Intelligent customer service robot question and answer method and system based on charging knowledge graph

Abstract

The invention provides an intelligent customer service robot question and answer method and system based on a charging knowledge graph, and the method comprises the steps: carrying out the intention recognition of a charging session inquiry, and obtaining an inquiry intention; based on the inquiry intention, performing entity linking in the charging knowledge graph to obtain an entity to which the inquiry intention is linked; based on the linked entities and inquiry intentions, executing multi-hop search in the charging knowledge graph, and obtaining nodes consistent with intention slot constraint and associated rule terms; generating an evidence chain based on the obtained nodes and the associated rule terms; based on the evidence chain and predefined decision logic, driving the large language model to generate a natural language answer; and combining the natural language answer with the rule term document information corresponding to the natural language answer to generate a charging session answer. According to the inquiry intention, the entity linked in the charging knowledge graph is combined to perform multi-hop search and generate the answer, so that cross-theme interference is avoided, and the accuracy of charging session answering is improved.

Description

Technical Field

[0001] This invention belongs to the field of charging knowledge graph construction technology, specifically relating to an intelligent customer service robot question-and-answer method and system based on charging knowledge graph. Background Technology

[0002] With the popularization of new energy vehicles and public charging services, the platform's user base has exceeded 50 million, and the demand for online customer service continues to grow; therefore, the charging Q&A system for new energy vehicles is particularly important in practical applications.

[0003] When existing charging question-answering systems lack the constraints of domain structure, the retrieval stage relies solely on similarity as a guide, which easily leads to cross-topic interference and recall drift. In the generation stage, seemingly smooth but unfounded conclusions may be "rationalized" in the absence of sufficient evidence. Summary of the Invention

[0004] To overcome the shortcomings of the prior art, in a first aspect, this invention proposes an intelligent customer service robot question-and-answer method based on a charging knowledge graph, comprising: An intelligent customer service robot is used to receive charging conversation inquiries input by users; the intent of the charging conversation inquiries is identified to obtain the inquiry intent; Based on the query intent, entity links are performed in the pre-built charging knowledge graph to obtain the entities linked to the query intent; based on the linked entities and the query intent, a multi-hop search is performed in the charging knowledge graph to obtain nodes consistent with the intent slot constraints and associated rule clauses; based on the obtained nodes and associated rule clauses, an evidence chain is generated. Based on the evidence chain and predefined decision logic, the large language model is driven to generate a natural language answer; the natural language answer and the corresponding rule clause document information are combined to generate a charging session answer, which is then fed back to the user through the intelligent customer service robot.

[0005] Preferably, the process of constructing the charging knowledge graph includes: From multiple data documents including FAQs, business procedures, billing rules, site information, and order information for the charging business, information extraction technology is used to obtain entities, relationships, entity attributes, and rule terms. The entity is defined as a node in the knowledge graph, the relationship is defined as an edge in the knowledge graph, the attribute is defined as a feature of the node, and the rule clause is associated and bound to the corresponding node. Determine the version of the rule clauses and add corresponding version identifiers to the rule clauses to form a charging knowledge graph.

[0006] Preferably, the entities include: sites, charging piles, connectors, billing, orders, payments, invoices, members, service packages, channels, and rules; the relationships include at least one of the following: Based on the ownership relationship between the site and the charging pile, a "site-owner-equipment" relationship is formed; The relationship between billing rules and the site or order is formed as "billing-applicable-site / order"; The relationship between members and rules is formed as a "member-constraint-rule" relationship. Based on the usage relationship between invoices and orders, a "invoice-use-order" relationship is formed; Based on the relationship between FAQs and rules or processes, a "FAQ-Answer-Rules or Processes" relationship is formed. Based on the support relationship between channels and payment, a "channel-support-payment" relationship is formed.

[0007] Preferably, the step of performing a multi-hop search in the charging knowledge graph based on the linked entity and the query intent to obtain nodes and associated rule clauses consistent with the intent slot constraints; and generating an evidence chain based on the obtained nodes and associated rule clauses, includes: starting from the linked entity, performing a multi-hop search in the corresponding subgraph of the charging knowledge graph using a graph traversal algorithm to obtain nodes consistent with the intent slot constraints; extracting associated rule clauses from the obtained nodes; vectorizing the text content of the associated rule clauses and performing semantic retrieval to filter out candidate rule clause fragments semantically similar to the query intent; and integrating the obtained nodes, the candidate rule clause fragments, the associated rule clause version information, and the corresponding multi-source data documents to generate an evidence chain.

[0008] Preferably, the step of using the linked entity as the starting point and employing a graph traversal algorithm to perform a multi-hop search in the corresponding subgraph of the charging knowledge graph to obtain nodes consistent with the intent slot constraint includes: Starting from the linked entity, a graph traversal algorithm is used to perform a multi-hop search with a limited depth in the corresponding subgraph of the charging knowledge graph. During each hop of the search, according to the intent slot constraint in the query intent, only the edges that are consistent with the intent slot constraint are traversed, and the nodes that are consistent with the intent slot constraint are selected.

[0009] Preferably, the step of integrating the acquired nodes, the candidate rule clause fragments, the associated rule clause version information, and the corresponding multi-source data documents to generate a chain of evidence includes: The acquired nodes, candidate rule clause fragments, associated rule clause version information, and corresponding multi-source data documents are packaged to form an initial evidence set. The initial evidence set is then subjected to consistency verification based on the charging session query, and evidence fragments that conflict with session constraints or have expired are removed. The evidence fragments retained after verification are reordered according to their relevance to the query intent to form the final evidence chain.

[0010] Preferably, the process of generating the charging session answer includes at least one of the following: When, based on the evidence chain, it is identified that a charging session query is ambiguous due to the lack of a key intent slot constraint, a charging session answer containing common scenarios is generated based on the natural language answer and the default rule clause corresponding to the natural language answer. A clarifying question regarding the key intent slot constraint is embedded in the charging session answer, along with the multi-source data document corresponding to the charging session answer. When the evidence chain indicates the existence of multiple effective versions of the rule clause, the latest effective rule clause version corresponding to the natural language answer is preferentially selected to generate the charging session answer, and the version of the rule clause on which it is based is marked in the charging session answer, along with the multi-source data document corresponding to the charging session answer. If the latest effective rule clause version differs in applicability from other historical rule clause versions, the difference is also marked in the charging session answer. The intent slot constraint and the key intent slot constraint both include at least one of the following: the channel used by the user to initiate the charging session inquiry, the user's location, or the time information involved in the inquiry.

[0011] Preferably, after the charging knowledge graph is constructed, the method further includes: The rule clauses in the charging knowledge graph are updated and conflict detected through an incremental update pipeline; In response to a successful update, the version number of the rule terms document is incremented, and the retrieval index associated with the charging knowledge graph is updated synchronously. Construct a thesaurus covering common term variations in entity and rule clause document information in the charging knowledge graph; The thesaurus is applied to entity linking, the charging knowledge graph retrieval, and the charging conversation answer generation process; When an erroneous update is detected in the charging knowledge graph, a version rollback operation is performed to restore the charging knowledge graph to its state before the update.

[0012] Preferably, the node includes at least one of a site, charging pile, order, payment tool, invoice, member, and service package; The rules and terms include at least one of the following: billing rules, membership benefits rules, refund rules, and invoicing rules; The multi-source data documents include: FAQs on charging services, operating procedures and processes for charging services, charging billing rules and pricing policies, equipment information and status documents for charging stations and charging piles, and user order and transaction record documents.

[0013] Secondly, this invention application also proposes an intelligent customer service robot question-and-answer system based on a charging knowledge graph, comprising: The system application layer is used to receive charging session inquiries input by users using an intelligent customer service robot; and to perform intent recognition on the charging session inquiries to obtain the inquiry intent. The functional component layer is used to perform entity linking in the pre-built charging knowledge graph based on the query intent to obtain the entity linked to by the query intent; based on the linked entity and the query intent, perform multi-hop search in the charging knowledge graph to obtain the node consistent with the intent slot constraint and the associated rule clause; and generate an evidence chain based on the obtained node and associated rule clause. The basic support layer is used to drive the large language model to generate natural language answers based on the evidence chain and predefined decision logic; combine the natural language answers with the rule clause document information corresponding to the natural language answers to generate charging session answers, and then feed them back to the user through the intelligent customer service robot.

[0014] Preferably, the system further includes: a model building layer, used for: From multiple data documents including FAQs, business procedures, billing rules, site information, and order information for the charging business, information extraction technology is used to obtain entities, relationships, entity attributes, and rule terms. The entity is defined as a node in the knowledge graph, the relationship is defined as an edge in the knowledge graph, the attribute is defined as a feature of the node, and the rule clause is associated and bound to the corresponding node. Determine the version of the rule clauses and add corresponding version identifiers to the rule clauses to form a charging knowledge graph.

[0015] Preferably, the entities include: sites, charging piles, connectors, billing, orders, payments, invoices, members, service packages, channels, and rules; the relationships include at least one of the following: a "site-owner-device" relationship based on the ownership relationship between sites and charging piles; a "billing-applies to-site / order" relationship based on the applicability relationship between billing rules and sites or orders; a "member-constraint-rule" relationship based on the constraint relationship between members and rules; a "invoice-used for-order" relationship based on the usage relationship between invoices and orders; a "FAQ-answer-rule or process" relationship based on the answer relationship between FAQs and rules or processes; and a "channel-support-payment" relationship based on the support relationship between channels and payments.

[0016] Preferably, the functional component layer includes: a node acquisition module, used to perform multi-hop search in the corresponding subgraph of the charging knowledge graph, starting from the linked entity, using a graph traversal algorithm to acquire nodes consistent with the intent slot constraint; a rule clause extraction module, used to extract the associated rule clauses from the acquired nodes; a rule clause fragment filtering module, used to vectorize the text content of the associated rule clauses and perform semantic retrieval to filter out candidate rule clause fragments that are semantically similar to the query intent; and an evidence chain generation module, used to integrate the acquired nodes, the candidate rule clause fragments, the associated rule clause version information, and the corresponding multi-source data documents to generate an evidence chain.

[0017] Preferably, the node acquisition module is specifically used for: Starting from the linked entity, a graph traversal algorithm is used to perform a multi-hop search with a limited depth in the corresponding subgraph of the charging knowledge graph. During each hop of the search, according to the intent slot constraint in the query intent, only the edges that are consistent with the intent slot constraint are traversed, and the nodes that are consistent with the intent slot constraint are selected.

[0018] Preferably, the evidence chain generation module is specifically used for: The acquired nodes, candidate rule clause fragments, associated rule clause version information, and corresponding multi-source data documents are packaged to form an initial evidence set. The initial evidence set is then subjected to consistency verification based on the charging session query, and evidence fragments that conflict with session constraints or have expired are removed. The evidence fragments retained after verification are reordered according to their relevance to the query intent to form the final evidence chain.

[0019] Preferably, the process of generating charging session answers in the basic support layer specifically includes at least one of the following: When, based on the evidence chain, it is identified that a charging session query is ambiguous due to the lack of a key intent slot constraint, a charging session answer containing common scenarios is generated based on the natural language answer and the default rule clause corresponding to the natural language answer. A clarifying question regarding the key intent slot constraint is embedded in the charging session answer, along with the multi-source data document corresponding to the charging session answer. When the evidence chain indicates the existence of multiple effective versions of the rule clause, the latest effective rule clause version corresponding to the natural language answer is preferentially selected to generate the charging session answer, and the version of the rule clause on which it is based is marked in the charging session answer, along with the multi-source data document corresponding to the charging session answer. If the latest effective rule clause version differs in applicability from other historical rule clause versions, the difference is also marked in the charging session answer. The intent slot constraint and the key intent slot constraint both include at least one of the following: the channel used by the user to initiate the charging session inquiry, the user's location, or the time information involved in the inquiry.

[0020] Preferably, the system further includes: a knowledge graph update module, used for: The rule clauses in the charging knowledge graph are updated and conflict detected through an incremental update pipeline; In response to a successful update, the version number of the rule terms document is incremented, and the retrieval index associated with the charging knowledge graph is updated synchronously. Construct a thesaurus covering common term variations in entity and rule clause document information in the charging knowledge graph; The thesaurus is applied to entity linking, the charging knowledge graph retrieval, and the charging conversation answer generation process; When an erroneous update is detected in the charging knowledge graph, a version rollback operation is performed to restore the charging knowledge graph to its state before the update.

[0021] Preferably, the node includes at least one of a site, charging pile, order, payment tool, invoice, member, and service package; The rules and terms include at least one of the following: billing rules, membership benefits rules, refund rules, and invoicing rules; The multi-source data documents include: FAQs on charging services, operating procedures and processes for charging services, charging billing rules and pricing policies, equipment information and status documents for charging stations and charging piles, and user order and transaction record documents.

[0022] Thirdly, this application also proposes an electronic device, comprising: at least one processor and a memory; wherein the memory and the processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, the intelligent customer service robot question-and-answer method based on charging knowledge graph is implemented.

[0023] Fourthly, this application also proposes a readable storage medium having an executable program stored thereon, wherein when the executable program is executed, it implements the aforementioned intelligent customer service robot question-and-answer method based on a charging knowledge graph.

[0024] Compared with the closest prior art, the present invention application has the following beneficial effects: This invention discloses an intelligent customer service robot question-answering method and system based on a charging knowledge graph, comprising: using an intelligent customer service robot to receive a charging conversation inquiry input by a user; performing intent recognition on the charging conversation inquiry to obtain the inquiry intent; based on the inquiry intent, performing entity linking in a pre-constructed charging knowledge graph to obtain the entities linked to the inquiry intent; based on the linked entities and the inquiry intent, performing a multi-hop search in the charging knowledge graph to obtain nodes consistent with the intent slot constraints and associated rule clauses; generating an evidence chain based on the obtained nodes and associated rule clauses; driving a large language model to generate a natural language answer based on the evidence chain and predefined decision logic; combining the natural language answer and the rule clause document information corresponding to the natural language answer to generate a charging conversation answer, and feeding it back to the user through the intelligent customer service robot. By performing a multi-hop search according to the inquiry intent, combining the entities linked in the charging knowledge graph, and generating an evidence chain to further generate the answer, cross-topic interference and recall drift are avoided, significantly improving the accuracy of charging conversation responses. Attached Figure Description

[0025] Figure 1 A flowchart of a question-and-answer method for an intelligent customer service robot based on a charging knowledge graph, provided for this invention application; Figure 2 A flowchart illustrating an intelligent customer service robot question-and-answer method based on a charging knowledge graph, provided in this invention application; Figure 3 An architecture diagram of an intelligent customer service robot question-and-answer system based on a charging knowledge graph provided for this invention application; Figure 4 A detailed architecture diagram of an intelligent customer service robot question-and-answer system based on a charging knowledge graph provided in this invention application; Figure 5This is a schematic diagram of the operation of an electronic device provided in this invention application. Detailed Implementation

[0026] The specific embodiments of this invention will be further described in detail below with reference to the accompanying drawings.

[0027] Example 1: like Figure 1 As shown, this invention application proposes a question-and-answer method for an intelligent customer service robot based on a charging knowledge graph, comprising the following steps: Step 1: Use an intelligent customer service robot to receive charging conversation inquiries from users; perform intent recognition on the charging conversation inquiries to obtain the inquiry intent; Step 2: Based on the query intent, perform entity linking in the pre-built charging knowledge graph to obtain the entities linked to the query intent; based on the linked entities and the query intent, perform a multi-hop search in the charging knowledge graph to obtain nodes consistent with the intent slot constraints and associated rule clauses; based on the obtained nodes and associated rule clauses, generate an evidence chain. Step 3: Based on the evidence chain and predefined decision logic, drive the large language model to generate a natural language answer; combine the natural language answer with the rule clause document information corresponding to the natural language answer to generate a charging session answer, and then feed it back to the user through the intelligent customer service robot.

[0028] The method described in steps 1-3 constructs an end-to-end intelligent question-answering process from user inquiry to credible answer. First, by linking the query intent with entities, the user's natural language charging inquiry is precisely anchored to a specific entity within a pre-built charging knowledge graph. Then, based on this entity and query intent, a multi-hop search is performed in the knowledge graph to locate relevant nodes and rule clauses that satisfy the constraints. A scenario-based charging business (Schema) is used to limit the search space and multi-hop fact location, generating a traceable structured evidence chain to ensure the transparency and verifiability of the reasoning process. Finally, the evidence chain drives a large language model to generate a natural language answer, which is then combined with the source information of the rule clauses. This source information, such as node IDs and rule clause versions, is used by an intelligent customer service robot to output an accurate, easy-to-understand, and auditable final response to the user. The entire process is interconnected, integrating the precise reasoning of the knowledge graph with the natural generation advantages of the large language model, achieving high reliability, strong interpretability, and continuous maintainability in the charging domain's question-answering mechanism.

[0029] In multi-turn dialogues, the subject entity, intent slot constraints, and referential links serve as constraints throughout the retrieval and generation process, ensuring contextual consistency across consecutive sessions. Evidence is provided through intent slot constraints, and the source is marked with node ID, rule clause version, and timestamp to achieve auditability and maintainability.

[0030] Furthermore, the node includes at least one of a site, charging pile, order, payment tool, invoice, member, and service package; the rules and terms include at least one of billing rules, membership rights rules, refund rules, and invoicing rules.

[0031] Furthermore, the process of constructing the charging knowledge graph in step 2 includes: Step A: From the multi-source data documents of charging business FAQs (Frequently Asked Questions), business procedures, billing rules, site information, and order information, information extraction technology is used to obtain entities, relationships, entity attributes, and rule terms; Step B: Define the entity as a node in the knowledge graph, the relationship as an edge in the knowledge graph, the attribute as a feature of the node, and associate and bind the rule clause with the corresponding node; Step C: Determine the version of the rule clause and add a corresponding version identifier to the rule clause to form a charging knowledge graph.

[0032] Further, in step 2 above, the entities include: sites, charging piles, connectors, billing, orders, payments, invoices, members, service packages, channels, and rules; the relationships include at least one of the following: a "site-owner-device" relationship based on the ownership relationship between sites and charging piles; a "billing-applies to-site / order" relationship based on the applicability relationship between billing rules and sites or orders; a "member-constraint-rule" relationship based on the constraint relationship between members and rules; a "invoice-used-order" relationship based on the usage relationship between invoices and orders; a "FAQ-answer-rule or process" relationship based on the answer relationship between FAQs and rules or processes; and a "channel-support-payment" relationship based on the support relationship between channels and payments.

[0033] In step 2 above, regarding domain modeling of the knowledge graph, subgraphs are organized around the user journey of "charging station search—charging—payment—after-sales service (invoicing / refund)," with stations, charging stations, connectors, billing, orders, payments, invoices, members (e.g., membership, PLUS paid membership), service packages, channels, and rules as core entities. These are linked through relationships such as "station—ownership—device," "billing—applicable to—station / order," "membership—constraint—rule," "FAQ—answer—rule or process," "invoice—used for—order," and "channel—support—payment," forming a multi-layered knowledge structure from business scope to operational path. Unlike traditional document-only retrieval, this charging knowledge graph emphasizes the scope of application, effective / expiratory time, regional differences, and exceptions of clauses, enabling rapid limitation of the retrieval space and output of verifiable clause fragments during question-and-answer sessions. This knowledge graph, constructed for the first time in the charging business in a scenario-based manner, provides a consistent fact engine for question-and-answer and operational decision-making, laying the knowledge foundation for the navigation and generation of this invention.

[0034] In terms of the aforementioned charging knowledge graph modeling, this invention organizes subgraphs around the user journey of "charging station search—charging—payment—after-sales service (invoicing / refund)," using stations, charging piles, connectors, billing, orders, payments, invoices, members (including PLUS), service packages, channels, and rules as core entities. These are linked through relationships such as "station-owner-charging pile," "billing-applicable-station / order," "member-constraint-rule," "FAQ-answer-rule or process," "invoice-used-order," and "channel-support-payment," forming a multi-layered knowledge structure from business scope to operational path. Unlike traditional document-only retrieval, this charging knowledge graph emphasizes the scope of application, effective / expiratory time, regional differences, and exceptions of clauses, enabling rapid limitation of the retrieval space and output of verifiable clause fragments during question-and-answer sessions. This scenario-based charging knowledge graph is the first to be constructed within the charging business, providing a consistent fact engine for question-and-answer and operational decision-making, laying the knowledge foundation for the retrieval and generation of this invention.

[0035] Furthermore, after the charging knowledge graph is constructed, the following is also included: Step S1: Update and conflict-detect the rule clauses in the charging knowledge graph through the incremental update pipeline; Step S2: In response to a successful update, increment the version number of the rule terms document and synchronously update the retrieval index associated with the charging knowledge graph; Step S3: Construct a thesaurus covering common term variations in entity and rule clause document information in the charging knowledge graph; Step S4: Apply the thesaurus to entity linking, the charging knowledge graph retrieval, and the charging conversation answer generation process; Step S5: When an erroneous update is detected in the charging knowledge graph, a version rollback operation is performed to restore the charging knowledge graph to its state before the update.

[0036] After the charging knowledge graph is built, the following maintenance and enhancement processes are further executed: Rule clauses in the knowledge graph are dynamically updated and automatically conflict detected through an incremental update pipeline; upon successful update, the version number of the rule clause is automatically incremented and the relevant search index is updated synchronously; simultaneously, a thesaurus covering common term variations such as "PLUS member / Plus member / member Plus" is built and applied to the entire process of entity linking, knowledge retrieval, and conversation answer generation to ensure consistency in terminology; if an erroneous update is detected, version rollback can be automatically performed to restore the correct state before the update. Through the above terminology synonym unification mechanism and versioned incremental update process, stable, consistent, and highly available output can be maintained even when business rules and definitions change.

[0037] Furthermore, the multi-source data documents include: frequently asked questions and answers documents for charging services, operating procedures and processes documents for charging services, charging billing rules and pricing policies documents, equipment information and status documents for charging stations and charging piles, and user order and transaction record documents.

[0038] In step 2 above, the step of performing a multi-hop search in the charging knowledge graph based on the linked entity and the query intent to obtain nodes and associated rule clauses consistent with the intent slot constraints, and generating an evidence chain based on the obtained nodes and associated rule clauses, includes the following steps: Step 2.1: Starting from the linked entity, a graph traversal algorithm is used to perform a multi-hop search in the corresponding subgraph of the charging knowledge graph to obtain nodes consistent with the intent slot constraints; Step 2.2: From the obtained nodes, the associated rule clauses are extracted; Step 2.3: The text content of the associated rule clauses is vectorized and semantic retrieval is performed to filter out candidate rule clause fragments that are semantically similar to the query intent; Step 2.4: Based on the obtained nodes, the candidate rule clause fragments, the associated rule clause version information, and the corresponding multi-source data documents, an evidence chain is generated.

[0039] In terms of evidence chain generation, this invention does not simply splice together several similar segments. Instead, it starts with the intent and anchor points in the session, performs a multi-hop search along the subgraph with a limited depth, prioritizes edges and nodes consistent with the current slot constraints (such as channel, region, or time), and packages the matched node attribute slices and rule clauses into evidence. The evidence includes node ID, rule version, timestamp, and source document pointer, which can be directly mapped to the knowledge base to support verification and tracing.

[0040] In step 2.1 above, the step of using the linked entity as the starting point and employing a graph traversal algorithm to perform a multi-hop search in the corresponding subgraph of the charging knowledge graph to obtain nodes consistent with the intent slot constraint includes: Step 2.1.1: Starting from the linked entity, perform a multi-hop search with a limited depth in the corresponding subgraph of the charging knowledge graph using a graph traversal algorithm; Step 2.1.2: During each hop of the search process, according to the intent slot constraint in the query intent, only traverse the edges that are consistent with the intent slot constraint, and select the nodes that are consistent with the intent slot constraint.

[0041] In step 2.4 above, the process of integrating the acquired nodes, candidate rule clause fragments, associated rule clause version information, and corresponding multi-source data documents to generate a chain of evidence includes: Step 2.4.1: Package the acquired nodes, candidate rule clause fragments, associated rule clause version information, and corresponding multi-source data documents to form an initial evidence set; Step 2.4.2: Perform consistency verification on the initial evidence set based on the charging session query, and remove evidence fragments that conflict with session constraints or have expired; Step 2.4.3: Reorder the retained evidence fragments according to their relevance to the query intent to form the final evidence chain.

[0042] Steps 2.4.1 to 2.4.3 above constitute the complete process of evidence chain generation: First, the retrieved nodes, rule clause fragments, and version information are packaged to form an initial evidence set; then, consistency verification is performed based on the current session query, eliminating conflicting or expired evidence fragments; finally, the evidence is rearranged according to its relevance to the query intent, forming a structured and traceable evidence chain. This process realizes the transformation from raw search results to high-quality evidence, providing a reliable basis for subsequently generating auditable and accurate answers.

[0043] In step 3 above, the process of generating the charging session answer includes at least one of the following: When, based on the evidence chain, it is identified that a charging session query is ambiguous due to the lack of a key intent slot constraint, a charging session answer containing common scenarios is generated based on the natural language answer and the default rule clause corresponding to the natural language answer. A clarifying question regarding the key intent slot constraint is embedded in the charging session answer, along with the multi-source data document corresponding to the charging session answer. When the evidence chain indicates the existence of multiple effective versions of the rule clause, the latest effective rule clause version corresponding to the natural language answer is preferentially selected to generate the charging session answer, and the version of the rule clause on which it is based is marked in the charging session answer, along with the multi-source data document corresponding to the charging session answer. If the latest effective rule clause version differs in applicability from other historical rule clause versions, the difference is also marked in the charging session answer. The intent slot constraint and the key intent slot constraint both include at least one of the following: the channel used by the user to initiate the charging session inquiry, the user's location, or the time information involved in the inquiry.

[0044] In step 3 above, the generation of answers for the charging session mainly follows two types of rules: First, when the evidence chain identifies that the query is ambiguous due to the lack of key intent slot constraints (such as channel, region, time), an answer covering the most common situations is generated based on the default rule clauses, and the answer embeds a clarifying question about the missing constraints and the corresponding data document reference; Second, when the evidence chain indicates that there are multiple effective versions of the rule clauses, the latest effective version is used to generate the answer and the version information is clearly marked. If there are differences in applicability between the old and new versions, they are marked together. All answers embed the corresponding multi-source data document source to ensure traceability.

[0045] In terms of multi-turn dialogue and session state management, topic entities, slot constraints, historical evidence, and version definitions are continuously maintained state units. A hybrid retrieval strategy integrates graph traversal algorithms (such as breadth-first search, BFS, or heuristic A* search) with vector recall techniques (such as k-nearest neighbors, kNN) within a unified reordering framework, and incorporates session constraints for consistent evidence screening. Session management continuously records topics and slots to handle cross-turn binding of pronouns, omissions, and references. When a user continues a question with an abbreviated phrase like "How do I get a refund?", the topic (e.g., PLUS membership) and intent (e.g., refund) can be recovered from the previous dialogue, guiding subsequent navigation and evidence extraction, thus leading the dialogue to a decidable knowledge branch. If the current question is ambiguous or lacks key information (e.g., channel or region), the most common scenario can be presented based on default rules, and clarification can be proactively requested in the answer, improving the one-time resolution rate and avoiding over-inference when evidence is insufficient. In cases where both old and new rules coexist, the latest effective version will be output first, and the effective date and major differences will be indicated in the response to ensure consistency and transparency in external explanations.

[0046] In summary, such as Figure 2 As shown, this invention adopts the overall framework of "Knowledge-Graph-guided Retrieval-Augmented Generation (KG-RAG)" to achieve verifiable and traceable intelligent question answering. This framework first extracts multi-source information such as FAQs, business procedures, billing and membership, invoices and billing, sites and equipment, orders and payments into entities, relationships, and attributes on the data side, forming a charging domain schema and charging knowledge graph with versioning and applicability labeling. First, user questions can be encoded using an embedding model and semantically matched with vector indexes. After successful matching, a reranker model can be used to re-rank the results and select the Top-K relevant fragments. Subsequently, on the indexing side, both graph indexes and vector indexes are maintained simultaneously, enabling structured navigation and semantic nearest neighbor retrieval to work together. On the conversation side, state variables such as topic entities, slot constraints, and referential links are continuously maintained, serving as hard constraints for retrieval and generation. On the navigation side, multi-hop searches with limited depth are performed based on the graph, filtering out nodes and terms consistent with the conversation constraints and packaging them into evidence chains with node IDs, versions, and timestamps. On the generation side, a large language model is driven by strong constraint templates to answer based solely on evidence, and source annotations are automatically inserted. If the evidence is insufficient or conflicting, clarification or rejection strategies are triggered, thereby achieving verifiable and traceable generation in the end-to-end question-and-answer process.

[0047] To facilitate understanding, two typical examples are provided. First, the "PLUS Membership Refund Entry" question is a high-frequency, multi-phrase type of question, with users often asking phrases like "How do I get a refund for my Plus membership?", "Where is the refund entry?", and "PLUS benefits refund process?". The system first anchors "Membership—Refund" during the intent recognition and entity linking stage. Then, it extracts clauses consistent with the current session constraints along the "Membership → Refund Rules → Channel Restrictions" link, explicitly providing the APP (Application, mini-program) or mini-program path and restrictions in the answer, along with source attribution. If channel information is missing, a default path is provided first, followed by clarification. Project materials show that this type of question has been used in training examples for "similar question focusing and preference alignment" to improve the ability to capture true semantics. Second, in the "Vehicle Battery Service Package Purchase Path" question, after first identifying the "Service Package—Purchase" topic, it directly extracts evidence from the "Service Package → Purchase Entry → Account Binding" link, generating an answer consistent with the knowledge base's preset criteria. Both of these examples are presented in the prototype demonstration.

[0048] In addition, the existing technology also has the following drawbacks: Currently, online channels still rely primarily on human customer service, which suffers from pain points such as untimely response, mediocre user experience, and high operating costs. Furthermore, existing FAQ and fuzzy search capabilities are limited, making it difficult to support natural language consultations with rich semantic variations, and lacking the ability to continuously understand context and handle multi-turn dialogues. While general RAG (Retrieval-Augmented Generation) solutions that simply rely on vector similarity are easy to deploy, they often suffer from recall drift and model illusions due to the lack of constraints from domain-structured knowledge. Moreover, answers are difficult to trace back to a stable source, failing to meet the accuracy, verifiability, and consistency requirements of the power charging business. 2. In existing practices, customer service robots based on FAQ retrieval or pure vector semantic retrieval have a low deployment threshold, but on the one hand, they struggle to handle real-world conversation phenomena such as paraphrasing, colloquial expressions, omissions, and referential inconsistencies; on the other hand, they lack structured constraints on business terms (such as billing, membership, and invoicing rules) and relationships between objects (site—device—order—payment), leading to unstable answers, granular inconsistencies, or even fabricated facts. In contrast, knowledge graph question answering has strong structured constraints and traceability, but it is difficult to cover the business architecture and differences in the scope of the charging field on general knowledge graphs, resulting in high migration costs and limited adaptability.

[0049] 3. When a system lacks domain structure constraints, the retrieval phase relies solely on similarity, making it highly susceptible to cross-topic interference and recall drift. The generation phase, lacking sufficient evidence, may "rationalize" seemingly smooth but unfounded conclusions, representing the most sensitive "illusion" risk in customer service. Furthermore, without continuous management of session state, the system cannot reliably handle references and omissions like "this peg" or "that invoice" in multi-turn conversations, leading to contextual breaks and repeated follow-up questions, significantly degrading the user experience. Moreover, frequent changes in business rules, without versioning and synonym unification mechanisms in the knowledge base, can result in contradictory answers when old and new interpretations coexist, disrupting consistency in customer service messaging. These deficiencies are particularly pronounced in high-business scenarios like electric vehicle charging, thus requiring a question-answering mechanism centered on graph structure constraints and evidence chain presentation to fundamentally mitigate these issues.

[0050] The method of this invention introduces a structured representation of entities, relationships, and rules from a knowledge graph into the entire retrieval and generation process. This enables the system to reliably identify intents and anchor points in multi-turn dialogues, perform multi-hop localization along the business journey subgraph, package hit nodes and clauses as evidence, and explicitly cite sources when generating answers. If insufficient evidence is encountered, the system proactively initiates clarification or refuses to answer. Through this method, the system can significantly improve the accuracy, consistency, and auditability of answers while maintaining a natural language interaction experience. This meets the practical needs of e-charging apps and vehicle networking platforms for online customer service to move from simply being able to answer to being truly trustworthy, aligning with the goals of constructing a scenario-based knowledge graph, applying RAG to implement a knowledge navigator, and realizing a question-and-answer prototype.

[0051] Compared to general RAG solutions that rely solely on vector recall, this invention introduces graph structure constraints and hard session state constraints during the retrieval phase, and uses evidence chains to drive output with source annotations during the generation phase, significantly improving the accuracy and consistency of answers. Prototype demonstrations show that the system can provide stable answers based on preset knowledge base definitions and effectively alleviate model illusions through RAG navigation. Furthermore, due to support for versioning and synonym unification, the system maintains consistent and traceable external responses even during periods of business definition changes. The handling of pronouns, omissions, and ambiguities in multi-turn dialogues is also improved through state management, resulting in an overall enhanced user experience and meeting the practical requirements of e-charging and vehicle networking platforms for "accurate, verifiable, and evolvable" online customer service.

[0052] This invention aims to deeply couple a knowledge graph in the charging domain as a "knowledge navigator" with a large language model, providing structured constraints and evidence chains. This enables robust semantic understanding, accurate knowledge retrieval, and reliable natural language generation in multi-turn dialogues, while retaining source annotations in the output to ensure auditability and maintainability. The method, as presented in mid-term and interim materials, proposes and validates the overall approach of "charging problem knowledge graph + RAG knowledge navigator + multi-turn question-answering prototype": on the one hand, constructing a scenario-based knowledge graph guided by the user journey; on the other hand, tightly binding structured evidence with the large model's generation process through the knowledge navigator at the generation end. This allows for accurate answers based on the knowledge base in the prototype and significantly alleviates the model illusion problem.

[0053] Example 2: like Figure 3 As shown, the present invention also provides an intelligent customer service robot question-and-answer system based on a charging knowledge graph, including: The system application layer is used to receive charging session inquiries input by users using an intelligent customer service robot; and to perform intent recognition on the charging session inquiries to obtain the inquiry intent. The functional component layer is used to perform entity linking in the pre-built charging knowledge graph based on the query intent to obtain the entity linked to by the query intent; based on the linked entity and the query intent, perform multi-hop search in the charging knowledge graph to obtain the node consistent with the intent slot constraint and the associated rule clause; and generate an evidence chain based on the obtained node and associated rule clause. The basic support layer is used to drive the large language model to generate natural language answers based on the evidence chain and predefined decision logic; combine the natural language answers with the rule clause document information corresponding to the natural language answers to generate charging session answers, and then feed them back to the user through the intelligent customer service robot.

[0054] Furthermore, the system also includes: a model building layer, used for: From multiple data documents including FAQs, business procedures, billing rules, site information, and order information for the charging business, information extraction technology is used to obtain entities, relationships, entity attributes, and rule terms. The entity is defined as a node in the knowledge graph, the relationship is defined as an edge in the knowledge graph, the attribute is defined as a feature of the node, and the rule clause is associated and bound to the corresponding node. Determine the version of the rule clauses and add corresponding version identifiers to the rule clauses to form a charging knowledge graph.

[0055] like Figure 4As shown, in terms of system architecture, this invention preferably adopts a four-layer structure to organize the implementation logic. The basic support layer can provide a graph database or relational database, object storage, vector retrieval engine, and index synchronization components to ensure unified management and high-concurrency access of structured data and document evidence. The model building layer includes graph construction, hybrid retrieval, entity recognition, relation extraction, attribute standardization, and conflict resolution, and performs version management of rule clauses; hybrid retrieval unifies graph traversal (such as BFS or heuristic A*) and vector retrieval (such as kNN) under the same reordering framework, and performs consistency screening based on comprehensive session constraints; session management records topics and slots, and handles cross-round binding of pronouns, omissions, and references; natural language output is performed under the constraints of the evidence chain, and source annotations and effective information are automatically inserted. Finally, the system unfolds along the main line of "knowledge question answering - multi-round clarification - evidence presentation - knowledge operation," forming a closed loop between question answering and knowledge maintenance. The technical organization of "overall architecture design - key technology research - prototype development" proposed by this system is consistent and has been verified in the prototype. A complete "retrieval-generation" closed-loop system is formed with the charging knowledge graph as the core of navigation.

[0056] Further, the entities include: sites, charging piles, connectors, billing, orders, payments, invoices, members, service packages, channels, and rules; the relationships include at least one of the following: a "site-owner-device" relationship based on the ownership relationship between sites and charging piles; a "billing-applies to-site / order" relationship based on the applicability relationship between billing rules and sites or orders; a "member-constraint-rule" relationship based on the constraint relationship between members and rules; a "invoice-used-order" relationship based on the usage relationship between invoices and orders; a "FAQ-answer-rule or process" relationship based on the answer relationship between FAQs and rules or processes; and a "channel-support-payment" relationship based on the support relationship between channels and payments.

[0057] Furthermore, the functional component layer includes: a node acquisition module, used to perform multi-hop search in the corresponding subgraph of the charging knowledge graph, starting from the linked entity, using a graph traversal algorithm to acquire nodes consistent with the intent slot constraint; a rule clause extraction module, used to extract the associated rule clauses from the acquired nodes; a rule clause fragment filtering module, used to vectorize the text content of the associated rule clauses and perform semantic retrieval to filter out candidate rule clause fragments that are semantically similar to the query intent; and an evidence chain generation module, used to integrate the acquired nodes, the candidate rule clause fragments, the associated rule clause version information, and the corresponding multi-source data documents to generate an evidence chain.

[0058] Furthermore, the node acquisition module is specifically used for: Starting from the linked entity, a graph traversal algorithm is used to perform a multi-hop search with a limited depth in the corresponding subgraph of the charging knowledge graph. During each hop of the search, according to the intent slot constraint in the query intent, only the edges that are consistent with the intent slot constraint are traversed, and the nodes that are consistent with the intent slot constraint are selected.

[0059] Furthermore, the evidence chain generation module is specifically used for: The acquired nodes, candidate rule clause fragments, associated rule clause version information, and corresponding multi-source data documents are packaged to form an initial evidence set. The initial evidence set is then subjected to consistency verification based on the charging session query, and evidence fragments that conflict with session constraints or have expired are removed. The evidence fragments retained after verification are reordered according to their relevance to the query intent to form the final evidence chain.

[0060] Furthermore, the process of generating charging session answers in the basic support layer specifically includes at least one of the following: When, based on the evidence chain, it is identified that a charging session query is ambiguous due to the lack of a key intent slot constraint, a charging session answer containing common scenarios is generated based on the natural language answer and the default rule clause corresponding to the natural language answer. A clarifying question regarding the key intent slot constraint is embedded in the charging session answer, along with the multi-source data document corresponding to the charging session answer. When the evidence chain indicates the existence of multiple effective versions of the rule clause, the latest effective rule clause version corresponding to the natural language answer is preferentially selected to generate the charging session answer, and the version of the rule clause on which it is based is marked in the charging session answer, along with the multi-source data document corresponding to the charging session answer. If the latest effective rule clause version differs in applicability from other historical rule clause versions, the difference is also marked in the charging session answer. The intent slot constraint and the key intent slot constraint both include at least one of the following: the channel used by the user to initiate the charging session inquiry, the user's location, or the time information involved in the inquiry.

[0061] Furthermore, the system also includes: a knowledge graph update module, used for: The rule clauses in the charging knowledge graph are updated and conflict detected through an incremental update pipeline; In response to a successful update, the version number of the rule terms document is incremented, and the retrieval index associated with the charging knowledge graph is updated synchronously. Construct a thesaurus covering common term variations in entity and rule clause document information in the charging knowledge graph; The thesaurus is applied to entity linking, the charging knowledge graph retrieval, and the charging conversation answer generation process; When an erroneous update is detected in the charging knowledge graph, a version rollback operation is performed to restore the charging knowledge graph to its state before the update.

[0062] Furthermore, the node includes at least one of the following: a site, a charging pile, an order, a payment tool, an invoice, a member, and a service package; The rules and terms include at least one of the following: billing rules, membership benefits rules, refund rules, and invoicing rules; The multi-source data documents include: FAQs on charging services, operating procedures and processes for charging services, charging billing rules and pricing policies, equipment information and status documents for charging stations and charging piles, and user order and transaction record documents.

[0063] Example 3: like Figure 5 As shown, the present invention also provides an electronic device, which may be a computer device, a microcontroller device, a smart mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, processor, and transceiver component are connected via a bus; the memory can be used to store executable programs, and an exemplary executable program may include instructions; the processor is used to execute the instructions stored in the memory. The memory can also be used to store data, which can be accessed and / or modified when instructions are executed.

[0064] The processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, and it is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the storage medium to realize the corresponding method flow or corresponding function, so as to realize the steps of the intelligent customer service robot question-and-answer method based on charging knowledge graph in the above embodiments.

[0065] Example 4: Based on the same inventive concept, this invention also provides a readable storage medium, specifically an electronic device readable storage medium (Memory). This readable storage medium is a memory device within an electronic device used to store programs and data. It is understood that the storage medium here can include both built-in storage media within the electronic device and extended storage media supported by the electronic device. The storage medium provides storage space, which stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more executable programs (including program code). It should be noted that the storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. Loading and executing one or more instructions stored in the storage medium by the processor can implement the steps of the intelligent customer service robot question-and-answer method based on a charging knowledge graph in the above embodiments.

[0066] Those skilled in the art will understand that embodiments of this invention can be provided as methods, systems, or computer program products. Therefore, this invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0067] This invention application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0068] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0069] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit its protection scope. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the application, but these changes, modifications or equivalent substitutions are all within the protection scope of the claims pending approval.

Claims

1. A question-and-answer method for an intelligent customer service robot based on a charging knowledge graph, characterized in that, include: An intelligent customer service robot is used to receive charging-related inquiries from users. The intent of the charging session query is identified by performing intent recognition. Based on the query intent, entity links are performed in the pre-built charging knowledge graph to obtain the entities linked to by the query intent; Based on the linked entities and the query intent, a multi-hop search is performed in the charging knowledge graph to obtain nodes consistent with the intent slot constraints and associated rule clauses; based on the obtained nodes and associated rule clauses, an evidence chain is generated. Based on the evidence chain and predefined decision logic, the large language model is driven to generate a natural language answer; the natural language answer and the corresponding rule clause document information are combined to generate a charging session answer, which is then fed back to the user through the intelligent customer service robot.

2. The method according to claim 1, characterized in that, The process of constructing the charging knowledge graph includes: From multiple data documents including FAQs, business procedures, billing rules, site information, and order information for the charging business, information extraction technology is used to obtain entities, relationships, entity attributes, and rule terms. The entity is defined as a node in the knowledge graph, the relationship is defined as an edge in the knowledge graph, the attribute is defined as a feature of the node, and the rule clause is associated and bound to the corresponding node. Determine the version of the rule clauses and add corresponding version identifiers to the rule clauses to form a charging knowledge graph.

3. The method according to claim 2, characterized in that, The entities include: sites, charging piles, connectors, billing, orders, payments, invoices, members, service packages, channels, and rules; the relationships include at least one of the following: a "site-owner-device" relationship based on the ownership relationship between sites and charging piles; a "billing-applies to-site / order" relationship based on the applicability relationship between billing rules and sites or orders; a "member-constraint-rule" relationship based on the constraint relationship between members and rules; a "invoice-used for-order" relationship based on the usage relationship between invoices and orders; a "FAQ-answer-rule or process" relationship based on the answer relationship between FAQs and rules or processes; and a "channel-support-payment" relationship based on the support relationship between channels and payments.

4. The method according to claim 1, characterized in that, Based on the linked entities and the query intent, a multi-hop search is performed in the charging knowledge graph to obtain nodes and associated rule clauses that are consistent with the intent slot constraints. Based on the acquired nodes and associated rule clauses, an evidence chain is generated, including: starting from the linked entity, performing a multi-hop search in the corresponding subgraph of the charging knowledge graph using a graph traversal algorithm to obtain nodes consistent with the intent slot constraints; extracting the associated rule clauses from the acquired nodes; vectorizing the text content of the associated rule clauses and performing semantic retrieval to filter out candidate rule clause fragments that are semantically similar to the query intent; and integrating the acquired nodes, the candidate rule clause fragments, the associated rule clause version information, and the corresponding multi-source data documents to generate an evidence chain.

5. The method according to claim 4, characterized in that, Starting from the linked entity, a graph traversal algorithm is used to perform a multi-hop search in the corresponding subgraph of the charging knowledge graph to obtain nodes consistent with the intent slot constraints, including: Starting from the linked entity, a graph traversal algorithm is used to perform a multi-hop search with a limited depth in the corresponding subgraph of the charging knowledge graph. During each hop of the search, according to the intent slot constraint in the query intent, only the edges that are consistent with the intent slot constraint are traversed, and the nodes that are consistent with the intent slot constraint are selected.

6. The method according to claim 4, characterized in that, The process of integrating the acquired nodes, candidate rule clause fragments, associated rule clause version information, and corresponding multi-source data documents to generate an evidence chain includes: The acquired nodes, candidate rule clause fragments, associated rule clause version information, and corresponding multi-source data documents are packaged to form an initial evidence set; the initial evidence set is then subjected to consistency verification based on the charging session query, and evidence fragments that conflict with session constraints or have expired are removed; the evidence fragments retained after verification are reordered according to their relevance to the query intent to form the final evidence chain.

7. The method according to any one of claims 1-6, characterized in that, The process of generating the charging session answer includes at least one of the following: When, based on the evidence chain, it is identified that a charging session query is ambiguous due to the lack of a key intent slot constraint, a charging session answer containing common scenarios is generated based on the natural language answer and the default rule clause corresponding to the natural language answer. A clarifying question regarding the key intent slot constraint is embedded in the charging session answer, along with the multi-source data document corresponding to the charging session answer. When the evidence chain indicates the existence of multiple effective versions of the rule clause, the latest effective version of the rule clause corresponding to the natural language answer is preferentially selected to generate the charging session answer, and the version of the rule clause on which it is based is marked in the charging session answer. The multi-source data document corresponding to the charging session answer is also embedded. If the latest effective version of the rule clause has applicability differences from other historical versions of the rule clause, these differences are also marked in the charging session answer. The intent slot constraint and the key intent slot constraint both include at least one of the following: the channel used by the user to initiate the charging session inquiry, the user's location, or the time information involved in the inquiry.

8. The method according to claim 1, characterized in that, Following the construction of the charging knowledge graph, the following is also included: The rule clauses in the charging knowledge graph are updated and conflict detected through an incremental update pipeline; In response to a successful update, the version number of the rule terms document is incremented, and the retrieval index associated with the charging knowledge graph is updated synchronously. Construct a thesaurus covering common term variations in entity and rule clause document information in the charging knowledge graph; The thesaurus is applied to entity linking, the charging knowledge graph retrieval, and the charging conversation answer generation process; When an erroneous update is detected in the charging knowledge graph, a version rollback operation is performed to restore the charging knowledge graph to its state before the update.

9. The method according to claim 1, characterized in that, The node includes at least one of the following: site, charging pile, order, payment tool, invoice, member, and service package; The rules and terms include at least one of the following: billing rules, membership benefits rules, refund rules, and invoicing rules; The multi-source data documents include: FAQs on charging services, operating procedures and processes for charging services, charging billing rules and pricing policies, equipment information and status documents for charging stations and charging piles, and user order and transaction record documents.

10. A smart customer service robot question-and-answer system based on a charging knowledge graph, characterized in that, include: The system application layer is used to employ intelligent customer service robots to receive charging conversation inquiries input by users; The intent of the charging session query is identified by performing intent recognition. The functional component layer is used to perform entity linking in the pre-built charging knowledge graph based on the query intent to obtain the entity linked to by the query intent; based on the linked entity and the query intent, perform multi-hop search in the charging knowledge graph to obtain the node consistent with the intent slot constraint and the associated rule clause; and generate an evidence chain based on the obtained node and associated rule clause. The basic support layer is used to drive the large language model to generate natural language answers based on the evidence chain and predefined decision logic; combine the natural language answers with the rule clause document information corresponding to the natural language answers to generate charging session answers, and then feed them back to the user through the intelligent customer service robot.

Images