Business knowledge intelligent assistance method based on power business service scene

Abstract

The application discloses a business knowledge intelligent assistance method based on a power business service scene, relates to the technical field of power business service knowledge assistance, acquires a power business service request and associated business context, generates a business scene signature, acquires a knowledge atom constructed by power business service business knowledge and containing a business applicable boundary and source information, recalls a candidate knowledge atom based on the request and the business scene signature, performs applicability access determination according to the business scene signature and the business applicable boundary, and obtains available knowledge atoms, constructs an evidence set according to the available knowledge atoms and generates a structured handling guide, establishes an evidence mapping relationship of an output item in the guide with the available knowledge atoms, and determines a release, withdrawal, supplementary question or manual review state according to the evidence mapping relationship. The application also discloses a corresponding system, an electronic device and a storage medium, which are used to reduce the risk of business conditions not adapting to knowledge participation in reply and no evidence handling conclusion release.

Description

Technical Field

[0001] This invention relates to the field of knowledge assistance technology for power business services, specifically to an intelligent assistance method for business knowledge based on power business service scenarios. Background Technology

[0002] Current power supply service systems typically rely on power marketing knowledge bases, business hall knowledge bases, intelligent search engines, or large-scale customer service systems to retrieve, match, and generate responses to customer inquiry texts, work order summaries, and content entered by salespersons, in order to support common business inquiries and processing prompts.

[0003] Chinese patent document CN118505237A discloses an intelligent customer service system for power business halls based on a domestically developed large-scale model and its usage method. It collects relevant data from the power industry and customer service interaction records, processes, classifies, annotates entities, and performs sentiment analysis on voice and text data. It also constructs a vector database and a decision tree for diagnosing power problem faults. Combined with query optimization, integration of the vector database and decision tree, construction of typical customer service dialogue scenarios, and model fine-tuning, it provides support for identifying customer inquiries and generating solutions. Another Chinese patent document, CN113094025A, discloses a knowledge-based power business hall service system. This system includes modules for application architecture, application performance, server technology architecture, data architecture, system integration, and security protection information interaction. The application performance component includes knowledge application, knowledge management and operation, an intelligent search engine, document processing services, and an artificial intelligence engine, used to realize knowledge application, guide business acceptance processes, and recommend related knowledge.

[0004] However, in electricity service inquiries, the same natural language question can receive different responses depending on the user's category, electricity usage attributes, geographical location, processing stage, policy version, and document status. Existing solutions primarily focus on three parts: question identification, relevant knowledge retrieval, and response generation. The retrieval or generation is typically based on keyword matching, semantic similarity, entity tags, or historical customer service sample distribution. For semantically similar knowledge fragments applicable to different objects / regions / versions, the system is prone to mixing in inappropriate content during context generation. Furthermore, for processing conclusions regarding materials, time limits, fees, or restrictive prompts, the lack of item-by-item source constraints may lead to untraceable responses.

[0005] If this issue remains unresolved, it will lead to problems such as inconsistent responses from customer service representatives, repeated business corrections, increased pressure on manual review, and contamination of the knowledge base. Therefore, the current challenge is to address how to reduce the risk of knowledge being mismatched with business conditions in responses during the knowledge-assisted process for electricity business services, and how to prevent the generation of conclusions lacking supporting evidence. Summary of the Invention

[0006] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a business knowledge-based intelligent assistance method for power business service scenarios. This method acquires knowledge atoms constructed from power business service knowledge, containing business applicability boundaries and source information; recalls candidate knowledge atoms based on requests and business scenario signatures, and performs applicability admission judgment according to the business scenario signature and business applicability boundaries to obtain usable knowledge atoms; constructs an evidence set based on the usable knowledge atoms and generates structured processing guidelines; establishes an evidence mapping relationship between the output items in the guidelines and the usable knowledge atoms, and determines the status of publication, withdrawal, supplementary inquiry, or manual review accordingly. This method reduces the risk of knowledge incompatible with business conditions participating in responses and the publication of processing conclusions without evidence, thus solving the technical problems described in the background art.

[0007] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: A business knowledge-based intelligent assistance method for power business service scenarios includes: obtaining power business service requests and associated business contexts, generating a business scenario signature; and obtaining knowledge atoms constructed from power business service business knowledge, wherein the knowledge atoms contain business application boundaries and source information. Candidate knowledge atoms are recalled based on request and business scenario signatures, and applicability admission determination is performed according to the business scenario signature and business applicability boundary to obtain usable knowledge atoms; Based on available knowledge atoms, construct an evidence set and generate structured processing guidelines; establish an evidence mapping relationship between the output items of the structured processing guidelines and available knowledge atoms; determine the release status when the mapping is valid and there is no conflict; determine one of the following statuses based on the missing type: withdrawal status, supplementary questioning status, or manual review status when the mapping is missing or there is a conflict.

[0008] Furthermore, the business scenario signature includes business intent, user category, electricity usage attribute, geographical region, processing channel, processing stage, request time, knowledge version, and material status; Among them, business intent, user category, geographical region, processing stage, request time, and knowledge version are key fields. If any key field is missing, a corresponding supplementary question status is generated. If all key fields are present, the process enters the recall of candidate knowledge atoms.

[0009] Furthermore, the business application boundaries of knowledge atoms include the set of applicable conditions, the set of prohibited conditions, the applicable geographical area, the effective time interval, the applicable processing stage, and the type of processing action; the source information includes the source terms and source priority. The applicability admission determination compares each key field with the business applicability boundary, and writes the comparison results into the gating record of the corresponding candidate knowledge atom.

[0010] Furthermore, the applicability admission determination sequentially performs condition matching, prohibition condition determination, region matching, version validity determination, and stage matching; when all determinations pass, the corresponding candidate knowledge atom is determined as a usable knowledge atom; when any determination fails, a gating failure code is generated, and the candidate knowledge atom is marked as a removed knowledge atom, preventing it from entering the evidence set.

[0011] Furthermore, the recall of candidate knowledge atoms includes: keyword recall based on standardized electricity business service requests, vector recall based on semantic similarity, graph adjacency recall based on business intent, processing actions, and material item relationships, and metadata filtering based on geographical location, processing stage, and knowledge version; each recall result is deduplicated to form candidate knowledge atoms.

[0012] Furthermore, when vector recall is in any of the states of response timeout, return exception, or unavailable, candidate knowledge atoms are formed by keyword recall and metadata filtering, and a downgraded recall flag is written into the structured processing guidelines. When vector recall returns normally, its recall results are deduplicated together with the results of keyword recall, graph adjacency recall, and metadata filtering.

[0013] A business knowledge-based intelligent assistance system for power business service scenarios includes a knowledge assistance server, which includes: a request access module for receiving power business service requests; and a business context parsing module for obtaining the business context associated with the power business service request. The business scenario signature generation module is used to generate a business scenario signature based on the power business service request and the business context. The knowledge atomization module is used to obtain knowledge atoms that contain information about the business application boundaries and their sources. The multi-channel candidate recall module is used to recall candidate knowledge atoms based on power business service requests and business scenario signatures. The applicable condition gating module is used to perform an applicability admission determination based on the signature of the business scenario and the business applicable boundary to obtain usable knowledge atoms; The evidence package construction module is used to construct evidence sets based on available knowledge atoms; The structured processing guidance generation module is used to generate structured processing guidance based on the evidence set; The output item evidence consistency verification module is used to establish an evidence mapping relationship between the output items of the structured processing guidelines and the available knowledge atoms, and determine one of the following statuses based on the evidence mapping relationship: release status, withdrawal status, supplementary inquiry status, and manual review status.

[0014] An electronic device includes a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface are connected via the communication bus; The communication interface is used to receive electricity business service requests, the business context associated with the electricity business service requests, and electricity business service knowledge. The memory stores a computer program; when the computer program is executed by the processor, it enables the electronic device to implement a business knowledge-based intelligent assistance method for power business service scenarios.

[0015] A computer-readable storage medium storing a computer program; when the computer program is executed by a processor, it enables the processor to implement a business knowledge-based intelligent assistance method for power business service scenarios.

[0016] (III) Beneficial Effects This invention provides a business knowledge-based intelligent assistance method for electricity business service scenarios, which has the following beneficial effects: By performing applicability admission judgment after recalling candidate knowledge atoms, the signature of the business scenario is compared with the business applicability boundary item by item. This prevents semantically related knowledge atoms that do not match in user category, region, version validity, or processing stage from being excluded from the evidence set, thus reducing the pollution of the response context. By constructing an evidence set based on available knowledge atoms and generating structured processing guidelines, the processing judgment, material list, processing steps, and risk warnings have a corresponding evidentiary basis, avoiding the free generation of responses that mix various types of evidence into a single conclusion.

[0017] By establishing an evidentiary mapping relationship between the output items in the structured processing guidelines and available knowledge atoms, and determining the withdrawal, manual review, or supplementary inquiry status based on missing mappings, knowledge conflicts, or missing business context, material items, time limit items, fee items, and restriction prompts without supporting sources are prevented from being published. By linking the applicability admission judgment with the consistency verification of output item evidence, and combining it with anomaly downgrading and pending review feedback sample pools, knowledge is constrained before entering the response and before the response is published, suppressing the participation of incompatible knowledge in the response and the pollution of the knowledge base by erroneous feedback. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of a knowledge-assisted application scenario for electricity business services provided in an embodiment of the present invention; Figure 2 The overall logical topology diagram of the knowledge evidence constraint system for electricity business services provided in this embodiment of the invention; Figure 3 The main flowchart of the knowledge evidence-constrained intelligent assistance method provided in the embodiments of the present invention; Figure 4 This is a schematic diagram of the signature field structure for a business scenario provided in an embodiment of the present invention; Figure 5 A schematic diagram of the knowledge atom field structure provided in an embodiment of the present invention; Figure 6 A flowchart of the multi-channel candidate knowledge recall process provided in this embodiment of the invention; Figure 7 A gating flowchart for applicable conditions provided in embodiments of the present invention; Figure 8 A hierarchical structure diagram of the role of the evidence package provided in the embodiments of the present invention; Figure 9 This is a flowchart of the output item evidence consistency verification process provided in an embodiment of the present invention; Figure 10 This is a server-side single-sided execution timing diagram provided in an embodiment of the present invention; Figure 11 This is a state machine diagram for anomaly degradation and manual review provided in an embodiment of the present invention; Figure 12 A schematic diagram comparing the ordinary RAG process provided in the embodiments of the present invention with the dual-gate process of the present invention. Detailed Implementation

[0019] 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.

[0020] Please see Figures 1-12 This invention provides a business knowledge-based intelligent assistance method for power business service scenarios.

[0021] Common solutions in existing power business service information systems typically involve using power marketing knowledge bases, power business hall knowledge bases, intelligent search engines, or large-scale customer service systems to handle customer inquiries, 95598 work orders, business hall transactions, and online business hall Q&A for knowledge retrieval or Q&A assistance. While this process can improve problem identification, knowledge retrieval, and natural language expression, it still primarily revolves around how to retrieve relevant content and generate responses. If the same power business service request is constrained by factors such as user category, location, processing stage, policy version, and document status, relying solely on keyword matching, full-text similarity, vector similarity, or ordinary generative Q&A can easily include semantically similar but unsuitable knowledge fragments in the response context. Furthermore, it may lead to the generation of document items, time limits, fee items, or processing conclusions without supporting evidence.

[0022] Therefore, based on the above-mentioned technical bottlenecks, the current challenge is to reduce the risks associated with knowledge-based responses that do not match business conditions and the risk of issuing conclusions without evidence in scenarios involving knowledge assistance for electricity business service consultation and business acceptance.

[0023] To this end, this implementation converts the power business service request and business context into a business scenario signature, constructs business knowledge into knowledge atoms with business applicability boundaries and source information, performs applicability access determination before candidate knowledge enters the response, and performs evidence consistency verification on each output item before the structured processing guidelines are issued.

[0024] like Figure 1 As shown, the power business service knowledge evidence constraint system 100 is deployed between the 95598 call center client 101, the business hall business acceptance client 102, the online business hall client 103, the work order system 104, the customer file system 105, the electricity business ledger 106, the knowledge assistance server 107, the knowledge base server 108, and the manual review terminal 109. Figure 1 The client, system, and server in the system are all abstract communication nodes. Their role is to represent the source of data, the source of knowledge, and the destination of verification. They are not limited to specific brand equipment, specific business hall physical layout, or specific server rack structure.

[0025] like Figure 1 and Figure 2 As shown, after receiving a power business service request, the knowledge-assisted server 107 does not directly send the request to the large model or ordinary knowledge base retrieval engine to generate a response. Instead, the request access module 110, business context parsing module 120, request standardization module 130, and business scenario signature generation module 140 first form a business scenario signature; for example... Figure 2 As shown, the slot integrity determination module 150, knowledge atomization module 160, multi-channel candidate recall module 170, applicable condition gating module 180, evidence package construction module 190, structured processing guidance generation module 200, output item evidence consistency verification module 210, anomaly degradation control module 220, pending review feedback sample pool 230, knowledge version management module 240, templated evidence response module 250 and log archiving module 260 are called sequentially or as needed to form a closed loop of input, processing, output, feedback and anomaly fallback in the system.

[0026] In this embodiment, an electricity service request refers to a natural language question, work order summary, text entered by a salesperson, or a business description input by the 95598 customer service client 101, the business hall service acceptance client 102, or the online business hall client 103. This request could be text such as "I want to apply for a charging pile, what materials are needed?", "How do I apply for peak-valley electricity pricing?", or "What documents are needed for distributed photovoltaic grid connection?". The electricity service request is not a customer satisfaction survey, marketing intention, or operational statistical indicator; rather, it is the raw business text input that enters the data processing flow in this embodiment.

[0027] In this embodiment, the business context refers to the structured or semi-structured business fields associated with the electricity business service request, which may include at least user category, electricity consumption attribute, geographical region, processing channel, processing stage, request time, knowledge version, and material status. User category may include, for example, residential users, non-residential users, micro and small enterprise users, high-voltage users, or temporary electricity users; geographical region may use city, district / county, power supply grid, or business hall area code; processing stage may include consultation, pre-acceptance, document correction, on-site verification, contract confirmation, or archiving. The above fields are read or inherited by the business context parsing module 120 from the work order system 104, customer file system 105, and electricity business ledger 106.

[0028] In this embodiment, the business scenario signature is a structured description of the applicable conditions for the current electricity business service request. Figure 4 In this process, the business scenario signature is formed by inputting the outputs of the request access module 110 and the business context parsing module 120 into the business scenario signature generation module 140, and then into the slot integrity determination module 150.

[0029] A business scenario signature can be represented as follows: S = 〈I, U, P, A, C, R, T, V, E〉 Wherein, S represents the business scenario signature, a set of fields used to describe the applicable boundaries of the current power business service request; I represents the business intent, such as new installation, capacity increase, transfer of ownership, name change, account cancellation, peak-valley electricity pricing application, charging pile application, or distributed photovoltaic grid connection; U represents the user category, such as residential user, non-residential user, micro and small enterprise user, high-voltage user, or temporary electricity user; P represents the electricity attribute, such as electricity usage nature, voltage level, or capacity range; A represents the geographical region, which can be the city, district / county, power supply grid, or business hall's regional code; C represents the processing channel, such as 95598 call center, business hall, or online business hall; R represents the processing stage, such as consultation, pre-acceptance, document correction, on-site verification, contract confirmation, or archiving; T represents the request time, which can be the system reception time or work order time; V represents the knowledge version constraint, which can be determined by T and the effective time and expiration time in the knowledge version management module 240; E represents the material status, such as submitted, missing, pending verification, or unknown. The symbols mentioned above are field identifiers and do not represent abstract scores or purely mathematical optimization variables.

[0030] In some embodiments, I, U, A, R, T, and V are key slots. When performing slot integrity determination, if any key slot is missing, the slot integrity determination module 150 generates a supplementary question or manual review mark, but does not output a definitive processing conclusion. This allows the business scenario signature to convert natural language requests into condition vectors that can be invoked in subsequent knowledge access determination, which helps avoid invoking unsuitable knowledge based solely on text similarity.

[0031] In this implementation, a knowledge atom is the smallest structured unit of evidence extracted from policies, regulations, business guidelines, FAQs, lists of required documents, or internal operating procedures. For example... Figure 5 As shown, the knowledge atom is generated by the knowledge atomization module 160 based on the knowledge documents in the knowledge base server 108, and may include the knowledge atom number 1701, knowledge topic 1702, original evidence text 1703, set of applicable conditions 1704, set of prohibition conditions 1705, scope of applicable area 1706, effective time 1707, expiration time 1708, applicable processing stage 1709, processing action type 1710, source clause 1711, source priority 1712, and conflict relationship 1713.

[0032] Among them, Knowledge Atom Number 1701 is used to uniquely identify a structured evidence unit; Knowledge Topic 1702 is used to identify which business topic this knowledge atom belongs to; Original Evidence Text 1703 is used to store the content of an original clause in the corresponding policy, guide, FAQ, or material list; Applicable Condition Set 1704 is used to limit the user's category, electricity attribute, geographical area, processing stage, or material status when using this knowledge atom; Prohibited Condition Set 1705 is used to limit the situations in which this knowledge atom cannot be used to generate a processable conclusion; Applicable Area Scope 1706 is used to limit the geographical rules; Effective Time and Expiration Time 1707 and 1708 are used to determine the validity of the version; Applicable Processing Stage 1709 is used to distinguish business statuses such as consultation, pre-acceptance, and document correction; Processing Action Type 1710 is used to indicate whether this knowledge atom supports a processing judgment, provides a material prompt, a correction prompt, a restriction prompt, or a manual verification; Source Clause 1711 is used to record the source of evidence; Source Priority 1712 is used to handle conflicts between multiple sources; Conflict Relationship 1713 is used to record the substitution, invalidation, or mutual exclusion relationships with other knowledge atoms.

[0033] Knowledge atoms are not ordinary text slices; ordinary text slices can only represent semantic content, while knowledge atoms simultaneously indicate which type of user, region, version, processing stage, and source of evidence the content applies to. Through the above fieldization process, knowledge atoms can serve as evidence units jointly invoked by the applicable condition gating module 180 and the output item evidence consistency verification module 210.

[0034] In this implementation, the applicability condition gating refers to the following: the applicability condition gating module 180 performs an admission determination based on the business scenario signature and the business applicability boundaries of the candidate knowledge atoms before they enter the evidence set. The admission determination may include user category matching, geographic region matching, version validity matching, processing stage matching, prohibition condition hit determination, and material status determination. Applicability condition gating is not a simple sorting, similarity filtering, or keyword filtering; its engineering function is to block semantically similar but business-conditionally incompatible knowledge from entering the subsequent response process.

[0035] In this embodiment, the evidence set, also known as the evidence package, is a structured evidence context composed of valid knowledge atoms 187 gated by applicable conditions. For example... Figure 8 As shown, the evidence package construction module 190 divides the effective knowledge atom set into layers to form the main evidence set 191, the conditional evidence set 192, the restrictive evidence set 193, the process evidence set 194, and the conflict evidence set 195, and outputs the layered evidence to the structured processing guidance generation module 200 and the output item evidence consistency verification module 210.

[0036] In this embodiment, the structured processing guidelines are generated by the structured processing guidelines generation module 200 based on... Figure 8 The output object generated from the evidence set shown, and further as follows: Figure 9 The output item evidence consistency verification module 210 verifies each item. The structured processing guide may include business type, confirmed conditions, missing conditions, processing judgment, material list, processing steps, risk warning, evidence number, and review mark. Unlike free customer service scripts, each substantive output item of the structured processing guide can be individually mapped to a knowledge atom by the output item evidence consistency verification module 210.

[0037] In this embodiment, output item evidence mapping refers to the output item evidence consistency verification module 210 associating each item in the structured processing guide—processing conclusion, material items, time limit items, fee items, and restriction prompts—with the knowledge atom number or source clause. For example... Figure 9 As shown, the output item mapping unit 211 completes the correspondence between output items and knowledge atoms, the evidence coverage calculation unit 212 determines the coverage status, the conflict output withdrawal unit 213 processes conflict outputs, and the manual review triggering unit 214 generates review tasks.

[0038] For example, the following mapping can be formed: M: oᵢ ↦ (atom_idⱼ, source_refⱼ) Where M represents the output item evidence mapping relationship; oᵢ represents the i-th substantive output item in the structured processing guide, where i is the output item number; atom_idⱼ represents the number of the j-th knowledge atom, where j is the knowledge atom number; and source_refⱼ represents the source clause or source document location of the j-th knowledge atom.

[0039] Furthermore, every publishable output item must be able to link back to its source of evidence; if an output item cannot establish a mapping or conflicts with the prohibition conditions, the output item is withdrawn through the output item evidence consistency verification module 210 or manually reviewed.

[0040] In some embodiments, evidence coverage can be calculated by dividing the number of substantive output items for which evidence mapping relationships have been established by the total number of all substantive output items, as shown below: C e = N m / N o Among them, C e Indicates evidence coverage; N m N represents the number of substantive output items for which an evidence mapping relationship has been established. o This indicates the total number of all substantive output items in the structured processing guide.

[0041] Substantive output items refer to output content that affects the processing action, such as processing conclusions, material items, time limits, fees, and restriction notices, but do not include formatted prompts, greetings, or fixed explanatory texts. When N o When the value is 0, the evidence coverage calculation unit 212 does not perform division calculations, but marks the current guidance as having no substantial output and transfers it to manual review or regeneration. This supplementary definition gives the evidence coverage a clear input object and triggering action.

[0042] In this embodiment, the gate failure code is a failure reason record generated by the applicable condition gate control module 180 for candidate knowledge atoms that have not passed the admission judgment. The gate failure code may include: user category mismatch, region mismatch, knowledge version failure, processing stage mismatch, prohibition condition hit or key slot missing, etc., and is used to write traceable records to the log archiving module 260, as well as for subsequent knowledge maintenance and manual review.

[0043] In this embodiment, the feedback sample pool 230 is used to store manually rewritten data by sales staff, complaint-related results, missed issues, knowledge conflicts, and manual review conclusions. The feedback sample pool 230 does not directly train the model or directly rewrite the knowledge base. Instead, it triggers the knowledge version management module 240 to revise the knowledge atoms after manual review, thereby reducing the risk of erroneous feedback polluting the knowledge base.

[0044] like Figure 1 As shown, the Electricity Business Service Knowledge Evidence Constraint System 100 includes a 95598 call center client 101, a business hall service acceptance client 102, an online business hall client 103, a work order system 104, a customer file system 105, an electricity business ledger 106, a knowledge assistance server 107, a knowledge base server 108, and a manual review terminal 109. The knowledge assistance server 107 can be deployed as an intranet server, cloud server, edge computing node, or a server cluster consisting of multiple service instances. Internally, it runs the following modules: request access module 110, business context parsing module 120, request standardization module 130, business scenario signature generation module 140, slot integrity judgment module 150, knowledge atomization module 160, multi-channel candidate recall module 170, applicable condition gating module 180, evidence package construction module 190, structured processing guidance generation module 200, output item evidence consistency verification module 210, anomaly degradation control module 220, pending review feedback sample pool 230, knowledge version management module 240, templated evidence response module 250, and log archiving module 260. These modules can be microservices, function services, rule engine components, database access components, or code logic encapsulation deployed within the same server process.

[0045] like Figure 2 As shown, the request access module 110, business context parsing module 120, request standardization module 130, and business scenario signature generation module 140 together constitute the request-side preprocessing link; the knowledge atomization module 160, multi-channel candidate recall module 170, and applicable condition gating module 180 together constitute the knowledge access link; the evidence package construction module 190, structured processing guidance generation module 200, and output item evidence consistency verification module 210 together constitute the pre-release evidence constraint link; the anomaly degradation control module 220, pending review feedback sample pool 230, knowledge version management module 240, templated evidence response module 250, and log archiving module 260 are used for anomaly fallback, feedback review, version revision, and log recording.

[0046] like Figure 1 As shown, the request access module 110 is communicatively connected to the 95598 customer service client 101, the business hall service acceptance client 102, and the online business hall client 103, respectively, to receive electricity business service requests. For example, when a customer submits a request through the 95598 customer service center asking what materials are needed, the 95598 customer service client 101 sends this text along with the session number, customer service number, and timestamp to the request access module 110. The request access module 110 does not directly generate a response to this text; instead, it generates a request number and sends the text and source information to the request standardization module 130. This process avoids the propagation of ambiguity caused by directly entering the knowledge retrieval process with the original, colloquial request.

[0047] like Figure 1 and Figure 4 As shown, the business context parsing module 120 uses the request number, customer number, or work order number as an index to read user category, electricity attribute, geographical location, processing channel, processing stage, request time, knowledge version, and material status from the work order system 104, customer file system 105, and electricity business ledger 106. For example, the customer file system 105 can return residential users, the electricity business ledger 106 can return low-voltage electricity, and the work order system 104 can return the consultation stage.

[0048] In one implementation, when the system cannot obtain the user category from the customer profile system 105, the business context parsing module 120 can mark the user category as an empty slot, and the slot integrity determination module 150 will trigger a replacement query, instead of using a default category. This prevents the system from incorrectly applying the bill of materials when the user category is missing.

[0049] like Figure 4As shown, after receiving the text output by the request access module 110, the request standardization module 130 performs operations such as sentence segmentation, noise reduction, synonym normalization, and business dictionary mapping. For example, "apply for peak and valley pricing," "open time-of-use pricing," and "time-of-use pricing application" can be normalized to "peak and valley pricing application"; "install electricity meter" and "new account opening" can be mapped to "low-voltage new installation" or "high-voltage new installation" based on the context. The request standardization module 130 can use a combination of rule dictionaries, word segmentation components, entity recognition models, or regular expressions to accomplish this. After obtaining the standardized request text, the business scenario signature generation module 140 then identifies the business intent and slot. The advantage of this approach is that it compresses colloquial expressions into standard items in the business dictionary, reducing the likelihood of missed or false recalls due to synonyms during subsequent recall.

[0050] like Figure 4 In this module, the business scenario signature generation module 140 generates a business scenario signature based on the standardized request text and the business context. For example, when inquiring about personal charging pile installation materials, the business scenario signature can be represented as follows: S = <Personal charging pile installation application, residential user, low-voltage electricity, a certain power supply area, 95598 customer service hotline, inquiry, current request time, current knowledge version, material status unknown> The fields in the above business scenario signature correspond to the business intent I, user category U, electricity attribute P, geographical region A, processing channel C, processing stage R, request time T, knowledge version V, and material status E, respectively.

[0051] The current request time and current knowledge version do not require a specific date or version number to be specified in the claims; their engineering significance lies in serving as input for determining version validity. If the request originates from the business hall client 102, the processing channel field can be written as "business hall"; if the request originates from the online business hall client 103, the processing channel field can be written as "online business hall". By using the business scenario signature as the basis for subsequent gating, the same natural language question can enter different knowledge access paths under different user categories, geographical regions, or processing stages.

[0052] like Figure 4 As shown, the slot integrity determination module 150 determines the key slots for the business scenario signature. These key slots may include information such as business intent, user category, geographic region, processing stage, request time, and knowledge version. For example, if the user category or geographic region is missing, the slot integrity determination module 150 will not allow the structured processing guidance generation module 200 to output a definitive material list, and will instead generate a supplementary question requesting confirmation of the processing region and user category before generating an applicable material list.

[0053] In one implementation, if only the processing channel is missing, and the processing channel does not affect the current business material list, the system can continue to recall candidates and mark the processing channel as unconfirmed in the structured processing guide. This not only prevents important missing information that would affect the applicability of knowledge, but also avoids unnecessary process interruptions caused by the absence of non-important fields.

[0054] like Figure 5 As shown, the policies, regulations, business guidelines, FAQs, application material lists, and internal operating procedures read from the knowledge base server 108 by the knowledge atomization module 160 are broken down into knowledge atoms. The knowledge atomization module 160 can segment each knowledge atom according to clause boundaries, table rows, material items, process nodes, or question-and-answer pairs. Each knowledge atom contains: original evidence text 1703; set of applicable conditions 1704; set of prohibition conditions 1705; applicable geographical scope 1706; effective date 1707; expiration date 1708; applicable processing stage 1709; processing action type 1710; source clause 1711; source priority 1712; and conflict relationship 1713, etc.

[0055] For example, a list of application materials for a residential personal charging pile can be marked as applicable to residential users, low-voltage electricity, designated power supply area, and currently valid version; a list of application materials for a corporate charging facility can be marked as applicable to non-residential users or corporate users. This field-based approach gives ordinary text paragraphs business applicability boundaries that can be determined by machines.

[0056] like Figure 5 As shown: In one implementation, for an old knowledge base that has not yet been finely annotated, the knowledge atomization module 160 can first use one or more paragraphs in the original document as a temporary knowledge atom, and require knowledge reviewers to supplement the applicable conditions, prohibition conditions, regional scope, version validity and source clauses before allowing it to enter the applicable conditions gating module 180. This approach is suitable for the migration of existing knowledge bases, and can avoid the old knowledge base text from directly participating in the response generation process without being annotated with the applicable boundaries, which can reduce the risk of erroneous output at the beginning of the migration.

[0057] like Figure 6As shown, the multi-channel candidate recall module 170 recalls candidate knowledge atoms based on business scenario signatures and standardized request text. The multi-channel candidate recall module 170 includes: a keyword recall unit 171, a vector recall unit 172, a graph adjacency recall unit 173, and a metadata filtering unit 174. The keyword recall unit 171 can recall candidates based on a business dictionary, policy name, or material name using an inverted index; the vector recall unit 172 can recall candidates based on the semantic similarity between the request text and the knowledge atom text; the graph adjacency recall unit 173 can recall relevant knowledge atoms based on business intent, processing actions, and relationships between material items; and the metadata filtering unit 174 can prioritize filtering candidates that do not belong to a specific region, business stage, or knowledge version.

[0058] In some embodiments, Top-K refers to the top K candidate knowledge atoms (where K is a configurable positive integer) retained during the candidate recall phase based on the results of keyword hit / semantic similarity / graph adjacency or metadata matching. The specific value of this parameter can be determined based on the knowledge base size, response latency requirements, recall coverage requirements, and subsequent gating computation load. Furthermore, Top-K is not used to directly publish the handling conclusion; it still needs to be approved by the applicable condition gating module 180.

[0059] In one implementation, when vector indexing cannot be used in the deployment environment, a combination of keyword recall and metadata filtering can be used to transfer low similarity or high ambiguity issues to manual review. This approach can support multi-channel candidate recall module 170 under conditions of low computing power and strong intranet isolation.

[0060] In one implementation, when vector indexes cannot be used in the deployment environment, a combination of keyword recall and metadata filtering can be used, and issues with low similarity or high ambiguity can be transferred to manual review, thus realizing the multi-channel candidate recall module 170. This approach can support situations with low computing power and strong intranet isolation.

[0061] like Figure 7 As shown, in the applicable condition gating module 180, after receiving the set of candidate knowledge atoms output from the multi-channel candidate recall module 170, the set of candidate knowledge atoms is sequentially passed through the condition matching unit 181, the prohibition condition judgment unit 182, the region matching unit 183, the version validity judgment unit 184, and the stage matching unit 185 for the admission judgment of applicability. If any key judgment fails, the gating failure encoding unit 186 will record the reason information of the failure. If all key judgments pass, the valid knowledge atom 187 is output to the evidence package construction module 190.

[0062] Gating failure coding can be represented as a structured record of the reasons for failure, such as: user category mismatch; region mismatch; knowledge version failure; processing stage mismatch; prohibited conditions hit or key slots missing, etc., and is written to the log archiving module 260 through gating failure coding unit 186, so that the specific reasons for the candidate knowledge being removed can be proved during subsequent manual review, knowledge maintenance and review.

[0063] like Figure 7 As shown, taking the consultation on personal charging pile installation materials as an example, the multi-channel candidate recall module 170 may recall candidate knowledge atoms such as personal charging pile installation materials for residents, charging facility installation materials for organizations, old versions of personal charging pile materials, material lists for areas outside the local area, and material missing correction requirements. The applicable condition gating module 180 determines the personal charging pile installation materials for residents as valid knowledge atoms 187; removes the charging facility installation materials for organizations based on user category mismatch; removes the old versions of personal charging pile materials based on version expiration; removes the material lists for areas outside the local area based on region mismatch; and retains the material missing correction requirements as limiting evidence or procedural evidence. Even if multiple candidate texts are semantically related to charging pile materials, only candidate knowledge that matches the business conditions can enter the evidence set.

[0064] In one implementation, such as Figure 7 As shown, the applicability condition gating module 180 can also be implemented using a parallel approach of rule gating and a lightweight classifier. Rule gating is applied to key fields such as user category, geographic region, request time, and processing stage. For non-rigid fields extracted from semi-structured text, such as material status and processing conditions, a lightweight classifier can output candidate applicability scores (but these are only candidate suggestions). Whether a field is included in the main evidence still depends on the rule-based judgment of the key fields. This alternative approach avoids the black-box nature caused by pure model output and considers the usability of semi-structured knowledge conditions.

[0065] like Figure 8 As shown, in the evidence package construction module 190, an evidence set is constructed through the applicable condition gating module 180 and valid knowledge atoms 187: The main evidence set 191 supports the judgment of whether the process is permissible / inapplicable or requires correction; the conditional evidence set 192 explains the applicable prerequisites; the restrictive evidence set 193 explains prohibitive conditions, risk warnings, or matters requiring manual verification; the process evidence set 194 explains the material list, processing steps, or time limits; and the conflict evidence set 195 retains information such as old and new versions, regional differences, and source priority conflicts. After role stratification, the structured processing guidance generation module 200 no longer deals with unstructured Top-K text fragments, but with evidence sets that can support different output items.

[0066] like Figure 8 As shown, the structured processing guidance generation module 200 generates structured processing guidance based on the evidence set. This guidance can include information such as: business type, confirmed conditions, missing conditions, processing eligibility / ineligibility judgment, required correction markers, material list, processing steps, risk warnings, evidence number, and review markers. For example, when inquiring about materials for personal charging pile installation, the structured processing guidance generation module 200 can output: business type is personal charging pile installation; user category is residential user; processing stage is inquiry; the material list is supported by the process evidence set; if the material status is unknown, the customer should be prompted to supplement the material status or transfer to manual confirmation.

[0067] The above output is generated in the form of structured fields, which facilitates the mapping of each output item by the evidence consistency verification module 210.

[0068] In one implementation, such as Figure 8 As shown, the structured processing guidance generation module 200 can also use a templated evidence response module 250 instead of a large model to replace the large model generation: the templated evidence response module 250 fills in the main evidence for processing judgment; fills in the material list and steps for process evidence; fills in the risk warning for restrictive evidence; and converts conflicting evidence into review marks. This approach is suitable for intranet deployment or low-computing-power deployment and situations where calling external large models is not allowed. Furthermore, because this approach still retains the access control for applicability and the consistency verification of output item evidence, its implementation mechanism is the same as the main embodiment and does not limit the scope of protection to a specific large model tool.

[0069] like Figure 9 As shown, in the output item evidence consistency verification module 210, after receiving the outputs from the structured processing guidance generation module 200 and the evidence package construction module 190, evidence mapping is performed for each substantive output item: The output item mapping unit 211 maps the processing conclusion, material items, time limit items, fee items, and restriction prompts to the available knowledge atom numbers and source clauses, respectively; the evidence coverage calculation unit 212 determines the evidence coverage status; the conflict output withdrawal unit 213 withdraws output items that have not been mapped or conflict with prohibition conditions; and the manual review triggering unit 214 generates a manual review task when knowledge conflicts, insufficient evidence, or missing key slots occur, and sends it to the manual review terminal 109. This ensures that the status of the generated result is determined by the source of evidence, rather than by the fluency or similarity results of the language model.

[0070] like Figure 9If the structured processing guidance generation module 200 outputs a material item, but the output item mapping unit 211 cannot find the corresponding source clause from the knowledge atom, the conflict output withdrawal unit 213 will mark the material item as unpublishable. When the material item conflicts with a prohibition condition of a knowledge atom, the manual review trigger unit 214 will hand the guidance over to the manual review terminal 109 for manual review. Here, withdrawal is not a simple alarm; it changes the publication status of the output item to prevent conclusions without evidence from being displayed to agents or customers.

[0071] like Figure 10 As shown, in the server-side single-sided execution embodiment, clients 101, 102, or 103 send a processing request to the knowledge assistance server 107; after obtaining the session context, the knowledge assistance server 107 initiates a knowledge query request to the knowledge base server 108; the knowledge base server 108 returns knowledge retrieval results such as rules, templates, cases, or evidence requirements; the knowledge assistance server 107 performs gating and reasoning processing, evidence package construction, and structured processing guidance output, and returns the structured processing guidance to the client; when there are conflicts in the output items or insufficient evidence, the knowledge assistance server 107 dispatches a manual review task to the manual review terminal 109, and generates a final return result based on the pass, withdrawal, supplementary question, or review opinion returned by the review terminal 109. This timing illustrates that the core processing of this embodiment can be completed unilaterally by the knowledge assistance server 107, facilitating the formation of server-side method, system, and electronic device embodiments.

[0072] like Figure 11 As shown, the anomaly degradation control module 220 controls the state machine for normal processing S1, key slot missing S2, knowledge version conflict S3, insufficient evidence coverage S4, index service failure S5, generation service failure S6, and feedback pollution risk S7. Specifically, for key slot missing S2, a supplementary question is generated; for knowledge version conflict S3, the manual review terminal 109 is triggered; for insufficient evidence coverage S4, the output item is withdrawn or an evidence summary is output; for index service failure S5, keyword recall is switched; for generation service failure S6, the response is downgraded to a template response or an empty response; and for feedback pollution risk S7, the response is written into the feedback sample pool 230 pending review. After the problem is resolved / review is passed / service is restored, the system returns to normal processing S1.

[0073] The above anomaly handling ensures that the system maintains a closed loop even when faced with weak data, weak retrieval, or weak generation, and does not output uncontrolled responses.

[0074] like Figure 11As shown, in one implementation, the anomaly degradation control module 220 can set the number of retries and fallback links. For example, when the vector recall unit 172 fails to return results within a preset response time limit in two consecutive requests, the anomaly degradation control module 220 does not continue waiting for vector recall. Instead, it calls the keyword recall unit 171 and the metadata filtering unit 174, and adds internal tags generated based on the degradation recall to the structured processing guidance. These tags are not displayed to the customer but are used by the manual review terminal 109 to determine whether manual confirmation is required, thus reducing the impact of search service failures on the overall process.

[0075] like Figure 1 and Figure 11 As shown, the feedback sample pool 230 receives review conclusions written back by the manual review terminal 109, records manually revised by sales staff, customer follow-up questions, complaint associations, missed questions, and knowledge conflict records. The feedback sample pool 230 does not directly rewrite the knowledge atoms in the knowledge base server 108; instead, it submits the feedback samples to the knowledge version management module 240 for review. Upon successful review, the knowledge version management module 240 updates the source terms, applicable conditions, validity period, or conflict relationships of the knowledge atoms. This process prevents erroneous feedback from directly contaminating the knowledge base and ensures version traceability in the knowledge revision process.

[0076] like Figure 12 As shown, a typical RAG process generally includes: request, Top-K recall, and response generation. However, the dual-gate process in this embodiment sequentially forms a business scenario signature and knowledge atoms after the request and business context. An applicability condition gating module 180 is set as the first gate before generating the structured processing guide, and an output item evidence consistency verification module 210 is set as the second gate before publication, withdrawal, or manual review. This avoids the problems of insufficient business applicability and evidence sources caused by relying solely on text relevance for recall and freely generating responses.

[0077] like Figure 1 and Figure 4 As shown, in the first parallel embodiment, the electricity business service request originates from the business hall business acceptance client 102, and the business scenario is the application for peak-valley electricity pricing for residential customers. After the salesperson enters the conditions required for a residential customer to activate peak-valley electricity pricing, the business context parsing module 120 reads the user category as residential user from the customer file system 105, the processing stage as consultation or pre-acceptance from the work order system 104, and the power supply area code from the location field. The business scenario signature generation module 140 generates a business scenario signature including peak-valley electricity pricing application, residential user, location, processing stage, request time, and current electricity price policy version.

[0078] In the above diagram, the multi-channel candidate recall module 170 may recall the application conditions for residential peak-valley electricity pricing, non-residential time-of-use electricity pricing rules, old versions of peak-valley electricity pricing policies, and application materials from other regions; the applicable condition gating module 180 will eliminate non-residential rules, old policies, and rules from other regions, leaving only the knowledge atoms related to residential peak-valley electricity pricing in the current region and the current version; the output item evidence consistency verification module 210 will verify the source clauses of the application conditions, material items, and effective period one by one. This parallel embodiment can prove the access control and evidence verification in this case, which is not limited to the charging pile application scenario, but can also be used in electricity price application scenarios.

[0079] like Figure 1 and Figure 8 As shown in the second parallel embodiment, the electricity business service request originates from the online business hall client 103, and the business scenario is distributed photovoltaic grid connection consultation: after the customer enters "How to apply for household photovoltaic grid connection", the business context parsing module 120 reads the user's category, location, and application channel information. If capacity verification, grid connection technical conditions, or on-site acceptance conditions are involved, the structured application guidance generation module 200 only outputs the list of materials and process guidance that can be publicly disclosed during the consultation stage, but marks the technical verification items regarding whether the capacity meets the grid connection requirements as requiring manual review. At this point, the limiting evidence set 193 and the conflicting evidence set 195 determine which output items cannot be automatically published. This avoids the system from overstepping its boundaries by outputting distribution network capacity verification or engineering design conclusions while still being able to output evidence-mapped business service application guidance.

[0080] like Figure 1 and Figure 7 As shown, in the third parallel embodiment, the electricity business service request originates from the work order system 104, and the business scenario is the supplementation of documents for a transfer or name change. The work order system 104 indicates that the current processing stage is document supplementation, and the material status shows that some certificates are missing. The business scenario signature generation module 140 sets the processing stage field to document supplementation and the material status field to missing. The applicable condition gating module 180 does not use the general processing procedure of the consultation stage as the main evidence, but prioritizes recalling and retaining the knowledge atoms of the document supplementation stage. The structured processing guidance generation module 200 outputs the material items that need to be supplemented and the corresponding source clauses. The processing stage field prevents the mixing of knowledge from the consultation stage and knowledge from the document supplementation stage.

[0081] like Figure 4 and Figure 10As shown, in a parallel implementation with different overall routes but the same goal, the business scenario signature generation module 140 can be deployed within the business hall acceptance client 102 on the edge side, while the applicable condition gating module 180 and the output item evidence consistency verification module 210 are still deployed within the knowledge assistance server 107. The business hall acceptance client 102 first completes the local business dictionary normalization and partial slot extraction, and then sends the business scenario signature draft to the knowledge assistance server 107. After reviewing the key slots, the knowledge assistance server 107 continues to perform candidate recall and gating.

[0082] This parallel implementation method is suitable for scenarios with high response requirements in the business hall intranet, and can reduce the load of repetitive text preprocessing on the server side, but does not change the dual-gate mechanism of this case.

[0083] like Figure 1 and Figure 10 As shown, in another parallel implementation, the system employs centralized server-side processing. The 95598 call center client 101, the business hall service client 102, and the online business hall client 103 are only responsible for uploading the request text and context fields. All request standardization, scenario signing, candidate recall, gating, evidence package construction, and output verification are performed by the knowledge-assisted server 107. This approach is suitable for deployment on provincial or municipal platforms with unified operation and maintenance, unified knowledge version management, and unified log auditing. Regardless of whether edge preprocessing or centralized server processing is used, the core principles are the access determination driven by the applicable boundaries of business scenario signatures and knowledge atoms, and the evidence mapping between output items and knowledge atoms, without altering the higher-level protection boundaries of this case.

[0084] like Figure 11 As shown: When a communication interruption occurs, if the 95598 customer service client 101, the business hall service client 102, or the online business hall client 103 sends a request to the knowledge assistance server 107 and does not receive a response within the set time range, the request access module 110 will set the request to a receipt timeout. The exception degradation control module 220 can first determine whether the request has been written to the request queue. If not, the client will resend the request. If it has been written but the processing fails later, the client will be informed that the request has been received and requires manual review. This can avoid the situation where repeated startups cause different processing instructions for the same customer's request.

[0085] like Figure 11As shown, in cases of missing data, if the business context parsing module 120 cannot obtain the user category from the customer file system 105, or the geographical region from the work order system 104, the slot integrity judgment module 150 will mark the business scenario signature as a critical slot missing; the exception degradation control module 220 will generate a supplementary question instead of allowing the applicable condition gating module 180 to continue matching with null values. For example, the system could return, "Please confirm the user category and processing region before generating the applicable materials list." This would prevent the issuance of a definitive response under missing conditions.

[0086] like Figure 11 As shown, in the case of inconsistent status, if the user category returned by the customer file system 105 is residential user, but the electricity usage attribute returned by the electricity usage business ledger 106 is high-voltage industrial and commercial electricity, then the business context parsing module 120 determines that there is a conflict in the context fields of this request. Furthermore, the exception degradation control module 220 will not call the ordinary candidate recall to generate the material list, but will generate the task and conflict fields for manual review. This can avoid the incorrect use of the rules for residential users or industrial and commercial users when the fields of different business systems are inconsistent.

[0087] like Figure 11 As shown, in the case of a data frame verification failure exception, if the request message sent by the client does not contain a request number or source channel and timestamp, the request access module 110 will refuse to let the message enter the subsequent gating process and will return a resubmission instruction to the client; if the main text in the message is present but the context field verification fails, the request access module 110 will retain the text and have the business context parsing module 120 retrieve the context again to avoid errors in the subsequent knowledge access judgment basis due to incomplete messages.

[0088] like Figure 11 As shown, in the abnormal situation of vector index service failure, if the vector recall unit 172 times out or returns an exception, the multi-channel candidate recall module 170 will not directly end the process, but will switch to the combination of keyword recall unit 171, graph adjacency recall unit 173 and metadata filtering unit 174 to recall candidates. At this time, the structured processing guidance generation module 200 can record the downgraded recall in its internal state, and the output item evidence consistency verification module 210 will improve the sensitivity of the review trigger. Even if semantic recall is temporarily unavailable, it can still generate controlled output or review tasks according to the evidence boundary.

[0089] like Figure 11As shown, in the event of a service failure exception, if the large model controlled generator of the structured processing guidance generation module 200 times out or becomes unavailable, the templated evidence response module 250 can take over the output process: the templated evidence response module 250 fills the main evidence set 191 into the processing judgment field, fills the process evidence set 194 into the material list and step fields, fills the restriction evidence set 193 into the risk warning field, and converts the conflict evidence set 195 into a manual review mark; this can prevent the generator failure from causing the entire knowledge assistance link to be interrupted, and can also prevent the system from outputting unverified free text when no results are generated.

[0090] like Figure 9 and Figure 11 As shown, when a parameter exceeds the limit, if the evidence coverage is less than the publishable limit, the output item evidence consistency verification module 210 will not publish the complete structured processing guide, but will withdraw the output items that were not mapped and trigger manual review. The specific publication threshold for evidence coverage can be determined by the deployment unit based on comparative testing and business expert review (a specific data should not be hardcoded before actual test data is obtained). This ensures that the parameter boundary is tied to the sampled evidence.

[0091] like Figure 1 and Figure 11 As shown, when feedback on abnormal pollution risks occurs, if the result of manual rewriting by the salesperson conflicts with the source terms of the existing knowledge atoms, the feedback sample pool 230 to be reviewed will not be directly written back to the knowledge base server 108, but will be submitted to the manual review terminal 109 for manual review. Only after the knowledge version management module 240 approves the feedback will the applicable conditions or source terms of the knowledge atoms, or the conflict relationship be updated accordingly, thereby reducing the risk of erroneous feedback, emotionally charged complaints, or temporary handling methods polluting the official knowledge base.

[0092] like Figure 1 and Figure 2 As shown, this embodiment can also be manifested as an intelligent auxiliary virtual device for knowledge evidence constraints in power business services. This virtual device can be deployed in the knowledge assistance server 107, or it can be deployed as multiple service components in a server cluster. The virtual device includes a request acquisition module, a context parsing module, a scene signature generation module, a knowledge atom construction module, a candidate recall module, an applicability admission judgment module, an evidence set construction module, a processing guidance generation module, an evidence consistency verification module, and an exception handling module. These modules are not merely abstract functional names, but can respectively correspond to the API gateway service, database access service, natural language processing service, rule engine service, retrieval service, evidence mapping service, and task queue service in the server.

[0093] like Figure 1 and Figure 4 As shown, the request acquisition module can correspond to the request access module 110. Its underlying form can be an HTTP interface service, a message queue consumer, an RPC server, or an intranet interface adapter. This module accepts request texts from the 95598 call center client 101, the business hall service acceptance client 102, or the online business hall client 103, and generates a request number for the context parsing module to process. In case of an exception, it can decide to return a retransmission instruction based on whether the fields in the message are complete, or write the request to the exception queue for subsequent scenario signature generation to provide a unique and traceable request entry point.

[0094] like Figure 1 and Figure 4 As shown, the context parsing module can correspond to the business context parsing module 120. Its underlying form can be a database query service, a cache access service, or a business interface orchestration service, etc. It is responsible for reading various fields of the business context from the work order system 104, the customer file system 105, and the electricity business ledger 106, and transmitting the reading results to the scenario signature generation module. When a certain field is missing, it does not make any guesses but hands over the missing state to the slot integrity judgment module 150 for judgment. This makes supplementary questions, reviews, and degradation have clear data sources.

[0095] like Figure 4 As shown, the scenario signature generation module can correspond to the business scenario signature generation module 140 and the slot integrity determination module 150. Its underlying form can be a combination of business dictionary, sequence labeling model, classifier and rule state machine. It generates business scenario signatures and outputs supplementary fields when key slots are missing. Its output directly enters the candidate recall module and the applicability admission determination module as the basis for subsequent gating.

[0096] like Figure 5 As shown, the knowledge atom building module corresponds to the knowledge atomization module 160, and its underlying structure can be document parsing service, table parsing service, manual review interface, knowledge version indexing service, and metadata writing service. This module breaks down policies, FAQs, business guidelines, and material lists into knowledge atoms, and writes applicable conditions, prohibition conditions, regional scope, version validity, processing stage, and source clauses for each knowledge atom. It converts ordinary knowledge base documents into evidence units with executable gating and executable mapping.

[0097] like Figure 6As shown, the candidate recall module can correspond to the multi-channel candidate recall module 170. Its underlying form can be a combination of keyword retrieval service, vector retrieval service, graph query service and metadata filtering service. It will output a set of candidate knowledge atoms, and will not output them directly to the generation module. Instead, it needs to be processed by the applicability admission judgment module to avoid the results of ordinary Top-K recall from entering the response context without being filtered by business conditions.

[0098] like Figure 7 As shown, the applicability admission judgment module can correspond to the applicability condition gating module 180. The underlying form can be a rule engine, a Boolean expression matcher, a version validity judge, and a conflict detector. It compares the business scenario signature with the applicability boundary of the candidate knowledge atom and outputs the valid knowledge atom 187, the removed knowledge atom, and the gating failure code. In this way, it can block those semantically similar but unsatisfactory knowledge from participating in the structured processing guidance.

[0099] like Figure 8 As shown, the evidence set construction module and the processing guidance generation module can correspond to the evidence package construction module 190 and the structured processing guidance generation module 200, respectively. The evidence set construction module divides the effective knowledge atoms 187 into main evidence, conditional evidence, limiting evidence, procedural evidence, and conflicting evidence; the processing guidance generation module outputs structured processing guidance based on the above evidence set. This structured output allows subsequent evidence mapping to be executed item by item.

[0100] like Figure 9 As shown, the evidence consistency verification module can correspond to the output item evidence consistency verification module 210. Its underlying form can be a field mapper, rule validator, coverage calculator and review task generator. It maps the output items in the structured processing guidelines to the number and source clause of the knowledge atom, and withdraws the output items or triggers manual review when the mapping fails or conflicts with the prohibition conditions.

[0101] like Figure 1 and Figure 10As shown, this embodiment can also be manifested as an electronic device. This electronic device includes a processor, a memory, a communication interface, and a communication bus. The processor can be a general-purpose CPU, a graphics processor, a dedicated AI acceleration unit, a server-level processing chip, or a multi-core processor; the memory can include volatile memory and non-volatile memory; the communication interface can communicate with the 95598 customer service client 101, the business hall service acceptance client 102, the online business hall client 103, the work order system 104, the customer file system 105, the electricity business ledger 106, and the knowledge base server 108. The processor calls the computer program in the memory through the communication bus to execute steps such as request acquisition, business context parsing, business scenario signature generation, knowledge atom construction, candidate recall, applicability admission determination, evidence set construction, structured processing guidance generation, and output item evidence consistency verification. In case of an anomaly, the processor also calls the anomaly degradation control program to generate supplementary inquiry information, withdrawal markers, evidence summaries, or manual review tasks.

[0102] like Figure 1 and Figure 10 As shown, this embodiment can also be a computer-readable storage medium. This storage medium can be a Flash memory, EEPROM, solid-state drive, hard disk drive, read-only memory, random access memory, optical disk, server storage array, or other medium capable of storing computer programs. When the computer program stored in this storage medium is executed by a processor, it implements the intelligent assistance method for knowledge evidence constraints in power business services as described in any of the foregoing embodiments. This storage medium embodiment can cover implementations delivered in the form of software products, server packages, intranet deployment images, or embedded service programs.

[0103] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A business knowledge-based intelligent assistance method for electricity business service scenarios, characterized by: include, Obtain the electricity business service request and associated business context, and generate a business scenario signature; Obtain knowledge atoms constructed from electricity business service knowledge. These knowledge atoms contain business application boundaries and source information. Candidate knowledge atoms are recalled based on request and business scenario signatures, and applicability admission determination is performed according to the business scenario signature and business applicability boundary to obtain usable knowledge atoms; Based on available knowledge atoms, construct an evidence set and generate structured processing guidelines; Establish an evidence mapping relationship between the output items of the structured processing guidelines and the available knowledge atoms. When the mapping is valid and there is no conflict, determine the release status. When the mapping is missing or a conflict occurs, determine one of the following statuses according to the type of missing item: withdrawal status, supplementary inquiry status, or manual review status.

2. The business knowledge intelligent assistance method according to claim 1, characterized in that: Business scenario signatures include business intent, user category, electricity usage attribute, geographical region, processing channel, processing stage, request time, knowledge version, and material status; Among them, business intent, user category, geographical region, processing stage, request time, and knowledge version are key fields. If any key field is missing, a corresponding supplementary question status is generated. If all key fields are present, the process enters the recall of candidate knowledge atoms.

3. The business knowledge intelligent assistance method according to claim 2, characterized in that: The business application boundaries of knowledge atoms include the set of applicable conditions, the set of prohibited conditions, the applicable geographical scope, the effective time interval, the applicable processing stage, and the type of processing action. The source information includes the source terms and source priority. The applicability admission determination compares each key field with the business applicability boundary, and writes the comparison results into the gating record of the corresponding candidate knowledge atom.

4. The business knowledge intelligent assistance method according to claim 3, characterized in that: The applicability admission determination process sequentially performs condition matching, prohibition condition determination, region matching, version validity determination, and stage matching. When all determinations pass, the corresponding candidate knowledge atom is identified as a usable knowledge atom. If any determination fails, a gating failure code is generated, and the candidate knowledge atom is marked as a removed knowledge atom, preventing it from entering the evidence set.

5. The business knowledge intelligent assistance method according to claim 4, characterized in that: The recall of candidate knowledge atoms includes: keyword recall based on standardized electricity business service requests, vector recall based on semantic similarity, graph adjacency recall based on business intent, processing actions and material item relationships, and metadata filtering based on geographical region, processing stage and knowledge version; Each recall result is deduplicated to form candidate knowledge atoms.

6. The business knowledge intelligent assistance method according to claim 5, characterized in that: When vector recall is in any of the states of response timeout, return exception, or unavailable, candidate knowledge atoms are formed by keyword recall and metadata filtering, and a downgraded recall flag is written into the structured processing guidelines. When vector recall returns normally, its recall results are deduplicated together with the results of keyword recall, graph adjacency recall, and metadata filtering.

7. The business knowledge intelligent assistance method according to claim 6, characterized in that: The evidence set includes the main evidence set, the conditional evidence set, the limiting evidence set, the procedural evidence set, and the conflicting evidence set; When generating structured processing guidelines, the main evidence set forms the processing judgment, the process evidence set forms the material list and processing steps, the limiting evidence set forms the risk warning, and the conflicting evidence set forms the manual review mark.

8. The business knowledge intelligent assistance method according to claim 7, characterized in that: The evidence mapping relationship includes output item identifier, knowledge atom number, and source clause; output items include processing conclusion, material items, time limit items, fee items, and restriction notices; For each output item, search for the corresponding knowledge atom. If a knowledge atom is found and no prohibition condition is met, determine the release status. If no knowledge atom is found, determine the withdrawal status. If the related business context is missing, determine the supplementary question status. If there are conflicts between knowledge atoms, determine the manual review status.

9. The business knowledge intelligent assistance method according to claim 8, characterized in that: After determining one of the following statuses—publishing, withdrawing, supplementary inquiry, or manual review—a log record is generated containing the request number, output item identifier, knowledge atom number, source clause, gating failure code, and status flag. The log record is then written to the log archiving module for the manual review terminal to read the corresponding review task.

10. The business knowledge intelligent assistance method according to claim 9, characterized in that: The structured processing guide includes the business type, confirmed conditions, missing conditions, processing judgment, list of materials, processing steps, risk warning, evidence number, and review mark; When the status is marked as "Published", the output includes the processing guide containing the evidence number; when the status is marked as "Withdrawn", the output includes the withdrawal item; when the status is marked as "Supplementary Question", the output includes the supplementary question field; and when the status is marked as "Manual Review", the output includes the review task number.

11. An intelligent auxiliary system for electricity business service knowledge evidence constraints, characterized in that, This includes knowledge-assisted servers, which include: The request access module is used to receive electricity business service requests; the business context parsing module is used to obtain the business context associated with the electricity business service request. The business scenario signature generation module is used to generate a business scenario signature based on the power business service request and the business context. The knowledge atomization module is used to obtain knowledge atoms that contain information about the business application boundaries and sources. The multi-channel candidate recall module is used to recall candidate knowledge atoms based on power business service requests and business scenario signatures. The applicable condition gating module is used to perform an applicability admission determination based on the signature of the business scenario and the business applicable boundary to obtain usable knowledge atoms; The evidence package construction module is used to construct evidence sets based on available knowledge atoms; The structured processing guidance generation module is used to generate structured processing guidance based on the evidence set; The output item evidence consistency verification module is used to establish an evidence mapping relationship between the output items of the structured processing guidelines and the available knowledge atoms, and determine one of the following statuses based on the evidence mapping relationship: release status, withdrawal status, supplementary inquiry status, and manual review status.

12. An electronic device, characterized in that, It includes a processor, memory, communication interface, and communication bus, with the processor, memory, and communication interface connected via the communication bus; The communication interface is used to receive electricity business service requests, the business context associated with the electricity business service requests, and electricity business service knowledge. The memory contains computer programs; When a computer program is executed by a processor, it causes an electronic device to perform the method as described in claim 1.

13. A computer-readable storage medium, characterized in that, A computer-readable storage medium stores a computer program; when executed by a processor, the computer program causes the processor to perform the method as described in claim 1.

Images