A method and system for manufacturing customer demand response program generation

Abstract

The application discloses a kind of manufacturing customer demand response scheme generation method and system, belong to artificial intelligence and manufacturing knowledge graph technical field.The method obtains the structured demand image of customer demand and manufacturing knowledge graph, generates structured query intent for graph retrieval;Fusion semantic retrieval and entity matching obtain hybrid retrieval result, and execute graph path expansion;To the retrieval triple and expansion path are deduplicated, evidence character labeling and business grouping, construct structured knowledge evidence package;Initial scheme containing knowledge basis, response step, information gap and risk explanation is generated using scheme generation agent;Through evaluation agent and refining agent, multi-dimensional evaluation and multi-round revision are executed;Finally, human-in-the-loop revision is carried out in combination with enterprise employee review and customer feedback, and the final scheme is output.The application improves the knowledge support, business executability and evidence traceability of manufacturing customer demand response scheme.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence, manufacturing knowledge graphs, and intelligent decision-making in manufacturing operations. Specifically, it relates to a method and system for generating manufacturing customer demand response solutions based on knowledge graph evidence enhancement and multi-agent iterative optimization. Background Technology

[0002] Against the backdrop of the manufacturing industry's accelerated transformation towards servitization, flexibility, and intelligence, the interaction between manufacturing enterprises and customers is no longer limited to the confirmation of a single product model or standard specification. Customer needs typically involve multiple dimensions, including material performance, specification customization, process handling, quality improvement, delivery cycle, after-sales service, and technical support, and are often expressed in the form of natural language, tables, illustrative images, or multimodal attachments. Customer need response solutions, as a key vehicle connecting external market demands with internal sales, technology, quality, production, supply chain, and after-sales activities, need to transform the customer's non-standard expressions into clear business intentions and form verifiable recommendations that include knowledge base, processing steps, information gaps, and risk boundaries.

[0003] Existing rule-based and data-driven approaches typically pre-convert customer needs into product attributes, technical parameters, business objectives, or constraints, and then generate product configurations, parameter recommendations, service combinations, or resource matching results through configuration rules, knowledge models, machine learning models, or multi-criteria decision-making methods. While these methods are applicable to scenarios with clearly defined input formats, stable business boundaries, and comprehensive rule sets, they often struggle to produce stable understandings of customer needs when faced with open-ended requirements that are verbally descriptive, lack information, have dynamically changing constraints, and implied intent. Furthermore, they are insufficient to cover comprehensive, cross-departmental collaborative response solutions.

[0004] Large language models possess strong semantic understanding, task decomposition, and natural language generation capabilities, enabling them to assist manufacturing enterprises in interpreting requirements, refactoring problems, and organizing solutions. However, directly using large language models to generate business response content based on original customer needs relies heavily on model parametric knowledge and cue constraints. This can easily lead to professional conclusions that are inconsistent with the enterprise's actual knowledge of products, materials, processes, quality, and services, and it is difficult to explain which enterprise knowledge facts the key judgments in the solution are derived from.

[0005] Retrieval-enhanced generation and knowledge graph-enhanced methods can introduce external documents, knowledge triples, or graph paths into the context of large language models, thereby providing a certain knowledge basis for the generated content. Agent-based methods further utilize role division, tool invocation, and feedback correction to improve the quality of the generated results. However, existing methods typically directly concatenate retrieved document fragments, knowledge triples, or graph paths into the model input, lacking mechanisms for evidence screening, evidence role labeling, business semantic grouping, and step-level evidence binding tailored to manufacturing operations. This easily introduces irrelevant knowledge, duplicate evidence, or contextual noise.

[0006] Furthermore, existing agent evaluations are mostly geared towards general question answering, design results, or scheduling performance, lacking a manufacturing customer demand response evaluation mechanism that simultaneously covers demand matching, knowledge support, structural integrity, business executability, illusion risk control, traceability, and consistency of response scenarios. Automatically generated or revised response plans still require enterprise employees to confirm capacity, process feasibility, quality requirements, and risk boundaries, and also require customer feedback on demand priorities, acceptance criteria, and new constraints. Therefore, there is an urgent need for a manufacturing customer demand response plan generation method that can unify non-standard customer demand parsing, manufacturing knowledge graph evidence acquisition, structured evidence organization, traceable solution generation, multi-agent evaluation revision, and human-in-the-loop feedback revision. Summary of the Invention

[0007] To address the challenges in existing technologies such as the difficulty in reliably parsing non-standard customer demands, the difficulty in effectively organizing retrieved knowledge, the lack of knowledge evidence binding in response schemes, the lack of closed-loop review in automatically generated content, and the difficulty in structurally integrating manual feedback, this invention provides a method and system for generating manufacturing customer demand response schemes based on knowledge graph evidence enhancement and multi-agent iterative optimization.

[0008] The present invention adopts the following technical solution:

[0009] A method for generating a customer demand response solution includes the following steps:

[0010] Step S1: Based on the structured demand profile of customer needs, map it to the business response scenario, and use the manufacturing knowledge graph to generate a structured query intent for graph retrieval.

[0011] Step S2: Perform semantic retrieval and entity matching on the triples in the knowledge graph, weighted fusion to obtain a mixed retrieval score, filter the top-k triples, and perform one-hop and two-hop path expansion in the knowledge graph starting from the query entity;

[0012] Step S3: Organize and process the retrieved triples and extension paths to form an evidence package containing information such as supporting triples, extension paths, evidence roles, and business groups;

[0013] Step S4: Input the requirement profile and evidence package into the solution generation agent to generate an initial solution containing a requirement summary, knowledge basis, response steps, risk description, etc.

[0014] Step S5: Perform structured analysis and verification on the initial solution. The evaluation agent performs multi-dimensional evaluation. If it fails the inspection, the refinement agent will revise it. This process is repeated until it passes or the iteration limit is reached.

[0015] Step S6: Integrate employee reviews and customer feedback, transform the feedback structured intelligent agent into structured human feedback, and trigger the feedback revision intelligent agent to revise the solution when necessary, and output the final solution.

[0016] Furthermore, step S1 includes the following steps:

[0017] Step S101: Obtain the structured requirement profile corresponding to the customer's requirements. The structured requirement profile includes at least the requirement identifier, requirement category, original requirement description, standardized requirement description, key attributes, requirement intent, explicit constraints, implicit requirements, multimodal evidence description, missing information, potential knowledge matching entity, and downstream query.

[0018] Step S102: Based on the requirement category, call the predefined requirement category-response scenario mapping set to map the customer requirement to the corresponding business response scenario;

[0019] Step S103: Extract candidate text from the standardized requirement description, key attributes, potential knowledge matching entities, downstream queries and original requirement description, and match the candidate text with the entity vocabulary of the manufacturing knowledge graph to obtain a candidate entity set.

[0020] Step S104: Extract a candidate relation set from the relation vocabulary of the knowledge graph.

[0021] Step S105: Input the demand profile, business response scenario, candidate entity set and candidate relationship set into the query agent to generate a structured query intent including query entity, query relationship, query attribute, missing slot, retrieval target and retrieval text.

[0022] Furthermore, step S2 includes the following steps:

[0023] Step S201: Encode the retrieval text in the structured query intent and the triplet text of each knowledge triplet in the knowledge graph, and calculate the semantic relevance score based on the cosine similarity between normalized vectors.

[0024] Step S202: Perform exact or partial matching between the query entity in the structured query intent and the head and tail entities of each knowledge triple to obtain the entity matching score;

[0025] Step S203: The semantic relevance score and the entity matching score are weighted and fused to obtain the mixed retrieval score, and the top-k manufacturing knowledge triples are retained according to the mixed retrieval score.

[0026] Step S204: Starting from the query entity in the structured query intent, the path is extended in the directed adjacency structure of the knowledge graph to obtain the one-hop path and the two-hop path.

[0027] Step S205: Perform path length constraints and duplicate path removal on the extended paths, and set a path identifier for each extended path.

[0028] Furthermore, step S3 includes the following steps:

[0029] Step S301: Based on the unique identifier of the knowledge triple and whether the head entity, relation, and tail entity are completely consistent, deduplication is performed on the Top-k manufacturing knowledge triples.

[0030] Step S302: Sort the deduplicated manufacturing knowledge triples according to the mixed retrieval score and query entity hit rate, and prioritize triples of different relation types to reduce evidence redundancy caused by excessive concentration of a single relation.

[0031] Step S303: Based on the role of the manufacturing knowledge triplet in the response scheme generation, label it with evidence roles;

[0032] Step S304: Based on business knowledge types such as product, material, performance, process, equipment, quality, defect, delivery, logistics, organization, and supply, perform business grouping of supporting triples and extension paths;

[0033] Step S305: Assemble the structured query intent, supporting triples, extended paths, evidence grouping results, retrieval summary, and evidence quality records into a structured knowledge evidence package, wherein each supporting triple stores triple_id, and each extended path stores path_id.

[0034] Furthermore, step S4 includes the following steps:

[0035] Step S401: Extract task information for solution generation from the requirement profile and structured knowledge evidence package. The task information includes at least requirement semantics, key attributes, explicit constraints, implicit requirements, missing information, supporting triples, extension paths, evidence roles, and evidence quality records.

[0036] Step S402: Use the solution generation agent to generate an initial manufacturing customer demand response solution. The initial manufacturing customer demand response solution includes at least a demand summary, knowledge basis, response steps, information gap, risk description, follow-up tasks, and traceability records.

[0037] Step S403: Bind the knowledge-based judgment to the corresponding triple_id or path_id, and bind the response step to the corresponding knowledge evidence, requirement profile, or missing information;

[0038] Step S404: Perform JSON parsing, predefined field completion, nested structure normalization, and response context alignment on the original output of the solution-generating agent;

[0039] Step S405: Verify whether the requirement identifier, requirement category, and response scenario in the response plan are consistent with the requirement profile, and verify whether the triple_id and path_id referenced in the plan exist in the current structured knowledge evidence package.

[0040] Furthermore, in step S5, the evaluation agent evaluates the current response plan from seven dimensions: demand alignment, knowledge support, structural integrity, business executability, illusion risk control, traceability, and response scenario consistency. It also outputs evaluation scores for each dimension, evaluation descriptions for each dimension, weak dimensions, main problems, modification suggestions, unsupported statements, and evidence usage issues.

[0041] Furthermore, the acceptance check includes a comprehensive evaluation score check, a single-dimensional evaluation score check, and a deterministic evidence check; the deterministic evidence check is used to check for invalid triplet identifiers, invalid path identifiers, response steps without bound evidence, and unsupported claims; when the comprehensive evaluation score is lower than the comprehensive evaluation acceptance threshold, any single-dimensional evaluation score is lower than the single-dimensional evaluation acceptance threshold, or the deterministic evidence check result is not empty, the current response scheme enters the refinement agent for revision.

[0042] Furthermore, step S6 includes the following steps:

[0043] Step S601: Obtain enterprise employee review information and customer feedback information. The enterprise employee review information includes at least supplementary business information, feasibility opinions, mandatory constraints, risk corrections, suggested adjustments, and approval status. The customer feedback information includes at least confirmed requirements, changed requirements, added constraints, rejection suggestions, and preferred directions.

[0044] Step S602: Using the feedback structured intelligent agent, extract only the information explicitly provided in the human feedback to generate a judgment result of structured employee review, structured customer feedback, feedback change type, revision goal, and whether the solution needs to be revised;

[0045] Step S603: When the employee approval status is approved and neither the employee nor the customer has raised any substantial changes, the automatic optimization response plan will be used as the final manufacturing customer demand response plan.

[0046] Step S604: When an employee requests a modification to the plan, or when the feedback includes new information, changes in constraints, corrections to risks, changes in customer preferences, or rejection of suggestions, the feedback revision agent is triggered.

[0047] Step S605: The feedback revision agent revises the solution based on the demand profile, structured knowledge evidence package, automatic optimization response plan and structured human feedback, and uses feedback_id to establish a traceability relationship between human feedback and solution changes.

[0048] Furthermore, the query agent, solution generation agent, evaluation agent, refinement agent, feedback structure agent, and feedback revision agent all employ a large language model prompting strategy to execute their respective tasks. The prompting strategy includes identity positioning, task description, precautions, and output specifications, and injects variable fields such as manufacturing domain name, demand profile, response scenario, candidate entity, candidate relationship, knowledge evidence package, current response solution, evaluation criteria, evaluation example, human feedback, structured feedback, and output structure.

[0049] This invention also discloses a manufacturing customer demand response solution generation system for executing the manufacturing customer demand response solution generation method, including a demand profile acquisition and response scenario mapping module, a structured query intent generation module, a hybrid knowledge retrieval module, a graph path expansion module, a knowledge evidence package construction module, an initial response solution generation module, a solution structure verification module, a multi-agent evaluation and refinement module, a human-in-the-loop feedback structuring and revision module, and a solution traceability and output module.

[0050] The advantages and beneficial effects of this invention are as follows:

[0051] First, this invention transforms non-standard customer needs into structured query intents, creating a stable association between customer expressions, need categories, business response scenarios, and manufacturing knowledge retrieval targets, thereby improving the controllability of need understanding.

[0052] Second, this invention integrates semantic retrieval, entity matching, and graph path expansion, which can simultaneously obtain directly related manufacturing knowledge triples and adjacent knowledge paths, thereby improving the accuracy and coverage of demand-related manufacturing knowledge acquisition.

[0053] Third, this invention organizes discrete triples and graph paths into a structured knowledge evidence package, and adds evidence role labeling, business knowledge grouping, retrieval summary and quality record, thereby reducing knowledge noise and the risk of knowledge misuse.

[0054] Fourth, this invention introduces triple_id, path_id, and feedback_id into the response scheme to establish a traceable relationship between knowledge basis, response steps, and human feedback.

[0055] Fifth, this invention improves the alignment of response solutions with requirements, the support of knowledge, the feasibility of business operations, and the risk control capabilities by forming an evaluation-revision closed loop through seven-dimensional evaluation, deterministic evidence inspection, and multiple rounds of refinement.

[0056] Sixth, this invention preserves the responsibility boundaries of enterprise employees and customers in the confirmation of production capacity, process feasibility, quality requirements, risk boundaries, demand priorities, and new constraints, so that the automatically generated solutions can be output after business review. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the overall process of generating a customer demand response solution in an embodiment of the present invention.

[0058] Figure 2 This is a flowchart of the structured query intent generation process in an embodiment of the present invention.

[0059] Figure 3 This is a flowchart of hybrid knowledge retrieval and graph path expansion in an embodiment of the present invention.

[0060] Figure 4 This is a flowchart illustrating the construction process of a structured knowledge evidence package in an embodiment of the present invention.

[0061] Figure 5 This is a flowchart of the initial response scheme generation and verification process in an embodiment of the present invention.

[0062] Figure 6 This is a flowchart of the multi-agent evaluation and iterative optimization process in an embodiment of the present invention.

[0063] Figure 7 This is a flowchart of employee review and customer feedback revision in an embodiment of the present invention.

[0064] Figure 8 This is a system execution flowchart in an embodiment of the present invention. Detailed Implementation

[0065] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0066] A method for generating manufacturing customer demand response solutions based on knowledge graph evidence enhancement and multi-agent iterative optimization includes the following steps:

[0067] Step S1: Obtain the structured demand profile and manufacturing knowledge graph corresponding to customer needs. Map customer needs to predefined business response scenarios according to the demand categories in the demand profile. Generate a structured query intent for manufacturing knowledge graph retrieval based on the demand profile, the entity terminology and relation terminology of the manufacturing knowledge graph.

[0068] Step S101: Obtain the structured requirement profile corresponding to the customer's requirements. The structured requirement profile includes at least the requirement identifier, requirement category, original requirement description, standardized requirement description, key attributes, requirement intent, explicit constraints, implicit requirements, multimodal evidence description, missing information, potential knowledge matching entity, and downstream query.

[0069] The structured requirement profile is represented as follows:

[0070]

[0071] In the formula, Indicates the first A customer needs profile; Indicates the demand identifier; Indicates the category of demand; Represents semantic information of demand; Indicates demand constraints; This represents a description of multimodal evidence. Indicates the set of missing information; This indicates guidance information for subsequent knowledge retrieval.

[0072] Step S102: Based on the requirement category, call the predefined requirement category-response scenario mapping set to map the customer requirement to the corresponding business response scenario;

[0073] The mapping between demand categories and response scenarios is represented as follows:

[0074]

[0075] In the formula, Indicates the first Response scenarios corresponding to each customer requirement instance. This represents a predefined requirement category-response scenario mapping function; This represents a set of mappings between demand categories and response scenarios.

[0076] Step S103: Extract candidate text from the standardized requirement description, key attributes, potential knowledge matching entities, downstream queries and original requirement description, and match the candidate text with the entity vocabulary of the manufacturing knowledge graph to obtain a candidate entity set.

[0077] Step S104: Extract a candidate relation set from the relation vocabulary of the knowledge graph.

[0078] Step S105: Input the demand profile, business response scenario, candidate entity set and candidate relationship set into the query agent to generate a structured query intent including query entity, query relationship, query attribute, missing slot, retrieval target and retrieval text.

[0079] The intent of a structured query is represented as:

[0080]

[0081] In the formula, Indicates the query to a set of entities. Indicates a query set of relations. This indicates a set of query attributes. This indicates a slot for information to be added. This indicates the business problem that knowledge retrieval needs to solve. This indicates the text being retrieved.

[0082] Step S2: Based on the structured query intent, perform semantic retrieval and entity matching on the knowledge triples in the manufacturing knowledge graph, obtaining semantic relevance scores and entity matching scores respectively. These two scores are then weighted and fused to obtain a mixed retrieval score. A preset number of manufacturing knowledge triples are selected based on the mixed retrieval score. Starting from the query entity in the structured query intent, one-hop and two-hop graph path expansion is performed in the adjacency structure of the manufacturing knowledge graph to obtain an expanded path set.

[0083] Step S201: Encode the retrieval text in the structured query intent and the triplet text of each knowledge triplet in the knowledge graph, and calculate the semantic relevance score based on the cosine similarity between normalized vectors.

[0084] The first in creating knowledge graphs A knowledge triple is represented as ,in Indicates the head entity. Indicates a relationship. This indicates the tail entity.

[0085] Step S202: Perform exact or partial matching between the query entity in the structured query intent and the head and tail entities of each knowledge triple to obtain the entity matching score;

[0086] The mixed retrieval score between structured query intent and knowledge triples is represented as follows:

[0087]

[0088] In the formula, sem(·) represents the semantic relevance score, ent(·) represents the entity matching score, α represents the semantic relevance score weight, and 1-α represents the entity matching score weight.

[0089] Step S203: The semantic relevance score and the entity matching score are weighted and fused to obtain the mixed retrieval score, and the top-k manufacturing knowledge triples are retained according to the mixed retrieval score.

[0090] The set of top-k manufacturing knowledge triples is represented as follows:

[0091]

[0092] Step S204: Starting from the query entity in the structured query intent, the path is extended in the directed adjacency structure of the knowledge graph to obtain the one-hop path and the two-hop path.

[0093] The extended set of one-hop and two-hop paths based on the query entity is represented as follows:

[0094]

[0095] In the formula, Indicates the first The set of extended paths corresponding to each requirement instance Indicates the starting point of the path. Indicates the path length. This indicates the creation of a knowledge graph.

[0096] Step S205: Perform path length constraints and duplicate path removal on the extended paths, and set a path identifier for each extended path.

[0097] Step S3: Perform evidence processing and organization on the selected manufacturing knowledge triples and extension paths to obtain a structured knowledge evidence package.

[0098] Step S301: Based on the unique identifier of the knowledge triple and whether the head entity, relation, and tail entity are completely consistent, deduplication is performed on the Top-k manufacturing knowledge triples.

[0099] Step S302: Sort the deduplicated manufacturing knowledge triples according to the mixed retrieval score and query entity hit rate, and prioritize triples of different relation types to reduce evidence redundancy caused by excessive concentration of a single relation.

[0100] Step S303: Based on the role of the manufacturing knowledge triplet in the response scheme generation, label it with evidence roles;

[0101] Step S304: Based on business knowledge types such as product, material, performance, process, equipment, quality, defect, delivery, logistics, organization, and supply, perform business grouping of supporting triples and extension paths;

[0102] Step S305: Assemble the structured query intent, supporting triples, extended paths, evidence grouping results, retrieval summary, and evidence quality records into a structured knowledge evidence package, wherein each supporting triple stores triple_id, and each extended path stores path_id.

[0103] The structured knowledge evidence package is represented as follows:

[0104]

[0105] In the formula, Indicates the set of supporting triples. Represents the set of extended paths. Indicates the results of evidence grouping. Indicates the search summary. This indicates the quality of evidence.

[0106] Step S4: Use the solution generation agent to generate an initial manufacturing customer demand response solution that includes knowledge basis, response steps, information gaps, risk descriptions, and traceability records; perform structured parsing, field normalization, context consistency alignment, and deterministic evidence verification on the initial solution to obtain a normalized initial response solution;

[0107] Step S401: Extract task information for solution generation from the requirement profile and structured knowledge evidence package. The task information includes at least requirement semantics, key attributes, explicit constraints, implicit requirements, missing information, supporting triples, extension paths, evidence roles, and evidence quality records.

[0108] Step S402: Use the solution generation agent to generate an initial manufacturing customer demand response solution. The initial manufacturing customer demand response solution includes at least a demand summary, knowledge basis, response steps, information gap, risk description, follow-up tasks, and traceability records.

[0109] The seven-dimensional comprehensive evaluation score of the response plan in round t is expressed as:

[0110]

[0111] In the formula, Represents the set of evaluation dimensions. The evaluation agent is represented in the 1st... Round of 1 The first requirement instance The evaluation scores are given for each dimension.

[0112] Step S403: Bind the knowledge-based judgment to the corresponding triple_id or path_id, and bind the response step to the corresponding knowledge evidence, requirement profile, or missing information;

[0113] Step S404: Perform JSON parsing, predefined field completion, nested structure normalization, and response context alignment on the original output of the solution-generating agent;

[0114] Step S405: Verify whether the requirement identifier, requirement category, and response scenario in the response plan are consistent with the requirement profile, and verify whether the triple_id and path_id referenced in the plan exist in the current structured knowledge evidence package.

[0115] Step S5: Utilize the evaluation agent and refinement agent to execute a multi-round evaluation-revision closed loop to obtain an automatically optimized response plan. The evaluation agent evaluates the current response plan from seven dimensions: requirement alignment, knowledge support, structural integrity, business executability, illusion risk control, traceability, and consistency with the response scenario. It outputs evaluation scores for each dimension, evaluation explanations for each dimension, weak dimensions, main problems, modification suggestions, unsupported statements, and evidence usage issues.

[0116] The acceptance check includes a comprehensive evaluation score check, a single-dimensional evaluation score check, and a deterministic evidence check. The deterministic evidence check is used to check for invalid triplet identifiers, invalid path identifiers, response steps without bound evidence, and unsupported claims. When the comprehensive evaluation score is lower than the comprehensive evaluation acceptance threshold, any single-dimensional evaluation score is lower than the single-dimensional evaluation acceptance threshold, or the deterministic evidence check result is not empty, the current response scheme enters the refinement agent for revision.

[0117] Accepting an inspection is indicated as follows:

[0118]

[0119] In the formula, This indicates the overall evaluation acceptance threshold. This indicates the acceptance threshold for a single-dimensional evaluation. This refers to the set of issues consisting of unsupported statements, invalid evidence citations, and unbound evidence steps.

[0120] Step S6: Integrate employee reviews and customer feedback using feedback structured agents and feedback revision agents to obtain the final manufacturing customer demand response plan.

[0121] Step S601: Obtain enterprise employee review information and customer feedback information. The enterprise employee review information includes at least supplementary business information, feasibility opinions, mandatory constraints, risk corrections, suggested adjustments, and approval status. The customer feedback information includes at least confirmed requirements, changed requirements, added constraints, rejection suggestions, and preferred directions.

[0122] Step S602: Using the feedback structured intelligent agent, extract only the information explicitly provided in the human feedback to generate a judgment result of structured employee review, structured customer feedback, feedback change type, revision goal, and whether the solution needs to be revised;

[0123] Step S603: When the employee approval status is approved and neither the employee nor the customer has raised any substantial changes, the automatic optimization response plan will be used as the final manufacturing customer demand response plan.

[0124] Step S604: When an employee requests a modification to the plan, or when the feedback includes new information, changes in constraints, corrections to risks, changes in customer preferences, or rejection of suggestions, the feedback revision agent is triggered.

[0125] Step S605: The feedback revision agent revises the solution based on the demand profile, structured knowledge evidence package, automatic optimization response plan and structured human feedback, and uses feedback_id to establish a traceability relationship between human feedback and solution changes.

[0126] Human-in-the-loop feedback revision is expressed as:

[0127]

[0128] In the formula, This indicates an automatic optimization response plan. This represents structured human feedback. This indicates the final manufacturing solution to meet customer needs.

[0129] The query agent, solution generation agent, evaluation agent, refinement agent, feedback structure agent, and feedback revision agent all employ a large language model prompting strategy to perform their respective tasks. This large language model prompting strategy includes identity positioning, task description, precautions, and output specifications, and injects variable fields such as manufacturing domain name, demand profile, response scenario, candidate entity, candidate relationship, knowledge evidence package, current response solution, evaluation criteria, evaluation example, human feedback, structured feedback, and output structure.

[0130] The automatic iteration termination condition is expressed as:

[0131] .

[0132] This invention also provides a manufacturing customer demand response solution generation system based on knowledge graph evidence enhancement and multi-agent iterative optimization, including a demand profile acquisition and response scenario mapping module, a structured query intent generation module, a hybrid knowledge retrieval module, a graph path expansion module, a knowledge evidence package construction module, an initial response solution generation module, a solution structure verification module, a multi-agent evaluation and refinement module, a human-in-the-loop feedback structuring and revision module, and a solution traceability and output module.

[0133] like Figure 1 As shown, this embodiment provides a method for generating manufacturing customer demand response solutions based on knowledge graph evidence enhancement and multi-agent iterative optimization. This method takes a structured demand profile and a manufacturing knowledge graph as input, and sequentially completes the generation of structured query intent, hybrid knowledge retrieval and graph path expansion, construction of structured knowledge evidence packages, generation of initial response solutions, multi-agent iterative optimization, and human-in-the-loop feedback revision, ultimately outputting a manufacturing customer demand response solution with traceable evidence and business verifiability.

[0134] In one specific embodiment, a customer demand response scenario in a steel manufacturing enterprise is used as an example for illustration. It should be noted that this invention is not limited to the steel manufacturing field, but can also be applied to equipment manufacturing, process manufacturing, discrete manufacturing, and other manufacturing scenarios with a manufacturing knowledge graph and customer demand response business foundation. In this embodiment, demand categories, response scenarios, manufacturing knowledge graph entity lexicons, relational lexicons, retrieval parameters, maximum graph path length, evaluation threshold, maximum iteration rounds, and response scheme output structure are pre-configured.

[0135] Table 1: Components and Functions of Demand Profile

[0136]

[0137] Table 2: Mapping of Demand Categories and Response Scenarios

[0138]

[0139] Step S1: Obtaining the demand profile and generating structured query intent. For example... Figure 2 As shown, the system first reads the structured requirement profile. The requirement profile includes at least requirement identifiers, requirement categories, requirement semantics, requirement boundaries, multimodal evidence, information gaps, and retrieval guidance information. The components of the requirement profile are shown in Table 1. Based on the requirement category, the system calls a predefined requirement category-response scenario mapping set, mapping customer requirements such as product selection, material performance, specification customization, process or surface treatment, quality improvement, delivery cycle, after-sales or technical services to corresponding response scenarios. The mapping relationship between requirement categories and response scenarios is shown in Table 2.

[0140] To reduce the open inference space of large language models and minimize the introduction of irrelevant entities, the system constructs a candidate entity set and a candidate relation set before invoking the query agent. Candidate entities are primarily derived from normalized requirement descriptions, key attributes, potential knowledge-matching entities, downstream queries, and original requirement text, and are matched and deduplicated against the entity vocabulary of the manufactured knowledge graph. The candidate relation set is directly extracted from the relation vocabulary of the manufactured knowledge graph.

[0141] The identity positioning content of the query agent prompt strategy includes: You are a domain expert used for manufacturing customer demand analysis and knowledge graph retrieval, and you are familiar with the demand semantics of the {Field_name} manufacturing domain and the entity, attribute and relation structure in the manufacturing knowledge graph.

[0142] The task description of the query agent prompting strategy includes: generating a structured query intent for manufacturing knowledge graph retrieval based on the demand profile {Demand_profile}, fixed demand category {Demand_category}, response scenario {Response_scenario}, candidate entities {Entity_candidates}, and candidate relations {Relation_candidates}.

[0143] The precautions for query agent suggestion strategies include: First, fixed demand categories must not be re-identified or modified; second, product brands, specifications, prices, delivery dates, customer names, or contract numbers must not be fabricated; third, query relationships should be selected from the candidate relationship set first; and fourth, the search text should be concise and suitable for direct knowledge retrieval.

[0144] The output specification of the query agent suggestion strategy includes:

[0145] {

[0146] "profile_id": "Requirement identifier",

[0147] "demand_category": "demand category",

[0148] "response_scenario": "response scenario",

[0149] "query_entities": ["query entities"],

[0150] "query_relations": ["query relations"],

[0151] "query_attributes": ["query attributes"],

[0152] "missing_slots": ["Information to be supplemented"],

[0153] "retrieval_intent": "retrieval target",

[0154] "retrieval_query": "Retrieve text"

[0155] }

[0156] Step S2: Demand-driven hybrid knowledge retrieval. For example... Figure 3 As shown, the system performs triple semantic retrieval and entity matching based on the structured query intent. Triple semantic retrieval measures the semantic relevance between the retrieved text and the triple text of the manufacturing knowledge graph; entity matching determines whether the query entity matches the head or tail entity of the triple. The two scores are weighted and fused to form a hybrid retrieval score, and the Top-k manufacturing knowledge triples are selected according to this score.

[0157] Step S3: Expanding Knowledge Graph Paths. A single triple typically only expresses local facts and struggles to represent the relationships between knowledge elements such as products, materials, performance, processes, quality, and delivery. Therefore, the system starts with the query entity and performs one-hop and two-hop path expansion within the directed adjacency structure of the knowledge graph, generating a path_id for each path. The expanded paths do not replace the Top-k triples but are used to provide local graph context for subsequent solution generation.

[0158] Step S4: Constructing a structured knowledge evidence package. For example... Figure 4 As shown, after acquiring candidate knowledge, the system further constructs a structured knowledge evidence package. This process includes two parts: evidence preparation and evidence organization. Evidence preparation sequentially completes triple deduplication and sorting, relation diversity selection, and evidence role labeling; evidence organization combines supporting triples, structured query intent, and extension paths to complete business grouping, evidence package assembly, and quality estimation. The composition of the structured knowledge evidence package is shown in Table 3.

[0159] Table 3: Composition and Functions of Structured Knowledge Evidence Packages

[0160]

[0161] During evidence processing, the system deduplicates evidence based on triple identifiers and whether the head entity, relation, and tail entity are completely identical. Subsequently, it sorts evidence according to mixed retrieval scores and query entity hit rates, prioritizing triples of different relation types to reduce evidence redundancy. Evidence roles can include supporting requirements, material properties, process solutions, quality requirements, delivery coordination, risk boundaries, and subsequent tasks. Business knowledge types can include product knowledge, material knowledge, attribute performance knowledge, process equipment knowledge, quality defect knowledge, delivery logistics knowledge, organizational supply knowledge, and general descriptive knowledge.

[0162] Step S5: Initial manufacturing customer requirement response solution generation. For example... Figure 5 As shown, the structured knowledge evidence package provides traceable factual evidence for demand response, but discrete evidence still needs to be further transformed into response content oriented towards business execution. The solution generation agent first extracts task information from the demand profile and knowledge evidence package, retaining demand semantics, key attributes, explicit constraints, implicit requirements, missing information, and multimodal evidence descriptions; the knowledge evidence package retains the response scenario, retrieval intent, demand anchor points, supporting triples, extension paths, retrieval summaries, and evidence quality records.

[0163] The response plan generated by the solution generation agent includes a requirements summary, knowledge basis, response steps, information gaps, risk description, follow-up tasks, and traceability records. The components of the plan are shown in Table 4.

[0164] Table 4: Components and Functions of the Response Scheme

[0165]

[0166] The identity positioning content of the solution generation intelligent agent prompt strategy includes: You are a domain expert for generating manufacturing business response solutions, and you are familiar with the interpretation of customer needs, the use of manufacturing knowledge evidence, and cross-departmental response plans in the {Field_name} manufacturing domain.

[0167] The task description of the intelligent agent prompting strategy for solution generation includes: generating a structured initial business response solution based on the demand profile {Demand_profile}, the structured knowledge evidence package {Evidence_package}, and the output structure {Output_schema}.

[0168] The precautions for the intelligent agent prompting strategy in the solution generation include: First, retain fixed demand categories and response scenarios; second, prioritize the use of supporting triples and extended paths; third, bind knowledge-based judgments and response steps to evidence identifiers; fourth, write unconfirmed information into information gaps or risk statements, and do not output it as a definite fact; fifth, do not fabricate delivery dates, inventory, prices, and contract information.

[0169] The output specifications for the agent prompting strategy generated by the solution include:

[0170] {

[0171] "profile_id": "Requirement identifier",

[0172] "demand_category": "demand category",

[0173] "response_scenario": "response scenario",

[0174] "demand_summary": "Demand Summary",

[0175] "knowledge_basis": [{"statement":"knowledge basis", "supporting_triples":[{"triple_id":"..."}], "supporting_paths":["path_id"]}],

[0176] "response_plan": [{"step":1, "action":"business action", "purpose":"purpose of the step", "evidence":[{"type":"triple / path / profile", "id":"..."}]}],

[0177] "missing_information": ["Information to be supplemented"],

[0178] "risk_notes": ["Risk Statements"],

[0179] "downstream_tasks": ["subsequent tasks"],

[0180] "traceability": {"used_triple_ids":["..."], "used_path_ids":["..."], "unsupported_claims":["..."]}

[0181] }

[0182] Step S6: Initial Solution Parsing and Deterministic Verification. Due to potential issues such as missing fields, inconsistent nesting structures, or rewritten response context in the original output of the large language model, the system sequentially performs JSON parsing, structure normalization, and context alignment operations on the original output. Context alignment uses trusted input as the benchmark, performing consistency calibration on requirement identifiers, requirement categories, and response scenarios. Subsequently, the system verifies whether the response steps are bound to knowledge evidence, whether the triple_id and path_id in the solution traceability record originate from the current knowledge evidence package, and whether the solution contains unfounded deterministic business information.

[0183] Step S7: Multi-agent evaluation and iterative optimization. For example... Figure 6 As shown, the system inputs the current response plan, requirement profile, and structured knowledge evidence package into the evaluation agent. The evaluation agent comprehensively evaluates the current plan from seven dimensions: requirement alignment, knowledge support, structural integrity, business executability, illusion risk control, traceability, and consistency with the response scenario. The seven-dimensional evaluation criteria are shown in Table 5.

[0184] Table 5: Evaluation Criteria for Seven-Dimensional Solutions

[0185]

[0186] The identity positioning content of the evaluation agent prompt strategy includes: You are a manufacturing customer demand response solution evaluation expert, familiar with demand interpretation, manufacturing knowledge use, business response process, and risks of large language model generation.

[0187] The task description for evaluating the agent's prompting strategy includes: performing a multi-dimensional evaluation based on the demand profile, the structured knowledge evidence package, the current response plan, and the evaluation rubric, and providing actionable modification suggestions.

[0188] The key considerations for evaluating agent prompting strategies include: first, evaluating the current solution rather than generating a new one; second, evaluating independently according to all dimensions in the evaluation criteria; third, verifying the consistency of evidence identification, step-level evidence binding, unsubstantiated claims, and response scenarios; and fourth, identifying specific information that lacks supporting evidence and providing modification suggestions for weak dimensions.

[0189] The output specification for evaluating agent prompting strategies includes:

[0190] {

[0191] "dimension_scores": {

[0192] "demand_alignment": "1.0-5.0",

[0193] "knowledge_grounding": "1.0-5.0",

[0194] "structural_completeness": "1.0-5.0",

[0195] "business_executability": "1.0-5.0",

[0196] "hallucination_risk_control": "1.0-5.0",

[0197] "traceability": "1.0-5.0",

[0198] "scenario_consistency": "1.0-5.0"

[0199] },

[0200] "weak dimensions": ["weak dimensions"],

[0201] "major_problems": ["Main problems"],

[0202] "revision_feedback": ["Revision suggestions"],

[0203] "unsupported_claims": ["Unsupported claims"],

[0204] "evidence_usage_issues": ["Issues related to the use of evidence"]

[0205] }

[0206] After the evaluation is completed, the system further performs a deterministic evidence check. This check verifies the existence of triple_id and path_id referenced in the response scheme in the current knowledge evidence package, identifies response steps without bound evidence, and merges unsupported claims from the scheme traceability record into the evaluation result. When the overall evaluation score is lower than the overall evaluation acceptance threshold, any single-dimensional evaluation score is lower than the single-dimensional evaluation acceptance threshold, or the deterministic evidence check finds problems, the current scheme enters the refinement agent for revision.

[0207] The identity positioning content of the refined intelligent agent prompt strategy includes: You are a manufacturing customer demand response solution refinement expert who can improve solution structure, evidence use and business steps based on evaluation results.

[0208] The task description for refining the agent's prompting strategy includes: revising the current response plan based on the demand profile, the structured knowledge evidence package, the current response plan, and the evaluation result.

[0209] The key considerations for refining the agent's prompting strategy include: 1) addressing weaknesses, key issues, and modification suggestions identified in the evaluation results; 2) retaining reasonable and evidence-supported content; 3) using only the triple_id and path_id present in the knowledge evidence package; 4) binding each response step to MKG evidence, requirement profiles, or missing information; 5) deleting or weakening unfounded conclusions; and 6) recording unresolved information in the information gap or risk statement.

[0210] Step S8: Employee review and customer feedback revision. For example... Figure 7 As shown, automatic optimization improves the consistency of requirements, the strength of evidence, and the quality of the structure of a solution, but it cannot completely replace the confirmation of actual capacity, business responsibilities, and user intent by company personnel. Therefore, after obtaining an automatically optimized solution, the system incorporates employee review and customer feedback for ongoing revision. Company employees mainly supplement missing business information, evaluate the feasibility of the solution, and correct risk boundaries and necessary constraints; customer feedback is mainly used to confirm or change requirements, add constraints, reject suggestions that do not meet expectations, and clarify the expected response direction.

[0211] The feedback structured intelligent agent prompt strategy's identity positioning content includes: You are a domain expert used to analyze manufacturing customer demand response solutions, employee review, and customer feedback.

[0212] The task description of the feedback structured intelligent agent prompting strategy includes: referring to the automatically optimized response plan {Optimized_response_plan}, structuring the human feedback {Human_feedback}, without directly revising the plan.

[0213] The key considerations for feedback structured intelligent agent prompting strategies include: first, retaining the distinct sources of employee and customer feedback; second, extracting only explicitly provided new information, changed requirements, new constraints, risk corrections, rejection suggestions, and response preferences; third, not inferring information not explicitly provided in human feedback; and fourth, determining whether feedback actually warrants triggering solution revisions.

[0214] The output specification of the feedback structured agent prompting strategy includes:

[0215] {

[0216] "profile_id": "Requirement identifier",

[0217] "feedback_id": "feedback identifier",

[0218] "structured_employee_review": {

[0219] "reviewer_role": "Reviewer role",

[0220] "filled_missing_information": ["Supplementary information"],

[0221] "feasibility_comments": ["feasibility comments"],

[0222] "mandatory_constraints": ["mandatory constraints"],

[0223] "risk_corrections": ["risk correction"],

[0224] "recommended_adjustments": ["Suggested adjustments"],

[0225] "approval_status": "approved / revision_required / rejected / pending"

[0226] },

[0227] "structured_customer_feedback": {

[0228] "confirmed_requirements": ["Confirmed requirements"],

[0229] "changed_requirements": ["Changed requirements"],

[0230] "additional_constraints": ["Added constraints"],

[0231] "rejected_suggestions": ["Rejected suggestions"],

[0232] "preferred_response_direction": "preferred direction"

[0233] },

[0234] "need_plan_revision": true

[0235] }

[0236] When an employee has approved a solution and neither the employee nor the client has proposed any substantial changes, the system retains the automatically optimized solution as the final response to the client's needs. When an employee requests modifications to the solution, or when the feedback includes new information, changes in constraints, corrections to risks, or changes in user preferences, the feedback revision agent is triggered.

[0237] The feedback and revision of the agent's prompts include the following identity positioning information: You are a human-in-the-loop revision expert who manufactures customer demand response solutions.

[0238] The task description for revising the intelligent agent's prompting strategy includes: revising the response plan based on the demand profile, the structured knowledge evidence package, the optimized response plan, and the structured human feedback.

[0239] The key considerations for implementing feedback-driven intelligent agent prompts include: 1. Only execute revisions requested by employees or based on customer feedback; 2. Utilize supplementary information from employees to fill information gaps, correct risk boundaries, and improve response steps; 3. Use customer feedback to update priorities and constraints or remove rejected suggestions; 4. Do not fabricate knowledge evidence packages or information not provided in human feedback; 5. Bind feedback-driven changes to the corresponding feedback_id; 6. Retain unresolved information and evidence-supported content not refuted by human feedback; 7. When customer preferences conflict with mandatory constraints related to feasibility, safety, quality, or compliance confirmed by employees, retain the mandatory constraints, write the conflict into the remaining uncertainties, and request further confirmation.

[0240] Step S9: System execution. (e.g.) Figure 8As shown, this invention also provides a manufacturing customer demand response solution generation system based on knowledge graph evidence enhancement and multi-agent iterative optimization. The system includes: a demand profile acquisition and response scenario mapping module for reading demand profiles and mapping demand categories to response scenarios; a structured query intent generation module for calling query agents to generate query entities, query relationships, query attributes, missing slots, retrieval targets, and retrieval text; a hybrid knowledge retrieval module for calculating semantic relevance scores, entity matching scores, and hybrid retrieval scores; a graph path expansion module for generating one-hop and two-hop expansion paths based on the adjacency structure of the manufacturing knowledge graph; a knowledge evidence package construction module for performing triple deduplication, sorting, diversity selection, evidence role labeling, business grouping, and quality analysis; an initial response solution generation module for generating structured response solutions; a solution structure verification module for performing JSON parsing, field normalization, context alignment, and evidence identifier verification; a multi-agent evaluation and refinement module for performing seven-dimensional evaluation, deterministic evidence checking, and solution revision; a human-in-the-loop feedback structuring and revision module for integrating employee review and customer feedback; and a solution traceability and output module for outputting the final customer demand response solution and its evidence traceability record.

[0241] In this embodiment, a customer requirement related to a process or surface treatment is selected as the subject of description. The system determines the requirement category as a process response scenario, generates a structured query intent, performs hybrid knowledge retrieval and two-hop graph path expansion, then constructs a structured knowledge evidence package and generates an initial response plan. After two rounds of evaluation and revision, the initial plan is supplemented with quality acceptance and process feasibility explanations by employees, and a final revision is completed based on customer feedback regarding surface quality expectations. The main processing results of this embodiment are shown in Table 6.

[0242] Table 6: Main Processing Results of Specific Embodiments

[0243]

[0244] As shown in Table 6, this embodiment can transform a non-standard customer requirement into a searchable, structured query intent, and organize the retrieved supporting triples and extension paths into a structured knowledge evidence package. The generated response plan can clearly distinguish between knowledge base, information gaps, and risk boundaries, and completes automatic optimization and manual confirmation through evaluation agent, refinement agent, feedback structured agent, and feedback revision agent, ultimately forming a manufacturing customer requirement response plan with traceable evidence.

[0245] Furthermore, the system inputs in this embodiment include three types of data. The first type is customer demand profiles, which are mainly used to describe customer demand objects, demand categories, key attributes, explicit constraints, potential intentions, known evidence, and missing information. The second type is manufacturing knowledge graphs, which are mainly used to provide structured knowledge of products, materials, processes, equipment, performance, quality, delivery, and services within the enterprise. The third type is parameter configuration, which mainly includes the number of Top-k searches, the maximum path extension length, evidence diversity selection rules, comprehensive evaluation acceptance threshold, single-dimensional evaluation acceptance threshold, maximum number of iterations, and human feedback triggering conditions.

[0246] During the requirement profiling input phase, the system does not require customers to fill in their requirements using a fixed template. Instead, it allows the preceding requirement collection module to output a pre-structured requirement profile. For requirements that are conversational, contain illustrative images, or lack key business parameters, the system prioritizes retaining the original semantics and missing information, and does not automatically add unconfirmed content as definitive conclusions. This ensures that when generating subsequent response solutions, confirmed facts, speculative requirements, and information awaiting confirmation can be distinguished.

[0247] During the response scenario mapping phase, the system selects the corresponding response scenario based on the requirement category. This mapping is not generated openly, but rather determined by a pre-defined mapping set. This approach allows different types of requirements to enter their respective business processes, avoiding confusion in the knowledge retrieval scope caused by using a uniform response model. For example, material performance-related requirements prioritize retrieving knowledge about materials, properties, composition, and application scenarios; delivery cycle-related requirements prioritize retrieving knowledge about delivery, logistics, inventory, and production collaboration; and quality improvement-related requirements prioritize retrieving knowledge about defects, causes, process control, and quality inspection.

[0248] During the candidate entity construction phase, the system first extracts words and phrases that may be related to the manufacturing knowledge graph from the demand profile, and then matches them with the manufacturing knowledge graph entity vocabulary. Matching methods can include exact match, alias match, string inclusion match, and normalized name match. For the matched candidate entities, the system performs deduplication, length filtering, and quantity truncation to avoid inputting too many low-relevance words into the query agent.

[0249] During the candidate relation construction phase, the system does not require the query agent to freely create retrieval relations. Instead, it prioritizes selecting relations from the manufacturing knowledge graph relation lexicon. The relation lexicon can include manufacturing business relations such as belonging to, containing, applying to, requiring, affecting, improving, reducing, causing, used for, generating, inspecting, controlling, supplying, delivering, and serving. By restricting the sources of candidate relations, the probability of non-canonical relation expressions appearing in query intent parsing can be reduced.

[0250] In the hybrid knowledge retrieval stage, semantic relevance scores are used to address the semantic differences between customer expressions and knowledge graph text, while entity matching scores are used to reinforce the presence of core target objects in the search results. The weighted fusion of these two scores avoids returning triples with similar topics but lacking core entities from simple semantic retrieval, and also avoids returning a large amount of knowledge containing only identical entities but irrelevant to the current target from simple entity matching.

[0251] During the graph path expansion phase, the system uses the query entity in the structured query intent as the starting point, rather than performing global path enumeration on the complete manufactured knowledge graph. The path expansion length is preferably set to one or two hops to balance local knowledge associations and computational efficiency. The path expansion results are mainly used to supplement the local graph context, rather than directly replacing the Top-k support triples. For each edge in the path, the system retains the entity, relation, and path order, and generates a unique path_id for the complete path.

[0252] During the evidence package construction phase, both supporting triples and extended paths are assigned traceable identifiers. The metadata for supporting triples includes at least `triple_id`, head entity, relation, tail entity, triple text, retrieval score, evidence role, and knowledge source; the metadata for extended paths includes at least `path_id`, path text, path length, starting entity, ending entity, and identifiers of the constituent triples. These identifiers are preferentially referenced in the response plan whenever specific knowledge is involved.

[0253] During the evidence grouping phase, the system organizes discrete triples into different evidence groups based on evidence roles and business knowledge types. Evidence groups can include groups for demand objects, products and materials, attributes and performance, processes and equipment, quality and defects, delivery and logistics, after-sales service, and risk boundaries. This grouping reduces contextual fragmentation between discrete triples, enabling the solution-generating agent to invoke knowledge evidence based on business semantics.

[0254] During the evidence quality analysis phase, the system generates a retrieval summary and an evidence quality record. The retrieval summary describes the main retrieval objects, the number of supporting triples, the number of extension paths, and the types of business knowledge covered. The evidence quality record describes the average mixed retrieval score, query entity coverage, evidence diversity, and evidence deficiencies. If the knowledge evidence package cannot cover the key attributes or constraints in the requirements, the system passes this content to the solution generation agent and requires it to write an information gap or risk statement.

[0255] During the initial response solution generation phase, the system does not require the solution-generating agent to provide definitive business conclusions beyond the scope of evidence. For content supported by triples or paths, the solution should bind triple_id or path_id in the knowledge basis or response steps; for content lacking evidence but requiring business confirmation, the solution should include missing_information; for content with uncertainties, conflicts, or requiring careful handling, the solution should include risk_notes.

[0256] During the solution structure verification phase, the system first determines whether the model output can be parsed into a structured object; secondly, it fills in any predefined but missing fields; thirdly, it calibrates the requirement identifier, requirement category, and response scenario to prevent the model from rewriting the fixed context during generation; and finally, it checks whether the evidence identifier in the solution traceability record comes from the current knowledge evidence package. If invalid evidence identifiers, unbound evidence steps, or unsupported business conclusions are found, the system records them as verification warnings and passes them to the evaluation agent or refinement agent.

[0257] During the multi-agent evaluation phase, the evaluation agent does not directly modify the solution, but instead outputs dimensional scores, problem descriptions, and modification suggestions. This separation of roles avoids confusion between the evaluation and revision processes and facilitates the storage of evaluation records for each round. Evaluation records should include at least the evaluation round, scores for each dimension, weak dimensions, main problems, unsupported claims, evidence usage issues, and suggested modification directions.

[0258] During the solution refinement phase, the refinement agent performs bounded revisions based on the evaluation results. The revision process prioritizes addressing low-scoring dimensions, unsupported claims, missing evidence bindings, and structural incompleteness. Content supported by evidence and not identified as problematic in the evaluation should be retained. Deterministic statements that cannot be supported by knowledge evidence packages should be deleted, weakened, or moved to the risk statement. Content requiring further business confirmation should be moved to the information gap section.

[0259] During the iteration termination phase, the system considers both the automatic evaluation results and the preset iteration limit. The automatic optimization process terminates when the solution meets the comprehensive evaluation threshold, the single-dimensional evaluation threshold, and the deterministic evidence check conditions. If the maximum number of iterations has been reached but the solution is still not fully successful, the system retains the most recently successfully generated and structurally valid solution and adds unresolved issues to the solution status description and risk description. This mechanism prevents abnormal revisions from overwriting existing available solutions.

[0260] During the employee review phase, employees can supplement missing information, revise the feasibility of solutions, clarify responsibility interfaces, add mandatory constraints, or adjust risk boundaries based on actual business data. Employee reviews do not require input in a fixed format; the system extracts information that can be used for solution revision from natural language review comments through a structured intelligent agent that provides feedback. Mandatory constraints in employee feedback have higher priority than customer preference information.

[0261] During the customer feedback phase, customers can confirm their needs, change their needs, add constraints, reject certain suggestions, or specify their desired response direction. The system only extracts explicitly stated information from the customer and does not automatically infer unexpressed preferences. When customer feedback conflicts with safety, quality, compliance, or feasibility constraints confirmed by employees, the system retains the mandatory constraints and records the conflict in `remaining_uncertainties`, indicating that further manual confirmation is required.

[0262] During the final solution output phase, the system outputs a response plan confirmed by employees or customers, and simultaneously outputs traceability records. These traceability records include not only used_triple_ids and used_path_ids, but also used_feedback_ids, which indicate which parts of the final solution originated from the manufacturing knowledge graph and which came from human feedback. This mechanism allows companies to subsequently review response steps, knowledge sources, and reasons for human intervention.

[0263] The system of this invention can be deployed in enterprise resource planning systems, manufacturing execution systems, advanced planning and scheduling systems, customer relationship management systems, manufacturing knowledge management platforms, or independent intelligent customer demand response servers. The system can read manufacturing knowledge graphs, order information, product knowledge, quality records, and delivery data through interfaces, and output the final response solution to business departments such as sales, technology, quality, production, supply chain, and after-sales service.

[0264] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Those skilled in the art can modify or make equivalent substitutions to the specific embodiments of the present invention without departing from the spirit and substance of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A method for generating a customer demand response solution, characterized in that, Includes the following steps: Step S1: Based on the structured demand profile of customer needs, map it to the business response scenario, and use the manufacturing knowledge graph to generate a structured query intent for graph retrieval. Step S2: Perform semantic retrieval and entity matching on the triples in the knowledge graph, weighted fusion to obtain a mixed retrieval score, filter the top-k triples, and perform one-hop and two-hop path expansion in the knowledge graph starting from the query entity; Step S3: Organize and process the retrieved triples and extension paths to form an evidence package containing information such as supporting triples, extension paths, evidence roles, and business groups; Step S4: Input the requirement profile and evidence package into the solution generation agent to generate an initial solution containing a requirement summary, knowledge basis, response steps, risk description, etc. Step S5: Perform structured analysis and verification on the initial solution. The evaluation agent performs multi-dimensional evaluation. If it fails the inspection, the refinement agent will revise it. This process is repeated until it passes or the iteration limit is reached. Step S6: Integrate employee reviews and customer feedback, transform the feedback structured intelligent agent into structured human feedback, and trigger the feedback revision intelligent agent to revise the solution when necessary, and output the final solution.

2. The method for generating a manufacturing customer demand response solution according to claim 1, characterized in that: Step S1 includes the following steps: Step S101: Obtain the structured requirement profile corresponding to the customer's requirements. The structured requirement profile includes at least the requirement identifier, requirement category, original requirement description, standardized requirement description, key attributes, requirement intent, explicit constraints, implicit requirements, multimodal evidence description, missing information, potential knowledge matching entity, and downstream query. Step S102: Based on the requirement category, call the predefined requirement category-response scenario mapping set to map the customer requirement to the corresponding business response scenario; Step S103: Extract candidate text from the standardized requirement description, key attributes, potential knowledge matching entities, downstream queries and original requirement description, and match the candidate text with the entity vocabulary of the manufacturing knowledge graph to obtain a candidate entity set. Step S104: Extract a candidate relation set from the relation vocabulary of the knowledge graph. Step S105: Input the demand profile, business response scenario, candidate entity set and candidate relationship set into the query agent to generate a structured query intent including query entity, query relationship, query attribute, missing slot, retrieval target and retrieval text.

3. The method for generating a manufacturing customer demand response solution according to claim 1, characterized in that: Step S2 includes the following steps: Step S201: Encode the retrieval text in the structured query intent and the triplet text of each knowledge triplet in the knowledge graph, and calculate the semantic relevance score based on the cosine similarity between normalized vectors. Step S202: Perform exact or partial matching between the query entity in the structured query intent and the head and tail entities of each knowledge triple to obtain the entity matching score; Step S203: The semantic relevance score and the entity matching score are weighted and fused to obtain the mixed retrieval score, and the top-k manufacturing knowledge triples are retained according to the mixed retrieval score. Step S204: Starting from the query entity in the structured query intent, the path is extended in the directed adjacency structure of the knowledge graph to obtain the one-hop path and the two-hop path. Step S205: Perform path length constraints and duplicate path removal on the extended paths, and set a path identifier for each extended path.

4. The method for generating a manufacturing customer demand response solution according to claim 1, characterized in that: Step S3 includes the following steps: Step S301: Based on the unique identifier of the knowledge triple and whether the head entity, relation, and tail entity are completely consistent, deduplication is performed on the Top-k manufacturing knowledge triples. Step S302: Sort the deduplicated manufacturing knowledge triples according to the mixed retrieval score and query entity hit rate, and prioritize triples of different relation types to reduce evidence redundancy caused by excessive concentration of a single relation. Step S303: Based on the role of the manufacturing knowledge triplet in the response scheme generation, label it with evidence roles; Step S304: Based on business knowledge types such as product, material, performance, process, equipment, quality, defect, delivery, logistics, organization, and supply, perform business grouping of supporting triples and extension paths; Step S305: Assemble the structured query intent, supporting triples, extended paths, evidence grouping results, retrieval summary, and evidence quality records into a structured knowledge evidence package, wherein each supporting triple stores triple_id, and each extended path stores path_id.

5. The method for generating a manufacturing customer demand response solution according to claim 1, characterized in that: Step S4 includes the following steps: Step S401: Extract task information for solution generation from the requirement profile and structured knowledge evidence package. The task information includes at least requirement semantics, key attributes, explicit constraints, implicit requirements, missing information, supporting triples, extension paths, evidence roles, and evidence quality records. Step S402: Use the solution generation agent to generate an initial manufacturing customer demand response solution. The initial manufacturing customer demand response solution includes at least a demand summary, knowledge basis, response steps, information gap, risk description, follow-up tasks, and traceability records. Step S403: Bind the knowledge-based judgment to the corresponding triple_id or path_id, and bind the response step to the corresponding knowledge evidence, requirement profile, or missing information; Step S404: Perform JSON parsing, predefined field completion, nested structure normalization, and response context alignment on the original output of the solution-generating agent; Step S405: Verify whether the requirement identifier, requirement category, and response scenario in the response plan are consistent with the requirement profile, and verify whether the triple_id and path_id referenced in the plan exist in the current structured knowledge evidence package.

6. The method for generating a manufacturing customer demand response solution according to claim 1, characterized in that: In step S5, the evaluation agent evaluates the current response plan from seven dimensions: demand alignment, knowledge support, structural integrity, business executability, illusion risk control, traceability, and response scenario consistency. It then outputs the evaluation scores for each dimension, evaluation descriptions for each dimension, weak dimensions, main problems, modification suggestions, unsupported statements, and evidence usage issues.

7. The method for generating a manufacturing customer demand response solution according to claim 6, characterized in that: The acceptance check includes a comprehensive evaluation score check, a single-dimensional evaluation score check, and a deterministic evidence check. The deterministic evidence check is used to check for invalid triplet identifiers, invalid path identifiers, response steps without bound evidence, and unsupported claims. When the comprehensive evaluation score is lower than the comprehensive evaluation acceptance threshold, any single-dimensional evaluation score is lower than the single-dimensional evaluation acceptance threshold, or the deterministic evidence check result is not empty, the current response scheme enters the refinement agent for revision.

8. The method for generating a manufacturing customer demand response solution according to claim 1, characterized in that: Step S6 includes the following steps: Step S601: Obtain enterprise employee review information and customer feedback information. The enterprise employee review information includes at least supplementary business information, feasibility opinions, mandatory constraints, risk corrections, suggested adjustments, and approval status. The customer feedback information includes at least confirmed requirements, changed requirements, added constraints, rejection suggestions, and preferred directions. Step S602: Using the feedback structured intelligent agent, extract only the information explicitly provided in the human feedback to generate a judgment result of structured employee review, structured customer feedback, feedback change type, revision goal, and whether the solution needs to be revised; Step S603: When the employee approval status is approved and neither the employee nor the customer has raised any substantial changes, the automatic optimization response plan will be used as the final manufacturing customer demand response plan. Step S604: When an employee requests a modification to the plan, or when the feedback includes new information, changes in constraints, corrections to risks, changes in customer preferences, or rejection of suggestions, the feedback revision agent is triggered. Step S605: The feedback revision agent revises the solution based on the demand profile, structured knowledge evidence package, automatic optimization response plan and structured human feedback, and uses feedback_id to establish a traceability relationship between human feedback and solution changes.

9. The method for generating a manufacturing customer demand response solution according to claim 2, characterized in that: The query agent, solution generation agent, evaluation agent, refinement agent, feedback structure agent, and feedback revision agent all employ a large language model prompting strategy to execute their respective tasks. The prompting strategy includes identity positioning, task description, precautions, and output specifications, and injects manufacturing domain name, demand profile, response scenario, candidate entity, candidate relationship, knowledge evidence package, current response solution, evaluation criteria, evaluation example, human feedback, structured feedback, and output structure through variable fields.

10. A manufacturing customer demand response solution generation system, used to execute the manufacturing customer demand response solution generation method as described in any one of claims 1-9, characterized in that, It includes a demand profile acquisition and response scenario mapping module, a structured query intent generation module, a hybrid knowledge retrieval module, a graph path expansion module, a knowledge evidence package construction module, an initial response scheme generation module, a scheme structure verification module, a multi-agent evaluation and refinement module, a human-in-the-loop feedback structuring and revision module, and a scheme traceability and output module.

Images