Intelligent customer service intelligent question-answering system fusing emotion and user behavior

Abstract

The invention relates to the technical field of internet data services, and particularly discloses an intelligent customer service intelligent question-answering system fusing emotion and user behaviors. The method is used for solving the problems of session cross-end confluence and identity adhesion, asynchronous knowledge caliber and transaction state, inconsistency between off-line evaluation and on-line routing, conflict between price insurance and price changing caliber and insufficient tool calling idempotent governance in high-concurrency scenes such as live broadcast and e-commerce. The system comprises a user behavior classifier, a state perception collector, a threshold determination unit, a cooperative decoder, a feedback control unit and a solution device. According to the method, user behavior classification feature element mapping according to the behavior ontology and the scene routing table is realized, retrieval, generation and execution arrangement are combined with threshold admission and double-variant backflow driving, and trace-leaving playback is given.

Description

Technical Field

[0001] This invention relates to the field of Internet data service technology, and in particular to an intelligent customer service question-and-answer system that integrates emotion and user behavior. Background Technology

[0002] The publicly available document "Design and Implementation of an LLM-Based Intelligent Customer Service System" on CNKI discloses an LLM-based intelligent customer service system for supply and marketing cooperatives. This system uses a large language model as its core, combining example interpolation, prompt engineering, and retrieval enhancement to cover knowledge. On the engineering side, it employs domain-driven design to achieve decoupled expansion. The quality inspection module introduces a collaborative framework of rule-based labeling, algorithmic labeling, and manual labeling. On the algorithm side, it uses a combination of a large model and locally fine-tuned small models to improve the accuracy and throughput of high-dimensional classification labeling. However, in the high-concurrency, high-timeliness, and high-state linkage scenarios of live-stream e-commerce, problems such as cross-platform merging of intelligent customer service conversations and identity drift, asynchrony between knowledge caliber and transaction state, inconsistency between offline evaluation and online routing, conflicts between price protection and price adjustment calibers, and insufficient idempotency governance of tool calls occur concurrently. Price protection and difference processing... The data sources, time granularity, permission boundaries, and version of statements involved in different user behaviors such as coupon validity and activity verification, multi-channel conversation merging, and identity resolution vary. This results in heterogeneous input semantics, incomparable indicators, inconsistent thresholds, and no clear target for feedback parameter tuning. Consequently, intelligent customer service cannot establish end-to-end causal consistency and strong consistency control that is thresholdable, alignable, and traceable for conversation identity merging, knowledge and strategy versions, real-time transaction status, online routing, and external tool execution within a single decision cycle. This manifests as distorted joint judgments, mismatched retrieval and generation, loss of idempotency in execution, and answer reversal. Consequently, within the system framework mentioned in the above literature, the intent recognition-decision management control-task distribution link of intelligent customer service cannot stably output a unified answer based on big data of user behavior and a large language model within a single decision cycle. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide an intelligent customer service question-and-answer system that integrates emotion and user behavior. Through five-channel state perception based on a conversation event bus, joint threshold access, collaborative decoding of dual-variable backflow, and solution orchestration decision chain, user behavior is classified and mapped for elements one to five. Two variables, causal consistency and strategy risk, are generated and backflowed for linkage retrieval, semantics and tool orchestration. Finally, the solution engine produces executable responses and instructions.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] This intelligent customer service and Q&A system integrates emotion and user behavior, including a state-aware collector, a user behavior classifier, a threshold judgment unit, a collaborative decoder, a feedback control unit, and a solution generator. The state-aware collector is equipped with transaction sequence, identity permissions, fulfillment resources, version definitions, and semantic slots to acquire elements one through five respectively. Elements one through five are then processed by the user behavior classifier to handle price protection and difference processing, coupon validity and activity definition verification, multi-channel conversation merging and identity resolution, order splitting and shipping adjustments, and user-related return / refund and after-sales applications. Behavior classification performs element mapping. The threshold determination unit compares element one to element five with the corresponding threshold to generate an admission flag. When the admission flag is true, the collaborative decoder receives element one to element five and outputs variable one and variable two. The feedback control unit feeds variable one back to the caliber version, semantic slot, and transaction timing channel, and synchronizes the verification item to the fulfillment resource channel. Variable two feeds back to the identity and permission, fulfillment resource, and transaction timing channels, and links with the caliber version channel. The solution strategist forms intelligent customer service response data and execution instructions based on variable one, variable two, and each element.

[0006] As a further aspect of this invention, the user behavior classifier takes data aggregated by the session event bus as input, and follows a cascaded process of rule template matching + local fine-tuning of small models for multi-class discrimination + large language model verification. Based on the preset mapping of behavior ontology and scenario routing table, it outputs user behavior classification labels for price protection and difference processing, coupon validity and activity calibrator verification, multi-channel session merging and identity resolution, order splitting and shipping adjustment, and return and refund and after-sales application. The preset behavior ontology and scenario routing table extract behavior ontology by offline aggregation of business logs and policy documents, and generate a preliminary routing draft by combining rule templates and local fine-tuning of small models for multi-class classification. Threshold and priority learning is performed on the playback set and solidified into a scenario routing table. It is then iteratively updated with grayscale calibration and online hinged loop. The local fine-tuning small model is a BERT model.

[0007] As a further aspect of this invention, when the state-aware collector acquires elements one through five, and the user behavior tags are price protection and difference processing, element one is the event sequence of events related to price change, order placement, price modification, coupon redemption and expiration, order splitting and shipping, and the current judgment time anchor point; element two is the membership level, price protection qualification switch, price protection frequency and limit, account and risk control mark, qualification validity period and authorization status; element three is the price caliber mapping of multi-warehouse projects and orders, the correspondence between the listed price and the settlement price, the status of in-transit and shipping nodes, and the constraints of price modification and price difference refund; element four is the price protection clause version, price list version, calculation caliber and scope of use and its effective time; element five is the price protection intent type, order and goods reference parsing, difference amount and currency slot, relevant time window and evidence citation slot.

[0008] As a further aspect of this invention, when the state-aware collector acquires elements one through five, and the user behavior tags are for coupons and activity caliber verification, element one is the time sequence of coupon activation, freezing, redemption, expiration, and activity window, as well as the current time anchor point; element two is the eligibility for use, audience group, number of uses and face value limits, account traffic restriction flag, and authorization status; element three is the set of compatible products, partitioned inventory, store and channel scope, and order compatibility mapping; element four is the activity rules, overlapping relationships, redemption logic, and version and scope of the display caliber; and element five is the activity identification, coupon type, threshold condition slot, bound order, and product reference.

[0009] As a further aspect of the present invention, when the state-aware collector acquires elements one to five, and the user behavior tags are multi-channel session merging and identity resolution, element one is the channel arrival order, cross-terminal timeline, message synchronization delay, and session start and end anchor points; element two is the account binding relationship, device fingerprint, login status and authorization chain, and identity merging strategy; element three is the work order archiving location and channel routing, the association of internal and external channel performance records, and the context carrying location; element four is the channel priority, routing table and display caliber version, and conflict resolution rule version; and element five is the identity matching confidence, referencing resolution result, context merging slot, and session tag.

[0010] As a further aspect of the present invention, when the status-aware collector acquires elements one to five, and the user behavior tag is order splitting and shipping adjustment, element one is the time sequence of order splitting, picking, shipping, and signing, and the most recent change anchor point; element two is change authorization, address and specification change permissions, risk control tags, and operation window restrictions; element three is the status of the line item in transit, inventory occupancy, batch or wave information, and path modifiability; element four is the version of the fulfillment change rule, order merging and splitting constraints, and address and specification change rule version; and element five is the change intent, address and specification slots, line item designation, and receiving information slots.

[0011] As a further aspect of the present invention, when the state-aware collector acquires elements one through five, and the user behavior tags are return / refund and after-sales application, element one is the time sequence of receipt, warranty, after-sales application, review and payment, and the time of voucher submission; element two is after-sales permissions, evidence viewing permissions, payment account verification and privacy desensitization authorization; element three is return / exchange judgment, serial number or batch binding, reverse logistics routing and pickup capability; element four is after-sales policy version, voucher requirement template, and billing and deduction caliber version; and element five is the problem type, quantity and reason slot, pickup and refund path slot, and associated order reference.

[0012] As a further aspect of this invention, the threshold determination unit performs deduplication alignment and standardized scoring on the session event and the five channel elements at the current time anchor point. It compares element one to element five with the threshold of the corresponding user behavior classification one by one, and superimposes hard gate verification, risk upper limit, and evidence completeness constraints. After short window hysteresis steady-state processing, it outputs the admission flag and the minimum blocking factor. When the admission flag is true, the collaborative decoder uses element one to element five and their channel weights and the nearest neighbor historical window as input to construct a joint representation and perform attention fusion solution on the causal dependency structure. It calculates and outputs variable one and variable two, which are the causal consistency coefficient and the policy risk potential field, respectively.

[0013] As a further embodiment of the present invention, the caliber version channel is provided with an access point 1, a version parameter mapper 1, a version configuration writer 1, and a strategy template selector; the semantic slot channel is provided with an access point 2, a semantic window parameter parser, and a semantic configuration writer; the transaction timing channel is provided with an access point 3, a time window parameter parser, a timing configuration writer, a mutual exclusion parameter parser, and a mutual exclusion configuration writer; the fulfillment resource channel is provided with an access point 4, a status mapping verifier, an execution parameter parser, and an execution configuration writer; and the identity and permission channel is provided with an access point 5, a permission parameter parser, and a permission configuration writer.

[0014] The feedback control unit sends variable 1 to: the first entry point of the entry version channel, which is processed by the first version parameter mapper and the first version configuration writer; the second entry point of the semantic slot channel, which is processed by the semantic window parameter parser and the semantic configuration writer; the third entry point of the transaction timing channel, which is processed by the time window parameter parser and the timing configuration writer; and sends the verification item of variable 1 to the fourth entry point of the fulfillment resource channel, which is processed by the status mapping verifier.

[0015] The feedback control unit sends variable 2 to: access point 5 of the identity and permission channel, which is processed by the permission parameter parser and permission configuration writer; access point 4 of the fulfillment resource channel, which is processed by the execution parameter parser and execution configuration writer; and access point 3 of the transaction timing channel, which is processed by the mutual exclusion parameter parser and mutual exclusion configuration writer, and in conjunction with the policy template selector of the caliber version channel, generates a template selection parameter set.

[0016] As a further aspect of this invention, the process by which the solution generator forms intelligent customer service response data and execution instructions based on Variable 1, Variable 2, and various elements includes: reading the corresponding rule template and response data structure according to user behavior classification tags; determining the version and scope of application of the terms and prices based on Element 4 and Variable 1; evaluating the fields of orders, execution items, and transaction sequences based on Element 1 and Element 3 and filling in the response fields and execution parameters; selecting an external tool set and determining the calling order, parameters, quota, concurrency, idempotency key, and time window order by referring to Element 2 and Variable 2; combining the aforementioned determination of the scope, evaluation of the fields, and orchestration of the calls into intelligent customer service response data and execution instructions; and recording the corresponding knowledge version, strategy version, time anchor, session identifier, and order identifier; when the admission flag is false, outputting the minimum supplementary item and blocking factor description, and sending it back to the corresponding channel by the feedback control unit.

[0017] The technical effects of the system proposed in this invention are as follows:

[0018] This invention introduces five-channel state awareness acquisition, joint threshold access, five-element collaborative decoding, feedback loop and solution orchestration into the intent recognition-conversation control-task distribution link of the customer service module. This enables end-to-end causal alignment and strong consistency collaboration of conversation identity, knowledge and policy version, real-time transaction status, online routing and external tool execution under key user behaviors in live streaming, such as price protection and difference processing, coupon validity and activity definition verification, multi-channel conversation merging and identity resolution, order splitting and shipping adjustment, returns and refunds and after-sales applications, within a single decision cycle. It stably forms a closed loop of response and execution that is unified in definition, clear in boundary, idempotent, controllable and traceable, and is effective under high concurrency and high-time conditions. Under effective conditions, it reduces answer flipping and rule conflicts, reduces duplicate execution and work order reflow, supports gray-scale changes and rapid retrospection, and improves the adaptability of intelligent customer service to complex business states. At the same time, through user behavior tagging element mapping and channelized reflow, it constrains the adaptive configuration of retrieval granularity, index refresh, semantic window, tool order and limit under the sufficiency of evidence and risk level, ensuring that data from different sources, different time granularities and different permission boundaries are parsed, merged and arranged under a unified standard. It maintains consistent input and output relationships and reproducible judgment criteria for price protection, split shipment, coupon verification and merged identity path, and supports audit traceability and playback review to ensure configuration rollback. Attached Figure Description

[0019] Figure 1 This invention provides a technical roadmap for an intelligent customer service and intelligent question-answering system that integrates emotion and user behavior.

[0020] Figure 2 This is a roadmap for the user behavior classifier technology of the present invention;

[0021] Figure 3This is a technical roadmap for the threshold determination unit of the present invention;

[0022] Figure 4 This is a schematic diagram of the data feedback control unit of the present invention. Figure 1 ;

[0023] Figure 5 This is a schematic diagram of the data feedback control unit of the present invention. Figure 2 ;

[0024] Figure 6 This is a technical roadmap for the solution generator of the present invention to generate intelligent customer service response data and execution instructions. Detailed Implementation

[0025] 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.

[0026] like Figure 1 As shown, the intelligent customer service question-and-answer system proposed in this invention integrates emotion and user behavior, including a state-aware collector, a user behavior classifier, a threshold judgment unit, a collaborative decoder, a feedback control unit, and a solution generator. The state-aware collector is equipped with transaction sequence, identity permissions, fulfillment resources, caliber version, and semantic slot channels to obtain elements one through five respectively. Elements one through five are obtained by the user behavior classifier for price protection and difference processing, coupon validity and activity caliber verification, multi-channel conversation merging and identity resolution, order splitting and shipping adjustment, returns and refunds, and after-sales service. The user behavior classification is used for element mapping. The threshold judgment unit compares element one to element five with the corresponding threshold to generate an admission flag. When the admission flag is true, the collaborative decoder receives element one to element five and outputs variable one and variable two. The feedback control unit feeds variable one back to the caliber version, semantic slot, and transaction timing channel, and synchronizes the verification item to the fulfillment resource channel. Variable two feeds back to the identity and permission, fulfillment resource, and transaction timing channels, and links with the caliber version channel. The solution strategist forms intelligent customer service response data and execution instructions based on variable one, variable two, and each element.

[0027] In the publicly available literature on background technology, the original customer service module's intent recognition-conversation central control management-task distribution chain, after channel access, sequentially performs intent recognition and slot extraction, query construction and retrieval enhancement, generation based on prompt templates, rule or small model rearrangement, strategy routing, and triggering external tools. Afterwards, an offline evaluation loop is formed by rule annotation, algorithm annotation, and manual sampling. Continuing to use this chain in high-concurrency, time-sensitive, and state-linked live streaming scenarios will lead to problems such as cross-platform conversation merging and identity drift, asynchrony between knowledge caliber and transaction status, inconsistency between offline evaluation and online routing, conflicts between price protection and price modification calibers, and insufficient idempotent governance of tool calls. This makes it difficult to establish end-to-end consistency in a single decision-making process. Causal consistency and strong consistency control can lead to issues such as answer reversal, duplicate execution, conflicting statements, and latency jitter. These issues are particularly pronounced in price protection and difference handling, coupon validity and activity statement verification, multi-channel conversation merging and identity resolution, order splitting and shipping adjustments, returns and refunds, and after-sales applications. This invention adds five-channel state awareness collection and scenario-based element mapping, joint threshold access and collaborative decoding, cross-channel feedback and causal dependency governance at the dialogue orchestration and strategy routing stages. This enables the retrieval granularity and refresh, semantic window, tool order and limit to be linked with causal consistency and risk level and aligned with the version time anchor point, achieving a stable online output with unified statements, idempotent governance, traceability, and gray-scale rollback capability.

[0028] It should be noted that, as Figure 2 As shown, the user behavior classifier takes data aggregated from the session event bus as input, and follows a cascaded process of rule template matching + local fine-tuning of small models for multi-class discrimination + large language model verification. Based on the preset behavior ontology and scenario routing table mapping, it outputs user behavior classification labels for price protection and difference processing, coupon validity and activity calibrator verification, multi-channel session merging and identity resolution, order splitting and shipping adjustment, and return and refund and after-sales application. The preset behavior ontology and scenario routing table extracts behavior ontology from offline aggregated business logs and policy documents, and generates a draft route by combining rule templates and local fine-tuning of small models for multi-class classification. Threshold and priority learning is performed on the playback set and solidified into a scenario routing table. It is iteratively updated with grayscale calibration and online hinged loop. The local fine-tuning small model is the BERT model. This cascaded process uses the unified timing of the session event bus as a benchmark. First, the candidate is converged by the rule template. Then, the score is given by the local fine-tuning of BERT multi-class discrimination. Finally, the large language model is verified and anchored by the behavioral ontology and the scene routing table. Combined with the learning of replay set thresholds and priorities and the online iteration of grayscale, it enables a consistent, robust and traceable single-label output for various user behaviors in a high-concurrency multi-channel environment, such as price protection and difference processing, coupon validity and activity caliber verification, multi-channel session merging and identity resolution, order splitting and shipping adjustment, return and refund and after-sales application. It also provides user behavior classification parameter entry for subsequent retrieval, generation and execution orchestration.

[0029] For example, at 22:18 on June 18th, a user asked in the live stream chat, "The price just dropped, can I get a refund for the price difference? Does the coupon still apply?" The same account simultaneously sent the order number and screenshot via the app's customer service IM. The order was paid for at 21:59, using a 20 RMB discount coupon for orders over 100 RMB. At 22:15, the same item dropped from 150 RMB to 130 RMB. In the multi-warehouse order splitting process, one item had already been shipped, and the other was awaiting picking. The session event bus converges messages and transaction events from both ends, unifying the time anchor point to 22:18, forming a timeline sorted by order → coupon redemption → price adjustment → order splitting and shipping. Rule templates that match orders with price changes and coupon redemptions within the recent window prioritize price protection and price difference processing. Simultaneously, due to the involvement of coupon inclusion, supplementary candidate coupon validity and activity definition verification are performed. Small model discrimination (local fine-tuning of BERT): Input comes only from the encoding of bus aggregation, outputting five categories of user behavior classification scores: coupon validity and activity caliber verification 0.84, price protection and difference handling 0.10, multi-channel conversation merging and identity resolution 0.03, order splitting and shipping adjustment 0.02, return and refund and after-sales application 0.01, and returning attention hotspots (fragments of text "can it be stacked", event "near expiration"). Large model verification: Semantic consistency checks and boundary condition reviews are performed on the user's original words and evidence extracts to confirm that the refund and difference request is the main one, the coupon calculation caliber needs to be determined within the price protection scenario, and the generation of verification tickets did not trigger reclassification suggestions. Gating decision: After the collaborative decoder provides the causal consistency coefficient and policy risk potential field, the classifier uses this gating to weight the small model results, and the final label is determined as price protection and difference handling. Results and Feedback: Output user behavior tags for subsequent links, and feed back channel attention weights to the state-aware collector to indicate that in the price protection scenario, priority should be given to strengthening the price table and terms version fields and maintaining the time window settings for order line item mapping.

[0030] It should be noted that, as Figure 3As shown, the threshold determination unit performs deduplication alignment and standardized scoring on the session event and the five channel elements at the current time anchor point. It compares element one to element five with the threshold of the corresponding user behavior classification one by one, and superimposes hard gate verification, risk upper limit, and evidence completeness constraints. After short window hysteresis steady-state processing, it outputs the admission flag and the minimum blocking factor. When the admission flag is true, the collaborative decoder takes element one to element five and their channel weights and the nearest neighbor historical window as input, constructs a joint representation, and performs attention fusion solution on the causal dependency structure. It calculates and outputs variable one and variable two, which are the causal consistency coefficient and the policy risk potential field, respectively. Each threshold determination unit first generates a verifiable single admission flag and a minimum blocking factor to prevent the premature entry of missing evidence and high-risk samples. Using a causal consistency coefficient and a strategy risk potential field at a unified time anchor, it provides parameterizable control over retrieval granularity, index refresh, semantic window, and tool call order and limits. Combined with short-window hysteresis and channel weight fusion, it mitigates jitter and cross-channel differences, thereby achieving consistent alignment of input evidence, an idempotent governance entry point for execution paths, and a stable interface for subsequent orchestration. The threshold determination unit facilitates evidence alignment within a single decision, version and permission unification, and controllable linkage between retrieval, generation, and execution. It stably outputs consistent, traceable, auditable, and idempotent rollback responses and instructions, forming a consistency control chain.

[0031] It should be noted that, as Figure 4 and Figure 5 As shown, the caliber version channel has an access point 1, a version parameter mapper 1, a version configuration writer 1, and a strategy template selector; the semantic slot channel has an access point 2, a semantic window parameter parser, and a semantic configuration writer; the transaction timing channel has an access point 3, a time window parameter parser, a timing configuration writer, a mutual exclusion parameter parser, and a mutual exclusion configuration writer; the fulfillment resource channel has an access point 4, a status mapping validator, an execution parameter parser, and an execution configuration writer; and the identity and permission channel has an access point 5, a permission parameter parser, and a permission configuration writer.

[0032] Specifically, the feedback control unit sends Variable 1 to two interfaces: Interface 1 of the caliber version channel, where it is processed by Version Parameter Mapper 1 and Version Configuration Writer 1. This resolves Variable 1 into caliber and version selection and binding parameters, determining the version number, switching method, and effective scope of the terms and prices, and writing the corresponding parameters into the subsequently referenced version pointer and configuration items. Interface 2 of the semantic slot channel is also sent to it, where it is processed by the Semantic Window Parameter Parser and Semantic Configuration Writer. This resolves Variable 1 into semantic processing parameters, setting the semantic window size, context capture boundary, rewrite strength, and candidate number, and writing these parameters into the prompt template and slot. The run configuration is filled in; it is sent to the third interface of the transaction timing channel, where it is processed by the time window parameter parser and timing configuration writer. The time window parameter parser parses the variable 1 into timing processing parameters, sets the event time window span, sorting and alignment strategy, water level and sampling rhythm, and writes them into the run configuration of the event queue and sorting unit. The verification items of variable 1 are sent to the fourth interface of the fulfillment resource channel, where they are processed by the status mapping verifier. The causal verification field carried by variable 1 is mapped into a checklist of status reading and field correspondence, generating reading consistency and field mapping rules for subsequent data retrieval and mapping configuration of resources such as orders and line items.

[0033] The feedback control unit sends variable two to: first, into access point five of the identity and permission channel, where the permission parameter parser and permission configuration writer process it, parsing variable two into permission parameters such as account and member verification depth, limit and frequency thresholds, whether to enable secondary verification and manual review, approval path, and blacklist / whitelist switch, and writes them into the corresponding permission configuration items; and second, into access point four of the fulfillment resource channel, where the execution parameter parser and execution configuration writer process it, parsing variable two into execution parameters such as external tool call order, concurrency and queuing weight, rate limiting and limit cap, retry and fallback strategies, and idempotent key generation and verification strategies, and writes them into the work... It has an orchestration configuration item; it is sent to the third interface of the transaction timing channel, and processed by the mutual exclusion parameter parser and mutual exclusion configuration writer. It parses variable two into mutual exclusion control parameters such as the mutual exclusion lock enable flag, lock granularity and locking duration, conflict handling and serialization strategy, and window entry and release conditions within the transaction time window, and writes them into the timing control configuration item. It also links with the policy template selector of the caliber version channel to generate a template selection parameter set. Based on the classification result of variable two, it generates a template selection parameter set to specify the policy template set, template priority and rollback template, switching method and effective boundary conditions, and sends it to the template selection configuration item of the caliber version channel.

[0034] It should be noted that, as Figure 6As shown, the process by which the solution generator forms intelligent customer service response data and execution instructions based on Variable 1, Variable 2, and various elements includes: reading the corresponding rule template and response data structure according to user behavior classification tags; determining the version and scope of application of the terms and prices based on Element 4 and Variable 1; evaluating the fields of orders, execution items, and transaction sequences based on Element 1 and Element 3 and filling in the response fields and execution parameters; selecting an external tool set and determining the calling order, parameters, limits, concurrency, idempotency keys, and time window order by referring to Element 2 and Variable 2; combining the aforementioned determination of the scope, evaluation of the fields, and orchestration of the calls into intelligent customer service response data and execution instructions, and recording the corresponding knowledge version, strategy version, time anchor, session identifier, and order identifier; when the admission flag is false, outputting the minimum supplementary item and blocking factor description, and sending it back to the corresponding channel by the feedback control unit.

[0035] Specifically, the joint targeting technology of Element 4 and Variable 1 is used to accurately select, bind versions, and limit the scope of application of terms and prices at the current time anchor point, constrain the reference boundaries of terms and unify field interpretation; the field calculation technology of Element 1 and Element 3 is used to evaluate on the same mapping of orders, execution projects, and transaction sequences, and to standardize the filling order and data source of amounts, nodes, and mapping relationships; referring to the execution orchestration technology of Element 2 and Variable 2, the set of external tools, the order of invocation, parameter templates, quotas and concurrency, idempotent keys, and time window order are determined under the constraints of permissions and risk levels, and the executable path is solidified; and the determination of terms, field evaluation, and execution orchestration are structured and synthesized, with knowledge version, strategy version, time anchor point, session and order identifiers, forming routable, recordable, and replayable response and instruction objects to support subsequent feedback and governance.

[0036] To clearly describe the five typical user behavior classifications for intelligent customers in live-streaming e-commerce within the system proposed in this invention, this invention introduces and explains the intelligent customer service dialogue generation and optimization scheme based on the corresponding elements one to five collected by the state-aware collector in the system.

[0037] Example 1

[0038] This implementation describes the elements one through five collected by the state-aware collector when the user behavior tags are price protection and difference processing.

[0039] In the system proposed in this invention, when the state-aware collector acquires elements one through five, and the user behavior tags are price protection and difference processing, element one is the event sequence of events related to price change, order placement, price modification, coupon redemption and expiration, order splitting and shipping, and the current judgment time anchor point; element two is the membership level, price protection qualification switch, price protection frequency and limit, account and risk control mark, qualification validity period and authorization status; element three is the price caliber mapping of multi-warehouse projects and orders, the correspondence between the listed price and the settlement price, the status of in-transit and shipping nodes, and the constraints of price modification and price difference refund; element four is the price protection clause version, price list version, calculation caliber and scope of use and its effective time; element five is the price protection intent type, order and goods reference parsing, difference amount and currency slot, relevant time window and evidence citation slot.

[0040] In this case, the intelligent customer service dialogue generation and optimization are completed through the following process:

[0041] (1) The channel access and conversation aggregation component aggregates events such as bullet comments and IM messages, orders and price changes, and establishes a unified timeline and time anchor point;

[0042] (2) The user behavior classifier uses bus data to determine the labels as price protection and difference processing through rule template → local BERT multi-classification → large model verification;

[0043] (3) The five channels of the state-aware data acquisition unit respectively pull out elements one to five of the price protection and difference processing mapping;

[0044] (4) The threshold determination unit compares the five elements with the scene threshold and the hard gate condition, and generates the admission mark and the blocking factor.

[0045] (5) The collaborative decoder generates variable one and variable two when the admission flag is true;

[0046] (6) The feedback control unit will feed back variable one to the caliber version, semantic slot, and transaction sequence and synchronize the verification items to the performance resources. It will feed back variable two to the identity and permissions, performance resources, and transaction sequence and link the template selection of the caliber version.

[0047] (7) The solution tool reads and binds the terms and price caliber versions according to the price protection template; calculates the line items and amount fields according to element one and element three; forms the external tool call order, parameters, limit, concurrency, idempotent key and time window order according to element two and variable two; and synthesizes them into response data and execution instructions.

[0048] (8) The tool arrangement and execution interface issues instructions such as refund, coupon issuance or price modification;

[0049] (9) Audit and record keeping knowledge and strategy versions, time anchors, session and order identifiers and instruction sequences.

[0050] To address user behavior characteristics related to price protection and difference handling (price changes after order placement, whether coupons are included, and the coexistence of branch line items and partial shipments), this technical approach, at the time anchor point of unified price change - order placement and verification - order splitting and shipment, first uses Element 4 and Variable 1 to complete the targeted binding of terms and price caliber versions, then uses Element 1 and Element 3 to establish a branch line item-level price caliber mapping and difference calculation path, and parametrically arranges the order, amount, concurrency, and idempotency of external actions (refund, coupon issuance, price modification) according to Element 2 and Variable 2, forming a price protection judgment and difference calculation and distribution path with unique caliber, homogeneous fields, and consistent execution chain within a single decision; the solutionr translates the causal consistency coefficient and strategy risk potential field into three types of control variables in real time: template selection, field evaluation, and tool arrangement, generating executable refund, coupon issuance, and price modification plans at the branch line item granularity and assigning idempotency keys, mutually exclusive time windows, and dependency topologies, so that responses and instructions are formed synchronously and can be directly issued and replayed.

[0051] Example 2

[0052] Unlike Example 1, this example describes the specific elements one to five collected by the state-aware collector when the user behavior tag is a coupon and the activity definition is being verified.

[0053] In the system proposed in this invention, when the state-aware collector acquires elements one through five, and the user behavior tags are coupons and activity caliber verification, element one is the time sequence of coupon activation, freezing, redemption, expiration, and activity window, as well as the current time anchor point; element two is the eligibility for use, audience package, number of uses and face value limits, account traffic restriction flag, and authorization status; element three is the set of compatible products, partitioned inventory, store and channel scope, and order compatibility mapping; element four is the activity rules, overlapping relationships, redemption logic, and version and scope of the display caliber; element five is the activity identification, coupon type, threshold condition slot, bound order, and product reference.

[0054] In this case, the intelligent customer service dialogue generation and optimization are completed through the following process:

[0055] (1) The channel access and conversation aggregation component aggregates bullet comments and IM messages, orders and coupons, and event events to establish a unified timeline and time anchor point;

[0056] (2) The user behavior classifier uses bus data to determine the labels for coupon validity and activity definition verification through rule template → local BERT multi-classification → large model verification.

[0057] (3) The five channels of the state-aware collector respectively pull element one to element five mapped by the label;

[0058] (4) The threshold determination unit compares the five elements with the scene threshold and the hard gate condition, and generates the admission mark and the blocking factor.

[0059] (5) The collaborative decoder generates variable one and variable two when the admission flag is true;

[0060] (6) The feedback control unit will feed back variable one to the caliber version, semantic slot, and transaction sequence and synchronize the verification items to the performance resources. It will feed back variable two to the identity and permissions, performance resources, and transaction sequence and link the template selection of the caliber version.

[0061] (7) The solution tool reads and binds the rules and reimbursement logic versions according to the activity and coupon template; evaluates the adaptation, threshold and window fields according to element one and element three; forms the verification, calculation call order, parameters, quota, concurrency, idempotency key and time window order according to element two and variable two; and synthesizes the response data and execution instructions.

[0062] (8) The tool arrangement and execution interface issues instructions for voucher verification, rights determination, reissue and cancellation;

[0063] (9) Audit and record keeping knowledge and strategy versions, time anchors, session and order identifiers and instruction sequences.

[0064] To address the characteristics of user behavior in verifying coupon validity and activity criteria, this technical approach uses a unified time anchor point—order-coupon event-activity window—to overlay and redeem versions with directional binding of element four and variable one. It establishes an adaptation and threshold evaluation path for order-goods-activity and coupon using elements one and three, and parametrically arranges verification and calculation actions based on element two and variable two. The solution interpreter translates the causal consistency coefficient and strategy risk potential field into template selection, threshold field evaluation and redemption, and rights call-up plan, generating an executable solution with time windows and overlay boundaries.

[0065] Example 3

[0066] Unlike Examples 1 and 2, this example provides a detailed description of elements one through five collected by the state-aware collector when the user behavior tags are multi-channel conversation merging and identity resolution.

[0067] In the system proposed in this invention, when the state-aware collector acquires elements one to five, and the user behavior tags are multi-channel session merging and identity resolution, element one is the channel arrival order, cross-terminal timeline, message synchronization delay, and session start and end anchor points; element two is the account binding relationship, device fingerprint, login state and authorization chain, and identity merging strategy; element three is the work order archiving location and channel routing, the association of internal and external channel performance records, and the context carrying location; element four is the channel priority, routing table and display caliber version, and conflict resolution rule version; and element five is the identity matching confidence, referencing resolution result, context merging slot, and session tag.

[0068] In this case, the intelligent customer service dialogue generation and optimization are completed through the following process:

[0069] (1) The channel access and conversation aggregation component aggregates messages and arrival sequences from multiple terminals such as live barrage and APP IM, and establishes a unified timeline and time anchor point;

[0070] (2) The user behavior classifier uses bus data to determine the labels for multi-channel conversation merging and identity resolution through rule templates → local BERT multi-classification → large model verification.

[0071] (3) The five channels of the state-aware collector respectively pull element one to element five mapped by the label;

[0072] (4) The threshold determination unit compares the five elements with the scene threshold and the hard gate condition, and generates the admission mark and the blocking factor.

[0073] (5) The collaborative decoder generates variable one and variable two when the admission flag is true;

[0074] (6) The feedback control unit will feed back variable one to the caliber version, semantic slot, and transaction sequence and synchronize the verification items to the performance resources. It will feed back variable two to the identity and permissions, performance resources, and transaction sequence and link the template selection of the caliber version.

[0075] (7) The resolver reads and binds the channel priority and conflict resolution rule version according to the merge template; calculates the merge key, context bearer and route number according to element one and element three; forms the identity verification depth, deduplication / merging and route adjustment order, parameters, concurrency, idempotent key and time window order according to element two and variable two; and synthesizes it into response data and execution instructions.

[0076] (8) The tool orchestration and execution interface issues instructions such as identity merging and writing, route adjustment and work order merging;

[0077] (9) Audit and record-keeping versions, time anchors, session and account identifiers and instruction sequences.

[0078] To address user behavior characteristics related to multi-channel session merging and identity resolution, this technical approach uses a unified cross-platform timeline—arrival order—session start and end time anchor point. It binds routing and display version with element four and variable one, establishes merging and carries the evaluation path with element one and element three, and performs parameterized orchestration, verification, and routing adjustment based on element two and variable two. The solution translates consistency and risk into merge key generation, route numbering, and mutual exclusion time window control, outputting an executable merging scheme.

[0079] Example 4

[0080] Unlike Examples 1, 2, and 3, this example provides a detailed description of elements one through five collected by the status-aware collector when the user behavior tags are order splitting and shipping adjustment.

[0081] In the system proposed in this invention, when the status-aware collector acquires elements one to five, and the user behavior label is order splitting and shipping adjustment, element one is the time sequence of order splitting, picking, shipping, and signing, and the most recent change anchor point; element two is change authorization, address and specification change permissions, risk control labels, and operation window restrictions; element three is the status of the line item in transit, inventory occupancy, batch or wave information, and path modifiability; element four is the version of the fulfillment change rule, order merging and splitting constraints, and address and specification change rule version; and element five is the change intent, address and specification slots, line item designation, and receiving information slots.

[0082] In this case, the intelligent customer service dialogue generation and optimization are completed through the following process:

[0083] (1) The channel access and session aggregation component aggregates address change / specification change requests and logistics node events to establish a unified timeline and time anchor point;

[0084] (2) The user behavior classifier uses bus data to determine the labels for order splitting and shipping adjustment through rule templates → local BERT multi-classification → large model verification.

[0085] (3) The five channels of the state-aware collector respectively pull element one to element five mapped by the label;

[0086] (4) The threshold determination unit compares the five elements with the scene threshold and the hard gate condition, and generates the admission mark and the blocking factor.

[0087] (5) The collaborative decoder generates variable one and variable two when the admission flag is true;

[0088] (6) The feedback control unit will feed back variable one to the caliber version, semantic slot, and transaction sequence and synchronize the verification items to the performance resources. It will feed back variable two to the identity and permissions, performance resources, and transaction sequence and link the template selection of the caliber version.

[0089] (7) The solution tool reads and binds the performance change and split / merge order constraint versions according to the change template; calculates the changeable nodes, paths and field mappings according to element one and element three; forms the call order, parameters, quota, concurrency, idempotent key and mutual exclusion time window order of address change / specification change / splitting / merge order according to element two and variable two; and synthesizes them into response data and execution instructions.

[0090] (8) The tool orchestration and execution interface issues and intercepts instructions such as reassignment, address change, specification change, and order splitting / merging;

[0091] (9) Audit and record keeping version, time anchor, session and order identifier and instruction sequence.

[0092] Based on the user behavior characteristics of order splitting and shipping adjustments, this technical solution uses Element 4 and Variable 1 to bind the change rule version at the unified time anchor points of order splitting-picking-outbound-in-transit-receipt, and Element 1 and Element 3 to establish a line item-level modifiable node evaluation path, and arranges change calls according to Element 2 and Variable 2 parameterization. The solution translates consistency and risk into a node permission matrix, call sequence and idempotent / mutually exclusive control, and outputs an executable adjustment plan.

[0093] Example 5

[0094] Unlike Examples 1, 2, 3, and 4, this example describes the collection of elements one through five by the status sensing collector when the user's behavior tags are return and refund and after-sales application.

[0095] In the system proposed in this invention, when the state-aware collector acquires elements one to five, and the user behavior is a return / refund or after-sales application, element one is the time sequence of receipt, warranty, after-sales application, review and payment, and the time of voucher submission; element two is after-sales permissions, evidence viewing permissions, payment account verification and privacy desensitization authorization; element three is return / exchange determination, serial number or batch binding, reverse logistics routing and pickup capability; element four is the after-sales policy version, voucher requirement template, and billing and deduction caliber version; element five is the problem type, quantity and reason slots, pickup and refund path slots, and associated order references.

[0096] In this case, the intelligent customer service dialogue generation and optimization are completed through the following process:

[0097] (1) The channel access and conversation aggregation component aggregates events such as after-sales consultation texts, receipts and voucher submissions, and establishes a unified timeline and time anchor point;

[0098] (2) The user behavior classifier uses bus data to determine the labels for return and refund and after-sales application through rule template → local BERT multi-classification → large model verification.

[0099] (3) The five channels of the state-aware collector respectively pull element one to element five mapped by the label;

[0100] (4) The threshold determination unit compares the five elements with the scene threshold and the hard gate condition, and generates the admission mark and the blocking factor.

[0101] (5) The collaborative decoder generates variable one and variable two when the admission flag is true;

[0102] (6) The feedback control unit will feed back variable one to the caliber version, semantic slot, and transaction sequence and synchronize the verification items to the performance resources. It will feed back variable two to the identity and permissions, performance resources, and transaction sequence and link the template selection of the caliber version.

[0103] (7) The solution tool reads and binds the after-sales policy and billing / deduction caliber version according to the after-sales template; calculates fields such as refundable / exchangeable judgment, reverse routing and pickup capability according to element one and element three; forms the pickup, quality inspection, refund / repayment call order, parameters, amount, concurrency, idempotent key and time window order according to element two and variable two; and synthesizes it into response data and execution instructions.

[0104] (8) The tool orchestration and execution interface issues instructions for trip creation, reverse logistics, refund / repayment, and review;

[0105] (9) Audit and record keeping version, time anchor, session and order identifier and instruction sequence.

[0106] Based on user behavior characteristics related to returns, refunds, and after-sales applications, this technical approach uses a unified time anchor point for receiving, applying, reviewing, and making payments. It binds the after-sales and billing versions with element four and variable one, establishes evidence and reverse path evaluation paths with element one and element three, and parametrically arranges pickup, quality inspection, and refund calls according to element two and variable two. The solution interprets consistency and risk into evidence requirements, reverse and funding instruction sequences, idempotency, and window parameters, and outputs an executable after-sales solution.

[0107] In summary, this invention introduces five-channel state perception acquisition, joint threshold access, five-element collaborative decoding, feedback loop and solution orchestration into the intent recognition-conversation control-task distribution link of the customer service module. This enables end-to-end causal alignment and strong consistency collaboration of conversation identity, knowledge and policy version, real-time transaction status, online routing and external tool execution under key user behaviors in live streaming, such as price protection and difference processing, coupon validity and activity definition verification, multi-channel conversation merging and identity resolution, order splitting and shipping adjustment, returns and refunds and after-sales applications. This stably forms a unified, clearly defined, idempotent, controllable, and traceable response and execution closed loop, and is effective under high concurrency and... Under conditions of high timeliness, it reduces answer reversal and rule conflicts, minimizes duplicate execution and work order reflow, supports gray-scale changes and rapid retrospective, and improves the adaptability of intelligent customer service to complex business states. At the same time, through user behavior tagging element mapping and channelized reflow, it constrains the adaptive configuration of retrieval granularity, index refresh, semantic window, tool order and limit under the sufficiency of evidence and risk level, ensuring that data from different sources, different time granularities and different permission boundaries are parsed, merged and arranged under a unified standard. It maintains consistent input-output relationships and reproducible judgment criteria for price protection, split shipment, coupon verification and merged identity paths, and supports audit traceability and playback review to ensure configuration rollback.

[0108] 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.

[0109] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent customer service intelligent question and answer system that fuses emotions and user behaviors, characterized in that, The state-aware collector, the user behavior classifier, the threshold determination unit, the collaborative decoder, the feedback control unit and the solution strategy generator are included. The state-aware collector is provided with transaction timing, identity permission, performance resource, caliber version and semantic slot channel to respectively acquire elements one to five. Elements one to five are mapped according to the user behavior classification of price protection and difference processing, coupon validity and activity caliber checking, multi-channel conversation merging and identity analysis, order splitting and shipping adjustment, return and refund and after-sales application acquired by the user behavior classifier. The threshold determination unit compares elements one to five with corresponding thresholds to generate an access flag. The collaborative decoder receives elements one to five and outputs variables one and two when the access flag is true. The feedback control unit returns variable one to the caliber version, semantic slot and transaction timing channels, and synchronizes the check items to the performance resource channel. Variable two is returned to the identity permission, performance resource and transaction timing channels, and is linked to the caliber version channel. The solution strategy generator generates intelligent customer service response data and execution instructions based on variables one and two and each element.

2. The intelligent customer service intelligent question and answer system fusing emotions and user behaviors according to claim 1, characterized in that, The user behavior classifier takes the data aggregated by the conversation event bus as input, and uses the cascaded process of rule template matching + local fine-tuning small model multi-classification discrimination + large language model checking. According to the preset behavior ontology and scene routing table mapping, the user behavior classification tags of price protection and difference processing, coupon validity and activity caliber checking, multi-channel conversation merging and identity analysis, order splitting and shipping adjustment, return and refund and after-sales application are output. The preset behavior ontology and scene routing table extract the behavior ontology by offline aggregation of business logs and system documents, generate a preliminary draft of the routing table in combination with rule templates and local fine-tuning small model multi-classification, learn the threshold and priority on the replay set, and solidify it into a scene routing table. The scene routing table is updated iteratively in a gray calibration and online reaming closed loop. 3.The intelligent customer service and intelligent question answering system fusing emotion and user behavior according to claim 1, characterized in that, When the state-aware collector acquires elements one to five and the user behavior label is price protection and difference processing, element one is the event sequence and current judgment time anchor point of price change, ordering, price change, coupon cancellation and expiration, order splitting and shipping related events; element two is the member level, price protection qualification switch, price protection frequency and limit, account and risk control mark, qualification validity period and authorization state; element three is the price caliber mapping of multi-warehouse line items and orders, the correspondence between the price and the settlement price, the in-transit and shipping node state, the price change and the difference price constraint; element four is the price protection clause version, the price table version, the calculation caliber and the use range and its effective time; element five is the price protection intent type, order and goods reference analysis, difference amount and currency slot, related time window and evidence reference slot.

4. The intelligent customer service intelligent question and answer system fusing emotion and user behavior according to claim 1, characterized in that, When the state-aware collector acquires elements one to five and the user behavior label is coupon and activity caliber checking, element one is the time sequence and current time anchor point of coupon validity, freezing, cancellation, expiration and activity window; Factor two is the use of qualifications, crowd package, times and denomination restrictions, account flow marks and authorized state; factor three is the adaptation of commodity set, partition inventory, shop and channel range, order adaptation mapping; factor four is the version and effective range of activity rules, superposition relationship, verification logic and display caliber; factor five is activity identification, coupon type, threshold condition slot, binding order and goods reference.

5. The intelligent customer service intelligent question and answer system fusing emotion and user behavior according to claim 1, characterized in that, When the state-aware collector acquires factors one to five, and the user behavior label is multi-channel conversation confluence and identity resolution, factor one is channel arrival order, cross-end timeline, message synchronization delay and conversation start and end anchor point, factor two is account binding relationship, device fingerprint, login state and authorization chain, identity merging strategy, factor three is work order archiving location and channel routing, channel internal and external performance record association, context bearing location, factor four is channel priority, routing table and display caliber version, conflict resolution rule version, factor five is identity matching configuration, reference resolution result, context merging slot and conversation label.

6. The intelligent customer service intelligent question and answer system fusing emotion and user behavior according to claim 1, characterized in that, When the state-aware collector acquires factors one to five, and the user behavior label is single and shipping adjustment, factor one is the time sequence and the latest change anchor point of single, picking, shipping and receipt, factor two is the change authorization, address and specification change permission, risk control label and operation window limit, factor three is the in-transit state of the line item, inventory occupation, batch or wave information, path modifiability, factor four is the version of the performance change rule, the constraint of single and single, the rule version of address and specification change, factor five is the change intention, address and specification slot, line item reference, and delivery information slot.

7. The intelligent customer service intelligent question answering system fusing emotions and user behaviors according to claim 1, characterized in that, When the state-aware collector acquires factors one to five, and the user behavior label is return and refund and after-sales application, factor one is the time sequence and voucher submission time of receipt, warranty, after-sales application, review and payment, factor two is the after-sales permission, evidence viewing permission, payment account verification and privacy desensitization authorization, factor three is the return and exchange determination, serial number or batch binding, reverse logistics routing and pickup capability, factor four is the version of the after-sales policy, the template of the evidence requirement, the caliber version of the billing and deduction, factor five is the problem type, quantity and reason slot, pickup and refund path slot, and associated order reference. 8.The intelligent customer service and smart Q&A system fusing emotion and user behavior according to claim 1, wherein, The threshold determination unit aligns and standardizes the scores of the conversation events and five-channel factors under the current time anchor point, compares factors one to five with the corresponding user behavior classification thresholds one by one, and superimposes hard gate verification and risk upper limit, evidence completeness constraint. After short window delay steady state processing, the access flag and the minimum blocking factor are output; the collaborative decoder takes factors one to five and their channel weights, the near neighbor history window as input when the access flag is true, constructs a joint representation and performs attention fusion solution on the causal dependency structure, calculates and outputs variable one and variable two. Variable one and variable two are causal consistency coefficient and strategy risk potential field respectively. 9.The intelligent customer service and smart Q&A system fusing emotion and user behavior according to claim 1, wherein, The caliber version channel is provided with an entry port one, a version parameter mapper one, a version configuration writer one and a policy template selector; the semantic slot channel is provided with an entry port two, a semantic window parameter parser and a semantic configuration writer; the transaction time sequence channel is provided with an entry port three, a time window parameter parser, a time sequence configuration writer, a mutual exclusion parameter parser and a mutual exclusion configuration writer; the performance resource channel is provided with an entry port four, a state mapping verifier, an execution parameter parser and an execution configuration writer; and the identity permission channel is provided with an entry port five, a permission parameter parser and a permission configuration writer. The feedback control unit sends the variable one to the entry port one of the caliber version channel, which is processed by the version parameter mapper one and the version configuration writer one; sends the variable one to the entry port two of the semantic slot channel, which is processed by the semantic window parameter parser and the semantic configuration writer; sends the variable one to the entry port three of the transaction time sequence channel, which is processed by the time window parameter parser and the time sequence configuration writer, and sends the check item of the variable one to the entry port four of the performance resource channel, which is processed by the state mapping verifier; The feedback control unit sends the variable two to the entry port five of the identity permission channel, which is processed by the permission parameter parser and the permission configuration writer; sends the variable two to the entry port four of the performance resource channel, which is processed by the execution parameter parser and the execution configuration writer; sends the variable two to the entry port three of the transaction time sequence channel, which is processed by the mutual exclusion parameter parser and the mutual exclusion configuration writer, and links the policy template selector of the caliber version channel to generate a template selection parameter set.

10. The intelligent customer service intelligent question answering system fusing emotion and user behavior according to claim 1, characterized in that, The process of the solution device to form the intelligent customer service response data and the execution instruction based on the variable one, the variable two and each element includes: reading the corresponding rule template and the response data structure according to the user behavior classification label, determining the version of the clause caliber and the price caliber and the applicable range combined with the element four and the variable one, performing field evaluation on the order, the execution item and the transaction sequence based on the element one and the element three and completing the filling of the response field and the execution parameter, selecting the external tool set and determining the calling order, the parameter, the quota, the concurrency and the idempotent key and the time window sequence with reference to the element two and the variable two, combining the caliber determination, the field evaluation and the calling arrangement into the intelligent customer service response data and the execution instruction, and recording the corresponding knowledge version, the policy version, the time anchor point, the session identifier and the order identifier; when the access flag is false, outputting the minimum supplement item and the blocking factor explanation, and returning to the corresponding channel by the feedback control unit.

Images