on the changed configuration items to generate a target generational configuration snapshot.

[0027] Based on the target generation configuration snapshot, backtrack along the configuration action path to the corresponding entry in the unified interface mapping scheme, and compare it item by item with the active generation configuration projection set to generate a generation difference relationship table.

[0028] As a preferred embodiment of the multi-model unified interface and dynamic configuration hot update method described in this invention, the specific steps of constructing a set of shadow invocation objects bound to the target generation and allocating new requests to the set of shadow invocation objects are as follows:

[0029] Based on the target generation configuration snapshot, generation difference relationship table and unified interface mapping scheme, the changed configuration items, affected mapping entries and call paths are associated and expanded to form a differential scope chain;

[0030] The differential scope chain is used to perform targeted reorganization of field rewriting relationships, interaction connection rules and exception normalization rules in the unified interface mapping scheme to generate shadow carrying slices.

[0031] Using shadow carrier slices as assembly units, generational binding is performed on connection context, authentication context, field rewriting relationship, interaction connection rules and exception normalization path to construct a set of shadow call objects;

[0032] Based on the generational and rewrite states of each call path in the shadow call object set, a path-level admission identifier is generated, and new requests are assigned to the shadow call object set according to the path-level admission identifier.

[0033] As a preferred embodiment of the multi-model unified interface and dynamic configuration hot update method described in this invention, the specific steps for generating generation confirmation information and generation rollback information are as follows:

[0034] Based on the shadow call object set, the requests allocated to the shadow call object set are rewritten and interactively orchestrated according to the unified interface mapping scheme, and the call is initiated through the shadow call object set to form the original response content;

[0035] The original response content is rewritten and reordered to form a unified response content. The operational status of the target generation is determined, and generational confirmation information and generational rollback information are generated.

[0036] Secondly, this invention provides a multi-model unified interface and dynamic configuration hot update system, comprising: a graph construction module for collecting capability parameters of candidate inference services and performing semantic normalization and field rewriting to construct a capability semantic graph; a semantic reorganization module for constraining and semantically reorganizing the call requests submitted by the business side based on the capability semantic graph to generate a unified request semantic carrier; a mapping generation module for mapping the unified request semantic carrier to the capability semantic graph to generate a unified interface mapping scheme; a generational snapshot module for reading the currently effective active generational configuration from the unified interface mapping scheme and generating a target generational configuration snapshot and a generational difference relationship table when a change in the active generational configuration is detected; a shadow assembly module for using the target generational configuration snapshot, the generational difference relationship table, and the unified interface mapping scheme as a basis to perform differential rewriting, construct a set of shadow call objects bound to the target generation, and allocate new requests to the set of shadow call objects; and a response determination module for initiating calls based on the set of shadow call objects according to the unified interface mapping scheme, performing unified response rewriting on the returned content, and generating generational confirmation information and generational rollback information.

[0037] The beneficial effects of this invention are as follows: By mapping the unified request semantic carrier to the capability semantic graph and generating a unified interface mapping scheme, a stable semantic mapping between business-side call requirements and heterogeneous inference service capability boundaries is achieved, resulting in unified field rewriting, unified interaction orchestration, and unified exception normalization. By combining the target generation configuration snapshot, generation difference relationship table, and unified interface mapping scheme to perform differential rewriting and construct a shadow call object set, the accurate transmission of configuration changes to changes in running objects and the smooth transition between old and new generations are achieved, resulting in low-disturbance switching, shadow verification carrying, and exception rollback control during configuration hot update. Attached Figure Description

[0038] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 A flowchart for a unified interface for multiple models and a dynamic configuration hot update method.

[0040] Figure 2 A flowchart for constructing a capability semantic graph.

[0041] Figure 3 A flowchart for generating a unified interface mapping scheme.

[0042] Figure 4A flowchart for constructing and requesting the shadow call object set. Detailed Implementation

[0043] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0044] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0045] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0046] Reference Figures 1-4 This is one embodiment of the present invention, which provides a multi-model unified interface and a dynamic configuration hot update method, including the following steps:

[0047] S1. Collect the capability parameters of candidate inference services, and perform semantic normalization and field rewriting to construct a capability semantic graph.

[0048] Collect the raw capability parameters of candidate inference services, and aggregate and organize them according to service identifiers to form a raw capability record set.

[0049] The specific process includes reading various capabilities from the candidate reasoning services and registering them under the corresponding service identifiers. Then, the capabilities that correspond to each other under the same service identifier are compared, organized, and merged so that the capabilities are centrally collected under the corresponding service identifiers.

[0050] Furthermore, a list of candidate inference services is determined, and each candidate inference service is assigned a unique service identifier. Based on the candidate inference service's context reading capacity, streaming output capability, structured output capability, tool invocation capability, input / output modality type, invocation restriction parameters, and abnormal feature parameters, the read capability content is registered under the corresponding service identifier name. All capability content under the same service identifier name is then arranged in a centralized manner, and records that can correspond to the same capability item range are collected and organized, and summarized and arranged according to the service identifier to form the original capability record set.

[0051] The original capability record set is used to organize different records of the same capability item, and semantic normalization and field rewriting are performed according to the capability semantic anchor point to form a standard capability item set.

[0052] The specific process includes extracting records with the same name, similar content, and corresponding to the same capability item from the original capability record set; arranging different records of the same capability item in a group; comparing and organizing the capability content, field names, and expression methods; grouping records with the same meaning but different writing styles into the same correspondence; using the capability semantic anchor as a unified reference for semantic unification and field rewriting; and unifying the expression of the same capability item in different candidate reasoning services to the same standard.

[0053] Furthermore, records that have undergone corresponding organization, semantic unification, and field rewriting are transcribed one by one into a unified entry format. The service identifier, the unified capability item name, the unified field name, and the corresponding capability content are fixed in the same entry structure, so that one record corresponds to one standard capability entry. Standard capability entries with the same service identifier and the same unified capability item name, field name, and capability content are repeatedly merged and then centrally collected according to the service identifier and the unified capability item name to form a set of standard capability entries.

[0054] It should be noted that the capability semantic anchor is a standardized semantic reference benchmark generated by extracting four types of data: the preset capability dictionary, the candidate inference service interface document, the historical call record, and the manually configured capability standard items.

[0055] Establish the association between candidate reasoning services and capability semantic anchors using a standard capability entry set, and generate a capability semantic graph.

[0056] The specific process includes extracting the service identifier, unified capability item name, unified field name, and corresponding capability content from the standard capability item set one by one; verifying the unified capability item name and unified field name against the capability semantic anchor; connecting the standard capability item with the corresponding capability semantic anchor one by one to form the association relationship between the candidate reasoning service and the capability semantic anchor; merging and organizing all the association relationships; and unifying the candidate reasoning service, capability semantic anchor, and standard capability item set content into the same relationship set to generate a capability semantic graph.

[0057] S2. Based on the capability semantic graph, the call requests submitted by the business side are constrained, decomposed, and semantically reorganized to generate a unified request semantic carrier.

[0058] Based on the capability semantic graph, the call request messages submitted by the business side are read, and the task information, request content and constraint information in the call requirements are aggregated and organized to form the original call requirement table.

[0059] The specific process includes extracting the fields related to the call requirements from the call request message submitted by the business side, verifying their correspondence with the capability semantic anchors in the capability semantic graph, and organizing the field content into task information, request content, and constraint information.

[0060] Furthermore, the aggregated task information, request content, and constraint information are registered one by one in the same organized table according to the corresponding source in the call request message, and the field name, field content, and category are written to keep the task information, request content, and constraint information clearly partitioned in the same organized table. After registration, content that can point to the same meaning is merged, and the organized task information, request content, and constraint information are arranged in the order of expression of call requirements to form the original call requirement table.

[0061] Based on the original call request table, each request item is constrained and semantically reorganized according to the capability semantic anchor points in the capability semantic graph, forming a reorganized constraint chain.

[0062] The specific process includes extracting the task information, request content, and constraint information from the original call request table one by one, verifying the correspondence between the request items and the capability semantic anchors in the capability semantic graph, merging the request items that correspond to the same capability semantic anchor into the same scope, organizing the request content into constraint items that can directly correspond to the capability semantic anchor, rearranging the already organized constraint items according to the correspondence between capability semantic anchors in the capability semantic graph, classifying and classifying the constraint items and verifying their correspondence, so that the constraint items formed by the task information, the constraint items formed by the request content, and the constraint items formed by the constraint information are sequentially connected, and the constraint items that are related, mutually supportive, and jointly act on the same call request are continuously organized into the same associated sequence to form a reorganized constraint chain.

[0063] It should be noted that the reorganized constraint chain refers to a continuous and orderly sequence of constraint items formed by breaking down the original business-side call requirements during the requirement analysis phase and rearranging and connecting them according to the underlying capability standards.

[0064] The reorganized constraint chain is encapsulated according to a unified carrier structure to generate a unified request semantic carrier.

[0065] The specific process includes classifying and sequentially writing each constraint item in the reorganized constraint chain according to its corresponding position in the unified carrier structure, so that each constraint item maintains its original order, correspondence, and attribution relationship in the unified carrier structure. The task-related constraint items, request content-related constraint items, and restriction-related constraint items written into the unified carrier structure are summarized and the entire constraint content corresponding to the same call request is checked to form a continuous and consistent semantic expression content among the constraint items, thereby generating a unified request semantic carrier.

[0066] It should be noted that the unified request semantic carrier refers to the standardized semantic data object that carries the complete call requirements after the constraints in the reorganized constraint chain are structurally encapsulated according to a preset standard template.

[0067] S3. Map the unified request semantic carrier to the capability semantic graph to generate a unified interface mapping scheme.

[0068] The task information, request content, and constraint information in the unified request semantic carrier are matched with the corresponding capability semantic anchors in the capability semantic graph, and the mapping association value between each requirement item and each capability semantic anchor is calculated to form a request anchor correspondence table.

[0069] The specific process includes matching the task information, request content, and constraint information in the unified request semantic carrier with the corresponding capability semantic anchors in the capability semantic graph. This involves breaking down the task information, request content, and constraint information in the unified request semantic carrier into independent requirement items, and then verifying the correspondence between the independent requirement items and the capability semantic anchors in the capability semantic graph, so that the requirement content in the unified request semantic carrier corresponds to a specific capability semantic anchor.

[0070] Furthermore, by calculating the mapping association values ​​between each requirement item and each capability semantic anchor point, a request anchor point correspondence table is formed. Based on the correspondence between independent requirement items and capability semantic anchor points, the mapping association values ​​are calculated. Independent requirement items that have been matched and have obtained mapping association values ​​are centrally arranged. The capability semantic anchor points and mapping association values ​​corresponding to the independent requirement items are written into the same sorting position according to the correspondence. Capability semantic anchor points that can fully express the meaning of the requirement are retained, and capability semantic anchor points that do not match are removed. The retained independent requirement items, corresponding capability semantic anchor points, and mapping association values ​​are registered in the same sorting table in a unified order to form a request anchor point correspondence table.

[0071] The expression for calculating the mapping association value between each requirement item and each capability semantic anchor is as follows:

[0072] ;

[0073] in, Indicates the first The first requirement item and the first The mapping association values ​​between each capability semantic anchor point Indicates the sequence number of the requirement item. The index of the semantic anchor point representing the capability. Indicates the first The first requirement item and the first Semantic distance between capability semantic anchors Indicates the first The first requirement item and the first Structural consistency among capability semantic anchors Indicates the first The first requirement item and the first Constraint deviation between semantic anchor points of each capability Indicates the first The first requirement item and the first The degree of interaction conflict between semantic anchors of capabilities.

[0074] It should be noted that, It is obtained by extracting common semantic features, constructing semantic feature vectors, and calculating cosine distance; It is obtained by extracting field hierarchy, field order, field quantity, distribution of required fields, distribution of optional fields, and nesting relationships, constructing a structural description vector, and calculating the similarity. It is obtained by extracting context range constraints, output format constraints, streaming return constraints, tool call constraints, latency constraints, and resource limit constraints, constructing a constraint parameter vector, and calculating the degree of deviation. It is obtained by extracting the request initiation method, return organization method, streaming segmentation method, tool triggering method, context continuation method, and exception feedback method, constructing an interaction feature vector, and calculating the degree of inconsistency.

[0075] Using the request anchor point mapping table as a clue, candidate associations connected to each capability semantic anchor point are extracted from the capability semantic graph to obtain a candidate mapping relationship set.

[0076] The specific process includes expanding each independent requirement item, capability semantic anchor, and mapping association value in the request anchor correspondence table. Starting from the capability semantic anchor, the process searches the capability semantic graph for candidate inference services, unified capability item names, unified field names, and corresponding capability content that have already established a correspondence with the capability semantic anchor. The found association content is then matched against the independent requirement items. The core intent of the task must be highly matched with the capability semantic anchor, and hard constraints such as context capacity, output format, tool calls, and latency must be fully implemented in the candidate service. Within the parameters and without exceeding limits, the call sequence and protocol stack of streaming and non-streaming control, fragmented return logic, status continuation and exception feedback must be seamlessly compatible; the system will quantify and evaluate these three items based on the mapping association value formula. Only when all three pass the preset threshold verification and there is no conflict will it be determined that they can be maintained and retained. Mismatched association content will be removed, and the retained association content will be merged and organized according to the order of the request anchor point correspondence table, so that independent requirement items can continue to be mapped to the association content in the capability semantic graph, and a candidate mapping relationship set will be obtained.

[0077] The candidate mapping relationship set is subjected to conflict resolution, position correction and order arrangement to form a mapping relationship chain.

[0078] The specific process includes: organizing the associated content in the candidate mapping relationship set according to independent requirement items; filtering out conflicting associated content; retaining associated content that can jointly support the expression of independent requirement items; checking the position and arranging the order of the retained associated content; resolving conflicts, correcting positions and arranging the order; merging the organized associated content into mapping units; connecting adjacent mapping units sequentially according to the sequential relationship between independent requirement items; eliminating mapping units that cannot form a continuous mapping relationship; and retaining mapping units that can maintain the correspondence between the preceding and following parts and the mapping connection, thereby obtaining the mapping relationship chain.

[0079] It should be noted that the mapping relationship chain refers to a continuous, ordered, and executable mapping relationship sequence formed by connecting the associated content in the candidate mapping relationship set in sequence according to the successive relationship of independent requirement items after conflict screening, position verification and order rearrangement.

[0080] The mapping relationship chain is combined and arranged with unified request fields, service fields, interaction rules, and exception normalization rules to generate a unified interface mapping scheme.

[0081] The specific process includes: expanding each mapping unit in the mapping relationship chain in sequence; matching the unified request fields already determined in the mapping units with the service fields in the candidate inference services; incorporating the associated content representing the relationship between request transmission, content acceptance, and capability invocation into the scope of interaction rules; incorporating the associated content representing the unified correspondence of different abnormal content into the scope of abnormal normalization rules; and arranging and centrally merging the unified request fields, service fields, interaction rules, and abnormal normalization rules continuously according to the sequential relationship in the mapping relationship chain, so that the same invocation requirement maintains a complete and consistent correspondence from the unified request fields to the service fields, from the service fields to the interaction rules, and from the interaction rules to the abnormal normalization rules, thereby generating a unified interface mapping scheme.

[0082] S4. Read the currently effective active generation configuration from the unified interface mapping scheme. When a change in the active generation configuration is detected, generate a target generation configuration snapshot and generation difference relationship table.

[0083] The configuration scope path is extracted from the unified interface mapping scheme, and the currently effective active generation configuration is read based on the configuration scope path to form an active generation configuration projection set.

[0084] The specific process includes expanding the unified request fields, service fields, interaction rules, and exception normalization rules in the unified interface mapping scheme one by one, connecting the relevant correspondences to form a configuration action path, and then searching, reading, and merging the currently effective generational configurations along the configuration action path.

[0085] Furthermore, the activity generation configurations read along the configuration action path are expanded one by one, and the corresponding positions of the activity generation configurations are checked against the corresponding positions in the configuration action path. The activity generation configurations that can support the establishment of the same configuration action path are grouped into the same aggregation range. After the check, the activity generation configurations with the same position and the same function under the same configuration action path are merged. The activity generation configurations that cannot support the establishment of the configuration action path or are inconsistent with the corresponding positions in the configuration action path are removed. Once the valid correspondence is determined, the configuration action path, the corresponding position, and the retained activity generation configurations are written into the same sorting result in a unified order to obtain the activity generation configuration projection set.

[0086] The active generation configuration projection set is associated and solidified to generate an active generation configuration fingerprint, and configuration changes are continuously captured to form a generation change candidate set.

[0087] The specific process includes: organizing the configuration action paths, corresponding positions, and activity generation configurations in the activity generation configuration projection set; fixing the activity generation configurations under the same configuration action path name to the same associated range according to their corresponding positions, so as to maintain a stable correspondence between the configuration action paths and the activity generation configurations; and collectively aggregating the fixed corresponding configuration action paths, corresponding positions, and activity generation configurations into an activity generation configuration fingerprint.

[0088] Furthermore, based on the correspondence in the generational configuration fingerprint, subsequent configuration changes are continuously checked. Configuration changes that are consistent with the generational configuration fingerprint and do not change the correspondence of the configuration action path and the generational configuration affiliation are excluded. Configuration changes that differ from the generational configuration fingerprint and change the corresponding position content or change the generational configuration correspondence under the same configuration action path are retained. The retained configuration changes are then centrally merged according to the configuration action path, and the corresponding configuration action path, corresponding position, and change content are written into the same sorting result to form a generational change candidate set.

[0089] The generational change candidate set and the active generational configuration projection set are merged in a targeted manner according to the configuration action path, and generational reorganization is performed on the changed configuration items to generate a target generational configuration snapshot.

[0090] The specific process includes: matching the configuration changes in the candidate set of generational changes with the active generational configurations in the projection set of active generational configurations; replacing and rearranging the changed configuration items; and centrally aggregating them according to their function paths, performing targeted merging and generational reorganization.

[0091] Furthermore, the configuration content to be merged and reorganized for generations will be expanded one by one according to the configuration action path. The configuration items, corresponding positions and configuration content after replacement and rearrangement under the same configuration action path will be arranged in a centralized manner. All configuration content in the configuration action path will be checked to ensure that the active generational configuration that has not changed and the configuration change content that has been replaced are consistent in position and sequence under the same configuration action path. All updated configuration content under the configuration action path will be collected separately and written into the same organized content in the order of the configuration action path. This will ensure that all updated configuration content is complete, continuous and clearly corresponding, forming a target generational configuration snapshot.

[0092] It should be noted that generational refactoring refers to the incremental configuration reconstruction process that precisely integrates newly captured configuration changes into the currently effective old configuration in order to assemble a new version configuration baseline.

[0093] Based on the target generation configuration snapshot, backtrack along the configuration action path to the corresponding entry in the unified interface mapping scheme, and compare it item by item with the active generation configuration projection set to generate a generation difference relationship table.

[0094] The specific process includes: expanding the configuration action path, corresponding position, and configuration content in the target generation configuration snapshot one by one; looking back at the corresponding entries in the unified interface mapping scheme along the configuration action path; and comparing them with the configuration action path, corresponding position, and active generation configuration in the active generation configuration projection set according to the same position; extracting and organizing the correspondence between different content, different positions, and different functions; and comparing them item by item.

[0095] Furthermore, the different correspondences extracted after item-by-item comparison will be expanded in a centralized manner. The corresponding entries in the configuration action path, corresponding position, configuration content in the target generation configuration snapshot, active generation configuration in the active generation configuration projection set, and unified interface mapping scheme in different correspondences will be kept in a corresponding and organized manner. This will merge the content that differs under the same configuration action path name into the same organized scope, and write it into the same organized information one by one according to the arrangement order of configuration action path and corresponding position, so that the configuration change content, change position and corresponding entry are kept in a continuous correspondence, and generate an intergenerational difference relationship table.

[0096] S5. Based on the target generation configuration snapshot, generation difference relationship table and unified interface mapping scheme, perform differential rewriting, construct a shadow call object set bound to the target generation, and allocate new requests to the shadow call object set.

[0097] Based on the target generation configuration snapshot, generation difference relationship table, and unified interface mapping scheme, the changed configuration items, affected mapping entries, and call paths are associated and expanded to form a differential scope chain.

[0098] The specific process includes: extracting each changed configuration item from the target generation configuration snapshot; verifying the changed configuration items against the difference positions and contents in the generational difference relationship table; merging affected mapping entries that correspond to the same change range into the same scope; tracing back along the affected mapping entries to the corresponding call paths in the unified interface mapping scheme; continuously connecting the changed configuration items, affected mapping entries, and call paths according to their correspondence; and arranging them in a centralized manner according to their correspondence, so that the affected mapping entries and call paths corresponding to the same changed configuration item are merged into the same scope; and verifying and continuously organizing the content of multiple affected mapping entries within the same call path scope, so that the changed configuration items, affected mapping entries, and call paths maintain a clear, complete, and continuous correspondence within the same call path scope, forming a differential scope chain.

[0099] It should be noted that the differential scope chain refers to the boundary chain of the scope of change impact formed by accurately associating and continuously connecting the changed configuration items, affected mapping entries and actual call paths caused by configuration changes.

[0100] The differential scope chain is used to perform targeted reorganization of field rewriting relationships, interaction connection rules and exception normalization rules in the unified interface mapping scheme to generate shadow carrying slices.

[0101] The specific process includes expanding each of the continuously corresponding changed configuration items, affected mapping entries, and call paths in the differential scope chain, extracting the field rewriting relationships, interaction connection rules, and exception normalization rules corresponding to the call paths in the unified interface mapping scheme, and merging the field rewriting relationships, interaction connection rules, and exception normalization rules that correspond to the changed configuration items and affected mapping entries into the same scope. Content that does not fall within the scope of the differential scope chain is removed. The remaining field rewriting relationships, interaction connection rules, and exception normalization rules are rearranged according to the before-after correspondence in the differential scope chain and merged and organized according to the call path, so that the field rewriting relationships, interaction connection rules, and exception normalization rules that need to participate in the change update within the same call path remain continuously corresponding and continuously organized, generating shadow carrying slices.

[0102] It should be noted that field rewriting relationship refers to the set of bidirectional mapping and conversion rules between standardized request fields on the business side and actual interface fields of the underlying specific inference service; interaction connection rule refers to the timing control and protocol adaptation rules between the unified interface and candidate inference services in the process of request initiation, data transmission, capability scheduling and response organization; exception normalization rule refers to the conversion rule that maps the heterogeneous error codes, exception formats and failure states returned by each candidate inference service to a unified standard exception system; shadow carrying slice refers to the targeted extraction and reorganization of target generation lightweight runtime fragments for call paths affected by configuration changes in the differential scope chain.

[0103] Using shadow carrier slices as assembly units, generational binding is performed on connection context, authentication context, field rewriting relationship, interaction connection rules and exception normalization path to construct a shadow call object set.

[0104] The specific process includes: expanding each item in the shadow carrier slice according to the call path; grouping the connection context, authentication context, field rewriting relationship, interaction connection rule, and exception normalization path corresponding to the same call path name into the same scope; verifying the correspondence between each item; and linking the connection context, authentication context, field rewriting relationship, interaction connection rule, and exception normalization path to the generational relationship corresponding to the target generational configuration snapshot item by item, so that all content within the same scope maintains a consistent generational affiliation; centrally aggregating according to the call path; merging the connection context, authentication context, field rewriting relationship, interaction connection rule, and exception normalization path with consistent generational affiliation into call carrier content; and centrally aggregating all call carrier content according to the order of the call paths, so that the call path corresponds to the complete call carrier content, thus constructing a shadow call object set.

[0105] Based on the generational and rewrite states of each call path in the shadow call object set, a path-level admission identifier is generated, and new requests are assigned to the shadow call object set according to the path-level admission identifier.

[0106] The specific process includes: expanding each call path in the shadow call object set one by one; merging the connection contexts, authentication contexts, field rewriting relationships, interaction connection rules, and exception unification paths that have been generationally bound under the same call path name; organizing the content that can indicate the target generation to which the call path belongs to form a generational status; organizing the content that can indicate the field rewriting and interaction connection adjustment of the call path to form a rewriting status; verifying the generational status and rewriting status corresponding to the call path; retaining the call paths with consistent generational status and implemented rewriting status; and writing the retained call paths, corresponding generational status, corresponding rewriting status, and corresponding shadow call object set into the same identifier content to generate a path-level admission identifier.

[0107] Furthermore, when allocating new requests to the shadow call object set based on the path-level admission identifier, the call requirements in the new request are checked item by item against the call path corresponding to the path-level admission identifier. New requests that fall within the same path-level admission identifier range and are consistent with the corresponding generational state and rewrite state are retained. The retained new requests are merged into the shadow call object set corresponding to the path-level admission identifier, and the new requests are allocated to the shadow call object set.

[0108] S6. Based on the shadow call object set, initiate the call according to the unified interface mapping scheme, and rewrite the returned content in a unified manner to generate generational confirmation information and generational rollback information.

[0109] Based on the shadow call object set, the requests allocated to the shadow call object set are rewritten and interactively orchestrated according to the unified interface mapping scheme, and the call is initiated through the shadow call object set to form the original response content.

[0110] The specific process includes expanding each request assigned to the shadow call object set, verifying the task information, request content, and constraint information in the request against the unified interface mapping scheme, and rewriting the content in the unified request fields to the corresponding service fields, so that the request content is consistent with the connection context, authentication context, field rewriting relationship, interaction connection rules, and exception normalization path.

[0111] Furthermore, the request content for field rewriting and interactive orchestration is categorized into the corresponding call path, and the corresponding call path in the shadow call object set initiates a call to the candidate inference service. All content returned by the candidate inference service along the corresponding call path is received one by one and arranged in the order of return. Text content, structured content, status content, and exception content returned under the same request name are merged into the same sorting scope. The returned content is checked to ensure that the service field content, interaction rules, and exception unification rules after field rewriting are consistent with the returned content. All returned content is continuously collected according to the call path and request affiliation to form the original response content.

[0112] The original response content is rewritten and reordered to form a unified response content. The operational status of the target generation is determined, and generational confirmation information and generational rollback information are generated.

[0113] The specific process includes verifying the correspondence between the original response content and the unified interface mapping scheme, rewriting the returned content to the corresponding position of the unified request field, and rearranging it according to the sequential relationship in the interaction rules to form the unified response content.

[0114] Furthermore, the operational status of the target generation is determined. The unified response content is expanded according to the call path and compared with the call path, generation status, and rewrite status in the shadow call object set. Unified response content that is complete, has a complete succession, is in the same position, and has a consistent generation status is included in the confirmation and organization scope. The corresponding call path, generation status, rewrite status, and unified response content are written into the confirmation record to generate generation confirmation information. Unified response content that is incomplete, has an interrupted succession, is in a different position, or is inconsistent with the generation status is included in the rollback and organization scope. The corresponding call path, generation status, rewrite status, unified response content, and abnormal content are written into the rollback record to generate generation rollback information.

[0115] This embodiment also provides a financial full-business world model construction system based on a heterogeneous hypergraph transformer, including: a graph construction module, used to collect capability parameters of candidate inference services and perform semantic normalization and field rewriting to construct a capability semantic graph; a semantic reorganization module, used to perform constraint decomposition and semantic reorganization on the call requests submitted by the business side based on the capability semantic graph to generate a unified request semantic carrier; a mapping generation module, used to map the unified request semantic carrier to the capability semantic graph to generate a unified interface mapping scheme; a generational snapshot module, used to read the currently effective active generational configuration from the unified interface mapping scheme, and generate a target generational configuration snapshot and a generational difference relationship table when a change in the active generational configuration is detected; a shadow assembly module, used to perform differential rewriting based on the target generational configuration snapshot, the generational difference relationship table, and the unified interface mapping scheme to construct a set of shadow call objects bound to the target generation, and allocate new requests to the set of shadow call objects; and a response determination module, used to initiate calls based on the set of shadow call objects according to the unified interface mapping scheme, and perform unified response rewriting on the returned content to generate generational confirmation information and generational rollback information.

[0116] In summary, this invention achieves stable semantic mapping between business-side call requirements and heterogeneous inference service capability boundaries by mapping a unified request semantic carrier to a capability semantic graph and generating a unified interface mapping scheme. This results in unified field rewriting, unified interaction orchestration, and unified exception normalization. Furthermore, by combining the target generation configuration snapshot, generation difference relationship table, and unified interface mapping scheme to perform differential rewriting and construct a shadow call object set, this invention achieves accurate transmission of configuration changes to changes in running objects and smooth transition between old and new generations. This results in low-disturbance switching during configuration hot updates, shadow verification support, and rollback control for exceptions.

[0117] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A multi-model unified interface and dynamic configuration hot update method, characterized in that, include: Collect the capability parameters of candidate inference services, and perform semantic normalization and field rewriting to construct a capability semantic graph; Based on the capability semantic graph, the call requests submitted by the business side are constrained, decomposed, and semantically reorganized to generate a unified request semantic carrier; Map the unified request semantic carrier to the capability semantic graph to generate a unified interface mapping scheme; Read the currently effective active generation configuration from the unified interface mapping scheme, and generate a target generation configuration snapshot and generation difference relationship table when a change in the active generation configuration is detected; Based on the target generation configuration snapshot, generation difference relationship table and unified interface mapping scheme, differential rewriting is performed to construct a set of shadow call objects bound to the target generation, and new requests are assigned to the set of shadow call objects. Based on the shadow call object set, calls are initiated according to a unified interface mapping scheme, and the returned content is uniformly rewritten to generate intergenerational confirmation information and intergenerational rollback information.

2. The multi-model unified interface and dynamic configuration hot update method as described in claim 1, characterized in that, The specific steps for constructing the capability semantic graph are as follows: Collect the raw capability parameters of candidate inference services, and aggregate and organize them according to service identifiers to form a raw capability record set; The original capability record set is used to organize different records of the same capability item, and semantic normalization and field rewriting are performed according to the capability semantic anchor point to form a standard capability item set; Establish the correspondence between candidate reasoning services and capability semantic anchors using a standard capability entry set, and generate a capability semantic graph.

3. The multi-model unified interface and dynamic configuration hot update method as described in claim 2, characterized in that, The specific steps for generating the unified request semantic carrier are as follows: Based on the capability semantic graph, the call request messages submitted by the business side are read, and the task information, request content and constraint information in the call requirements are aggregated and organized to form the original call requirement table; Based on the original call request table, each request item is decomposed and semantically reorganized according to the capability semantic anchor points in the capability semantic graph to form a reorganized constraint chain; The reorganized constraint chain is encapsulated according to a unified carrier structure to generate a unified request semantic carrier.

4. The multi-model unified interface and dynamic configuration hot update method as described in claim 1 or 3, characterized in that, The specific steps for mapping the unified request semantic carrier to the capability semantic graph are as follows: The task information, request content and constraint information in the unified request semantic carrier are matched with the corresponding capability semantic anchors in the capability semantic graph, and the mapping association value between each requirement item and each capability semantic anchor is calculated to form a request anchor correspondence table. Using the request anchor point mapping table as a clue, candidate associations connected to each capability semantic anchor point are extracted from the capability semantic graph to obtain a candidate mapping relationship set.

5. The multi-model unified interface and dynamic configuration hot update method as described in claim 1, characterized in that, The specific steps for generating the unified interface mapping scheme are as follows: The candidate mapping relationship set is subjected to conflict resolution, position correction and order reorganization to form a mapping relationship chain; The mapping relationship chain is combined and arranged with unified request fields, service fields, interaction rules, and exception normalization rules to generate a unified interface mapping scheme.

6. The multi-model unified interface and dynamic configuration hot update method as described in claim 5, characterized in that, The steps for reading the currently effective active generation configuration from the unified interface mapping scheme and detecting a change in the active generation configuration are as follows: Extract the configuration scope path from the unified interface mapping scheme, and read the currently effective active generation configuration based on the configuration scope path to form an active generation configuration projection set; The active generation configuration projection set is associated and solidified to generate an active generation configuration fingerprint, and configuration changes are continuously captured to form a generation change candidate set.

7. The multi-model unified interface and dynamic configuration hot update method as described in claim 1 or 6, characterized in that, The specific steps for generating the target generational configuration snapshot and generational difference relationship table are as follows: The generational change candidate set and the active generational configuration projection set are merged in a targeted manner according to the configuration action path, and generational reorganization is performed on the changed configuration items to generate a target generational configuration snapshot. Based on the target generation configuration snapshot, backtrack along the configuration action path to the corresponding entry in the unified interface mapping scheme, and compare it item by item with the active generation configuration projection set to generate a generation difference relationship table.

8. The multi-model unified interface and dynamic configuration hot update method as described in claim 7, characterized in that, The specific steps for constructing a set of shadow invocation objects bound to the target generation and allocating new requests to the set of shadow invocation objects are as follows: Based on the target generation configuration snapshot, generation difference relationship table and unified interface mapping scheme, the changed configuration items, affected mapping entries and call paths are associated and expanded to form a differential scope chain; The differential scope chain is used to perform targeted reorganization of field rewriting relationships, interaction connection rules and exception normalization rules in the unified interface mapping scheme to generate shadow carrying slices. Using shadow carrier slices as assembly units, generational binding is performed on connection context, authentication context, field rewriting relationship, interaction connection rules and exception normalization path to construct a set of shadow call objects; Based on the generational and rewrite states of each call path in the shadow call object set, a path-level admission identifier is generated, and new requests are assigned to the shadow call object set according to the path-level admission identifier.

9. The multi-model unified interface and dynamic configuration hot update method as described in claim 1, characterized in that, The specific steps for generating intergenerational confirmation information and intergenerational rollback information are as follows: Based on the shadow call object set, the requests allocated to the shadow call object set are rewritten and interactively orchestrated according to the unified interface mapping scheme, and the call is initiated through the shadow call object set to form the original response content; The original response content is rewritten and reordered to form a unified response content. The operational status of the target generation is determined, and generational confirmation information and generational rollback information are generated.

10. A multi-model unified interface and dynamic configuration hot update system, based on the multi-model unified interface and dynamic configuration hot update method according to any one of claims 1 to 9, characterized in that, include: The graph construction module is used to collect the capability parameters of candidate inference services, and perform semantic normalization and field rewriting to construct a capability semantic graph. The semantic reorganization module is used to decompose and semantically reorganize the call requests submitted by the business side based on the capability semantic graph, and generate a unified request semantic carrier. The mapping generation module is used to map the unified request semantic carrier to the capability semantic graph and generate a unified interface mapping scheme. The generational snapshot module is used to read the currently effective active generational configuration from the unified interface mapping scheme, and generate a target generational configuration snapshot and generational difference relationship table when a change in the active generational configuration is detected. The shadow assembly module is used to perform differential rewriting based on the target generation configuration snapshot, generation difference relationship table and unified interface mapping scheme, construct a shadow call object set bound to the target generation, and allocate new requests to the shadow call object set; The response determination module is used to initiate calls based on the shadow call object set according to the unified interface mapping scheme, and to uniformly rewrite the returned content to generate generational confirmation information and generational rollback information.

Images