An intelligent question and answer method for research and development services based on industry evidence and standard anchor points

Abstract

The application discloses a kind of based on industry evidence and standard anchor point R&D service intelligent question and answer method, comprising: evidence fragmenting analysis is carried out to standard, regulation and other multi-source data, and the traceable industry multimodal knowledge graph with evidence as core is constructed;Standard anchor point set is formed by abstracting from standard specification, and user multimodal input is parsed into structured design intent element;With design intent and standard anchor point as constraint, execute explicit correlation, bridge and anchor gap driven complementary association retrieval, and build evidence subgraph supporting determination;Design scheme is generated under the constraint of evidence aggregation and anchor point projection, and through anchor point consistency check and directional correction, the final output that meets or clearly labeled information insufficient is formed.The application introduces industry evidence and standard constraint into intelligent question and answer, realizes the effective completion of implicit engineering constraint and the controllable generation of key parameters, provides technical support for scheme executability, compliance and traceability in R&D design service scene.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to an intelligent question-answering method for R&D services based on industry evidence and standard anchors. Background Technology

[0002] In the product innovation design and R&D implementation process of discrete manufacturing industries such as mold making and home appliances, design solutions must simultaneously meet multiple constraints, including engineering feasibility, material and process windows, testing methods, and standard / specification clauses. Especially for SMEs, design requirements are often input in unstructured forms such as verbal descriptions, reference images, or sketches. This information commonly suffers from missing key parameter slots and implicit, unexpressed constraints, leading to repeated rework in subsequent solution derivation, parameter determination, and compliance assessment, thus reducing R&D efficiency and implementation success rates.

[0003] With the development of artificial intelligence technology, intelligent question answering based on large models is gradually being applied in product design and R&D services. This type of technology, through the understanding and generation of multimodal information such as natural language text and images, combined with resources such as enterprise knowledge bases, can assist users in conceptualizing solutions, providing technical consultation, and design reasoning to a certain extent. At the same time, the introduction of methods such as retrieval-enhanced generation enables large models to acquire information from external knowledge and participate in content generation, providing a new intelligent solution to complex design problems.

[0004] Chinese patent application CN120371993A discloses a search enhancement generation method and device. It evaluates and scores search resources from multiple perspectives across several preset dimensions. These dimensions include the relationship between the resource and the query input (e.g., relevance, usefulness, supplementation) as well as the resource's own characteristics (e.g., authority, timeliness). The multi-dimensional scores are then combined to select at least one target information resource. Finally, a large model generates a response based on the target resource, thereby improving the quality of RAG (Search Enhancement Generation) search input and the final response. This patented technology focuses on addressing the inconsistent quality of retrieved documents, which leads to unstable response quality. However, it lacks clause-level location evidence binding and determinable standard constraint expressions to impose hard constraints on the generation process, and it also fails to provide a closed-loop cross-modal consistency verification for multimodal results, resulting in insufficient engineering-level traceability and compliance. Chinese patent application CN118535984A discloses a method for measuring the semantic similarity of standard clauses in the power industry based on contrastive learning. It extracts and decomposes clause text from power industry documents, uses a RoBERTa model for sentence encoding, optimizes the clause vector representation through contrastive learning, and finally outputs a semantic similarity score for the clauses using methods such as cosine similarity. This method is used for automated clause search / comparison, assisting clause analysis, and mitigating conflicts and overlaps in standard documents. While this patented technology focuses on standard clause vectorization and similarity measurement / retrieval support, it lacks a mechanism to further translate clauses into determinable constraint / threshold rules for constraint generation output. It also does not address multimodal structural landmark constraints and cross-modal consistency verification closed-loops, making it difficult to directly support standard-driven engineering solutions.

[0005] In summary, existing large-scale intelligent question-answering technologies still have shortcomings in R&D design service scenarios. Current methods often treat design problems as a one-time, open-ended generation process, relying primarily on linguistic and visual priors for inference. They lack explicit modeling of industry evidence sources, engineering parameter constraints, and standard clauses, making them prone to professional biases or neglecting mandatory specifications during the generation process. Furthermore, even when combined with retrieval mechanisms, relevant standards and specifications are mostly used for post-generation result verification, failing to effectively constrain key parameters and compliance requirements during the generation stage. They also lack the ability to complete implicit engineering constraints in user input, and cross-text and image constraint alignment is difficult to achieve. Consequently, the generated results fail to meet the actual needs of industrial product design in terms of executability, compliance, and traceability. Summary of the Invention

[0006] In view of the above, this invention provides an intelligent question-answering method for R&D services based on industry evidence and standard anchors, aiming to solve the problems of insufficient engineering constraint modeling, difficulty in implementing generated results, and insufficient compliance and traceability in existing large-scale intelligent question-answering technologies in product R&D and design services. By constructing an industry multimodal knowledge graph with evidence fragments as the core object, and abstracting the decidable clauses in standards / specifications into standard anchors, the solution space is constrained by evidence and anchors throughout the retrieval, reasoning, and generation processes. User text, diagrams, and historical context are parsed into structured design intentions. Complementary association retrieval is triggered based on anchor coverage gaps to complete implicit constraint fields. During the generation stage, key conclusions are forcibly bound to evidence pointers, and after consistency verification, targeted supplementary retrieval and local correction are performed on conflict points, thereby enhancing the engineering applicability of intelligent question answering in product innovation and design service scenarios.

[0007] A smart question-answering method for R&D services based on industry evidence and standard anchors includes the following steps: (1) Perform evidence fragment analysis on multi-source data including standards, regulations, process specifications, testing methods and engineering cases, and construct an industry multimodal knowledge graph with evidence as the core and traceability; (2) Abstract a set of definable standard anchor points from standards and specifications, and parse user multimodal input into structured design intent elements, and explicitly label constraints, conflicts and missing information; (3) Perform explicit relevance retrieval, bridging retrieval and complementary relevance retrieval driven by anchor gap with design intent and standard anchor points as constraints, and construct evidence subgraphs to support the judgment; (4) Generate an executable design scheme under the constraints of evidence aggregation and anchor point projection, and form a final output design scheme that satisfies or clearly indicates insufficient information through anchor point consistency verification and orientation correction.

[0008] Furthermore, the specific implementation of step (1) is as follows: S11: Perform layout-aware analysis and segmentation on multi-source materials including standards, regulations, process specifications, testing methods and engineering cases. Abstract multimodal content, including text paragraphs, table units, process items and illustrations, into a minimum set of citationable evidence fragments, and generate a unique evidence pointer for each evidence fragment, including source type, document identifier, version number, location information, applicable domain and key parameters. S12: For each evidence fragment and its evidence pointer, construct text semantic representation and visual / structural semantic representation respectively, and generate a unified embedding representation through a controlled fusion function based on meta-features; S13: Identify entity mentions from evidence fragments using embedded representations and evidence pointers, map entity mentions to normalized entity nodes based on industry-controlled vocabularies and synonym mapping rules, construct candidate sets for entity mentions that cannot be uniquely determined, and generate inter-entity relationships supported by at least one evidence fragment under the constraints of a predefined relationship template. S14: Integrate entity nodes and relationships between entities into an industry multimodal knowledge graph, and establish a traceable link between graph conclusions and original evidence fragments through evidence pointers; at the same time, maintain time-sliced ​​subgraphs according to document identifiers and version numbers, and only perform incremental updates on the affected parts when the version is updated, to ensure the long-term consistency and traceability of the knowledge graph.

[0009] Furthermore, the specific implementation of step (2) is as follows: S21: Based on the knowledge graph, locate clauses containing judgment statements including thresholds, levels, prohibitions and mandatory inspection processes from standards, regulations and enterprise specifications, and generate a standard anchor point set containing applicable domains, constraint types, constraint expressions and anchor point evidence pointers through field extraction and constraint template mapping. S22: Parse the user-input text, images, and historical dialogue context respectively; extract candidate values ​​and confidence scores for each design intent slot from the text; generate an engineered structural description containing structural objects, connection relationships, size cues, and material characteristics from the image and extract corresponding candidate values; extract candidate values ​​with time markers from the historical context to form a multimodal candidate set for each slot. S23: Merge the multimodal candidate sets of each slot, calculate a comprehensive score for each candidate value, the comprehensive score is based on modal confidence weighting, historical time decay and consistency function, wherein the consistency function comprehensively evaluates the degree of matching between the candidate value and the text semantics, image structure description and standard anchor point identifiable alignment, and imposes a penalty on mutual exclusion conflicts with other candidate values; S24: Filter candidate values ​​based on slot threshold gating. When there are candidates that meet the scoring threshold, minimum interval, consistency threshold and evidence sufficiency constraints, determine the principal value of the slot and retain multimodal information in the form of principal value plus supplementary value; otherwise, mark the slot as unknown and output the supplementary information items and recommended acquisition methods. Finally, output the structured design intent elements aligned with knowledge graph fields and standard anchor points.

[0010] Furthermore, the specific implementation of step (3) is as follows: S31: Based on the structured design intent slots and image engineering structure description, construct explicit relevant queries and bridging queries, execute the retrieval to obtain a preliminary evidence set, and calculate the basic score of each piece of evidence; S32: Align the preliminary evidence set with the standard anchor set, calculate the coverage of the fields required for anchor determination by the current evidence, and when the coverage is lower than the preset threshold, identify the set of missing fields necessary to trigger anchor determination as input for complementary retrieval; S33: For the identified missing fields, perform a breadth-first search on the industry multimodal knowledge graph with limited relation types to obtain evidence that can fill in the missing fields, and merge the newly added evidence to update the evidence set, while retaining the scoring information of each piece of evidence; S34: Repeat steps S32~S33 until the coverage of the anchor decision field of the evidence set meets the threshold or the maximum number of iterations is reached, and output the final evidence set and retrieval status marker. S35: Extract evidence subgraphs from the knowledge graph based on the final evidence set, rearrange them by combining the basic score and anchor point hit items, and mark the nodes that are aligned with the standard anchor point pointers for subsequent evidence aggregation and hard anchor point filtering.

[0011] Furthermore, the specific implementation of step (4) is as follows: S41: Construct an evidence package for the relevant evidence for each design intent slot, determine the evidence weight based on the source credibility, retrieval score and anchor point association strength, and normalize the unit, scope and version to generate a two-layer constraint representation for each slot. S42: After filtering based on the rules of the applicable domain of anchor points, the prior distribution of anchor points is formed by combining the proximity of evidence subgraphs and the authority of the source. The neighborhood of anchor points is screened and aggregated according to weight to obtain the normalized constraint representation, and the set of hard anchor points is determined at the same time. S43: Encapsulate the design intent slot, evidence set, evidence subgraph, normalized constraint representation, and hard anchor point set into a generation constraint package. Use this constraint package as controlled input to generate a structured draft, requiring key fields to be bound to evidence pointers or anchor point pointers. S44: Render the structured draft into a human-readable solution text, and perform consistency checks on each hard anchor point set, marking the state of being satisfied, not satisfied, or lacking information, and output the design solution with traceability pointers. S45: Perform consistency verification on the design scheme, detect violations and missing information, perform targeted supplementary retrieval and local rewriting for conflict points, iteratively update the scheme and evidence subgraphs until the termination conditions are met, and output the final design scheme and a list of information to be supplemented.

[0012] A computer device includes a memory and a processor, wherein the memory stores a computer program and the processor executes the computer program to implement the above-described intelligent question-answering method for R&D services based on industry evidence and standard anchors.

[0013] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned intelligent question-answering method for R&D services based on industry evidence and standard anchors.

[0014] Based on the above technical solution, the present invention has the following beneficial technical effects: 1. This invention abstracts the "must be satisfied and can be determined" clauses in standards / specifications into a set of standard anchor points, and introduces anchor point constraints from the evidence recall and solution generation stages to limit the feasible solution space, so that the output solution can be determined as "satisfied / not satisfied / insufficient information", thereby improving the engineering executability and compliance consistency of the solution and reducing the risk of non-implementation caused by relying solely on free generation.

[0015] 2. This invention constructs an industry multimodal knowledge graph, abstracts standard clauses and other data into evidence objects that can be independently indexed and version-managed, and establishes supporting associations between evidence, entities and relational conclusions through evidence pointers, realizing traceable mapping between solutions and the terms of reference, thereby improving the verifiability and reviewability of design decisions.

[0016] 3. To address the issues of unstructured, implicit constraints, and missing information in multimodal inputs, this invention parses text, images, and historical context into structured design intent slots. When the anchor point judgment field coverage is insufficient, it triggers gap-driven complementary association retrieval to form "information items to be supplemented / supplementary retrieval evidence," thereby improving the constraint recognition and coverage completeness for complex design scenarios.

[0017] 4. This invention unifies units, standards, and versions through evidence aggregation and normalization, and employs evidence whitelist injection and mandatory citation rules during the generation stage. After generation, it combines consistency verification to perform targeted supplementary retrieval and local rewriting for conflict points, reducing drift caused by overall regeneration. It also follows the control principle of not outputting hard parameters such as thresholds / levels when evidence is insufficient, thereby improving the stability and reliability of the output results. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the intelligent question-answering method for R&D services based on industry evidence and standard anchors, as presented in this invention.

[0019] Figure 2 This is a bar chart showing the comparison of ablation experiments between the method of this invention and existing algorithm models. Detailed Implementation

[0020] To describe the present invention in more detail, the technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] like Figure 1As shown, this invention provides an intelligent question-answering method for R&D services based on industry evidence and standard anchors, specifically including the following steps: (1) Perform evidence fragment analysis on multi-source data such as standards, regulations, process specifications, testing methods and engineering cases, and construct an industry multimodal knowledge graph with evidence as the core and traceability.

[0022] First, given the user's multimodal input set:

[0023] in: This refers to the user's description of their needs in natural language. This refers to user-provided diagrams, reference images, or structural pictures. This indicates the context information of the historical dialogue.

[0024] The output of the model in this invention is a feasible design solution. Design scheme It includes: component composition, material recommendations, process routes, key parameter ranges, testing items and judgment methods, and compliance explanations, and provides traceable industry evidence for key conclusions.

[0025] Unlike unrestrained free generation, this invention models the task as a retrieval-enhanced reasoning process constrained by industry evidence and standard anchors. This process satisfies the following principles: key conclusions in the solution should be supported by the retrieved evidence; the solution must not violate the explicitly abstracted set of standard anchors; when the evidence is insufficient to support the judgment, it is prohibited to output hard parameters such as specific thresholds or levels, and instead outputs "information to be supplemented + recommended acquisition method".

[0026] To unify the management of information scattered across standards / regulations, process specifications, testing method documents, corporate specifications, patents, and proven cases, this invention constructs an industry multimodal knowledge graph:

[0027] Where: node set Used to represent entities such as functional objectives, key components, materials, processes, parameters, testing methods, standard clauses, typical failure modes, and engineering cases; edge set Used to represent relationships between nodes, including but not limited to "adaptation", "dependency", "constraint", "detection correspondence", "scope of application", "mutual exclusion / compatibility", "substitution", etc.

[0028] Any node or edge in the graph can be used to attach evidence fragments. Fragments of evidence Evidence pointers are formed by recording evidence metadata, including text paragraphs, table entries, diagrams and their descriptions, and process entries. Evidence pointers It includes fields such as source type, document identifier, version information, location information, applicable domain information, and key value information, thereby providing a unified basis for the traceability and consistency determination of subsequent outputs.

[0029] In this invention, the industry multimodal knowledge graph adopts an evidence-driven graphing mechanism with evidence fragments as primary objects. Specifically, it first performs layout-aware parsing and segmentation on multi-source documents such as standards, regulations, process specifications, testing methods, patents, and engineering cases, and then abstracts multimodal content such as text paragraphs, table cells, process items, and diagrams and their descriptions into a set of minimally identifiable evidence fragments.

[0030] And generate a unique evidence pointer for each piece of evidence. This is used to record the source document identifier, version number, location information, applicable domain, and explicitly given key parameters or judgment information. For graphic evidence, engineering intermediate representations (structural objects, connection relationships, dimensional clues, material / surface features, etc.) are extracted and bound to the corresponding figure captions as the same evidence fragment to achieve consistent management of multimodal evidence.

[0031] At the evidence representation level, this invention simultaneously constructs textual semantic representation and visual / structural semantic representation for each evidence fragment, and forms a unified embedded representation through a controlled fusion function:

[0032] in: and These represent the vector representations of the text and image channels, respectively. Meta-features include evidence type, source authority, and domain labels. Textual semantic representation is obtained through word segmentation, semantic encoding, and contextual modeling of the text content in the evidence fragments. Visual / structural semantic representation is obtained by structurally analyzing the diagrams, tables, or structured information contained in the evidence fragments, extracting structural objects, connections, size cues, material or surface features, and then vectorizing the above structured results. This method of multi-channel modeling and controlled fusion based on evidence meta-features aims to address the problem of inconsistent cross-modal expressions of the same engineering fact and significant differences in evidence reliability in industrial evidence. Through this mechanism, evidence expressing the same engineering fact in different modalities can be aligned in the same representation space while preserving their modal source differences, thereby reducing the interference of weak evidence on subsequent retrieval, evidence aggregation, and standard anchor point determination.

[0033] Based on this, the present invention uses evidence fragments as the constraint source to perform controlled normalization and relation generation on candidate entities; through an industry-controlled lexicon and synonym mapping rules, mentions of functional objectives, components, materials, processes, parameters, detection methods, and failure modes appearing in the evidence are mapped into standardized entity nodes. For entities whose mapping results cannot be uniquely determined, mention Construct a candidate set:

[0034] in: This represents the set of candidate normalized entity nodes obtained by matching a controlled vocabulary with synonym mapping rules. Indicates mention Normalization to candidate nodes Confidence level, This represents the number of candidates. The generation of relationships between entities is limited to a predefined set of relationship templates, and the relationships must... There must be at least one piece of evidence. This provides support, thereby forming a relational representation with evidentiary annotations. , and These are all normalized entity nodes in the graph.

[0035] To enhance the traceability of the graph, this invention treats evidence fragments as explicit objects and establishes associations with entities and their relationships, enabling any conclusion in the graph to point back to its source evidence. For the version evolution of standards and specifications, sub-graphs are sliced ​​according to document identifiers and version numbers, and maintenance time. In addition, incremental updates are performed only on the affected evidence and its associated subgraphs during version updates to support long-term consistency maintenance.

[0036] (2) Abstract a set of standard anchor points that can be judged from standards and specifications, and parse the user's multimodal input into structured design intent elements, and explicitly label constraints, conflicts and missing information.

[0037] Based on the aforementioned knowledge graph, this invention abstracts the constraints that are "must be met and can be determined" in standards, regulations, and enterprise design specifications into a set of standard anchor points:

[0038] in: For the number of anchor points, each anchor point Including the scope of application Constraint Types Constraint Expressions and anchor point evidence pointers , Used to describe the product category, component object, and operating / environment range to which the anchor point applies; Used to indicate constraint categories (threshold constraints, level constraints, prohibited item constraints, mandatory inspection process constraints, etc.); Used for formal representation of decidable rules; A location marker indicating the source of the evidence at this anchor point. Evidence pointer. For any piece of evidence Unified positioning and metadata encapsulation; anchor point evidence pointer It is a standard anchor point The "source location identifier" stored therein is used to store the constraint expression of the anchor point. It refers back to which clause / form cell it comes from.

[0039] Anchor set The construction and implementation method is as follows: First, locate clause paragraphs and table entries containing judgmental statements such as thresholds, levels, prohibitions, and mandatory inspection processes in standards / regulations / enterprise specifications; then, extract fields from the clause content, including extraction objects (components / materials / parameters), conditions (applicable domains), constraint values, units, and judgment criteria; finally, map the extraction results to a constraint template library to generate... The template library includes numerical comparison, interval inclusion, set membership, process inclusion / sequence constraints, mutual exclusion constraints, etc.; finally, version information and applicable domains are bound to anchor points, and source anchor point evidence pointers are recorded. When the standard version is updated, the anchor point set is incrementally updated according to the document identifier and version field, and historical versions are retained to support traceability and comparison.

[0040] It should be noted that the standard anchor point in this invention is not only used for post-generation verification, but also for defining the feasible solution space from the retrieval stage. Subsequent evidence retrieval, evidence aggregation, and solution generation all need to be aligned with the anchor point so that the final output can be directly determined as "satisfied / not satisfied / insufficient information".

[0041] Industrial design requirements often suffer from vague descriptions, implicit constraints, or missing information. To reduce the bias caused by directly retrieving data from raw text / images, this invention will input... Parsed into a structured set of design intent slots:

[0042] in: Indicates functional objectives. Represents key components / structural objects. Indicates the usage scenario and working conditions. Indicates key parameter items and their ranges. Indicates material preferences and disabling. Indicates process constraints. Indicates compliance / standards requirements, Indicates the evaluation indicators.

[0043] This invention is based on text. ,image With historical context For each design intent slot Extract candidates and form a candidate set and its confidence level representation. The historical context candidates are accompanied by time stamps. This is used to characterize its recentity. For image input... Optimize the generation of engineering structure description It includes structural objects, connection relationships, size clues, and material or surface features, and is used for slot candidate generation and consistency judgment.

[0044] To avoid misjudgments caused by fixed priorities, this invention adopts a fusion approach of "supplementary priority, conflict re-decision" to merge multimodal candidates into:

[0045] And for any candidate value Calculate the overall score:

[0046] in: Representing candidate values Extracting confidence scores from text / image / historical context For the current moment and Time difference, Used to determine candidate values ​​and other modalities (especially image structure descriptions). Whether they are consistent, consistency function A computable multi-factor decision is adopted: combining candidate values ​​with text / historical semantic consistency and image engineering structure description. The structure is consistent with the standard anchor points within the applicable domain, and penalties are imposed for mutual exclusion conflicts with other candidate values, thereby obtaining... Consistency score; where structural consistency is calculated based on the overlap and fit of the set of structural objects, the set of connection relationships, the set of size clues, and the set of material / surface features; parameters The preset weights are selected based on the slot type.

[0047] When a candidate value meets the slot-specific threshold gate, the slot master value is determined and retained in the form of "master value + supplementary value" to explicitly represent the supplementation of image or historical information to the text description; the threshold gate includes at least: the comprehensive score is not lower than a preset threshold, and there is a minimum score interval with the second-best candidate. The slot must meet the following conditions: it must be at least as high as the consistency threshold, the number of conflicts with other candidates must not exceed the upper limit, and the sufficiency of evidence constraint must be satisfied; otherwise, the slot will be marked as... It also outputs the information items that need to be supplemented and the recommended ways to obtain them, so as to avoid giving a definitive conclusion under the condition of insufficient information.

[0048] Through the above element-based representation, industry constraints implicit in natural language and images are made explicit and aligned with knowledge graph fields and standard anchors.

[0049] (3) Perform explicit relevance retrieval, bridging retrieval and complementary relevance retrieval driven by anchor gap with design intent and standard anchor points as constraints, and construct evidence subgraphs to support the judgment.

[0050] The retrieval objective of this invention is not merely to return semantically similar segments, but to extract evidence subgraphs that can support subsequent "satisfaction / dissatisfaction / insufficient information" determinations. Therefore, in addition to relevance, the retrieval process also uses a standard anchor set. The field can be determined as a coverage constraint: when the coverage... When the targeted supplementary search is triggered; The supplementary retrieval ends and the evidence subgraph extraction phase begins, with coverage... Defined as: for the current intent slot Evidence subgraph in the applicable set of standard anchor points The proportions of the fields required for anchor point determination have been provided. For each slot... Constructing query representation Explicit relevance search results Its basic score is:

[0051] in: and This is a non-negative weighting coefficient used to balance the influence of semantic similarity and term matching in search scoring, and satisfies... .

[0052] To reduce discrepancies between image cues and clause descriptions, the image engineering structure description is reused. And form a bridging query:

[0053]

[0054] in: .

[0055] For information that is "not explicitly stated by the user but necessary for anchor point determination," this invention changes the complementary retrieval trigger from general association expansion to anchor point gap driving, and defines the set to be completed:

[0056] in: The fields required to trigger anchor point determination and in Missing or low confidence.

[0057] For each field that needs to be completed In knowledge graphs Above, based on a predefined set of relation types This includes structural compatibility, material adaptation, process constraints, inspection compatibility, and clause references, and is traversed using a breadth-first search approach. The maximum search depth is set to [value missing]. During the traversal, only those paths that ultimately reach the "Standard Clause" or "Detection Method" node are retained, and the evidence fragments associated with these paths are summarized as fields. Supplementary evidence set .

[0058] To ensure that the search results can be used for anchor point determination, this invention introduces anchor point hit items during the rearrangement stage:

[0059] in: Used to adjust anchor point hit rate. The strength of the influence on the rearrangement results, base score Determined according to the following rules: for the same query With evidence Calculate their explicit relevance search scores respectively. Bridged retrieval score The scores from the two categories were then normalized and weighted according to preset coefficients. and Perform linear weighted fusion to obtain the base score. , and .

[0060] The anchor point hit item This is used to measure whether evidence aligns with the decision field of the standard anchor point and whether it has traceable evidence pointers. The evidence pointer Source pointers that can be linked to the corresponding anchor points Or generate verifiable mappings. For Performing a bridging search yields and with As a base ranking score, when the proportion of anchor points matched with the current slot that can be supported by evidence is below a threshold. When a complementary search is executed, and a supplementary search is triggered, the following closed-loop process is followed: When And the number of iterations At that time, based on Generate query and update ;when or The search will stop and the status will be marked as "coverage satisfied" or "insufficient information".

[0061] Final evidence set It consists of direct evidence obtained through explicit relevance searches for each design intent slot, supplementary evidence obtained through bridging searches combined with engineering structural descriptions, and supplementary evidence obtained through complementary correlation searches for fields necessary for anchor point determination but currently lacking information. After unifying and summarizing the above evidence, it is analyzed from an industry multimodal knowledge graph. Extracting evidence subgraphs from the middle And in this subgraph, it is possible to use standard anchor pointers. The standard clause nodes and detection-related nodes are marked to make them candidate sources for subsequent evidence aggregation, anchor projection, and hard anchor screening processes.

[0062] (4) Generate an executable design scheme under the constraints of evidence aggregation and anchor point projection, and form the final output that satisfies or clearly indicates insufficient information through anchor point consistency verification and orientation correction.

[0063] Evidence from different sources may have differences in units, threshold expressions, versions, or conflicts. To enable deducible reasoning based on standard criteria during the generation stage, this invention proposes an evidence aggregation and anchor point projection process, outputting a normalized constraint representation. With hard anchor set .

[0064] First, for the same slot The relevant evidence forms an evidence package:

[0065] in: For the first A fragment of evidence, The weight of evidence is determined and normalized as follows:

[0066] in: For source credibility function; This is the anchor point association strength function, used to measure the degree of matching between evidence and anchor point fields.

[0067] Subsequently, the information execution units, definitions, and versions in the evidence package were normalized. Unit normalization used a unit mapping table and conversion rules to unify different units into the target unit; definition normalization converted different expressions such as "maximum value," "not less than," and "range" into unified range or upper and lower bound representations; version adjudication was executed according to fixed rules: if the user... or If a version number is specified, that version will be selected; otherwise, the highest version number that applies to the domain will be selected; if the version numbers are the same, the one with the latest release date will be selected.

[0068] After completing the normalization of units, calibers, and versions, the slots will be... evidence package Summarized as a two-level constraint representation:

[0069] in: For the continuous semantic representation of the evidence package, For determinate structured constraint fields (such as upper and lower bounds of intervals, set membership / prohibited items, process inclusion / sequence requirements, etc.). For each standard anchor point The same structure ,in From anchor point constraint expression The feasible set obtained through induction is preserved, along with its anchor evidence pointers. .

[0070] To avoid cross-domain misalignment caused by "only vector nearest neighbor", this invention first considers the applicable domain of the anchor point. (Fields such as product category, component object, and operating condition / environment range) and the domain information obtained from slot parsing (from...) (and its controlled vocabulary normalization results) are filtered by rules: only anchors with consistent or non-conflicting applicable domains are retained for subsequent alignment; when the above domain field is in the slot If the confidence level is insufficient, hard filtering is not performed; instead, a soft alignment process using the nearest neighbor property and source authority of the evidence subgraph is initiated for adjudication. The evidence subgraph is then introduced after rule filtering is completed. The proximity and source authority form a differentiable anchor point prior distribution:

[0071] Based on this, take Top- Forming a neighborhood .

[0072] After completing the anchor point neighborhood selection, the prior distribution of the neighborhood anchor points is... In Top- Weights are obtained by normalization within the neighborhood. ,in and Based on this, feasible set projection is performed on the structured fields: for each field , to induce neighborhood anchor points By weight Perform aggregation constraints to obtain field results that meet the standard criteria. Ultimately, the following is formed:

[0073] The generation phase of this invention aims to output an executable, compliant, and traceable design solution, rather than a free textual restatement of retrieved evidence. Therefore, this invention employs a controlled generation process constrained by evidence and standard anchor points. Specifically, before generation, the system sets up design intent slots. Collection of evidence and the evidence subgraphs it constitutes Normalized constraint set and hard anchor point set Encapsulate as a constraint package This serves as the controlled input context for the generative model. To avoid untraceability caused by free expansion, this invention uses a set of evidence. The organization creates an evidence whitelist: for each piece of evidence. Assign a unique evidence ID and provide its evidence pointer. .

[0074] The generation process begins with the generative model outputting a draft structured scheme under the aforementioned structural constraints. Populate the fields in each module and bind evidence pointers or anchor pointers to key fields:

[0075] in: The field values ​​in the document are only allowed to originate from the evidence whitelist. Subsequently, without adding any new parameters, thresholds, or judgment conclusions, the structured draft will be... Render the solution text to be human-readable. It also preserves the correspondence between fields and evidence pointers to support traceability.

[0076] To prevent the generation of indeterminate conclusions under insufficient information, this invention introduces minimum necessary generation control rules during the generation stage. For key fields such as parameter range, level, detection threshold, prohibited materials, and compliance determination, the output must be bound to an evidence pointer. or anchor pointer If binding fails, outputting specific values ​​or levels will be prohibited, and the field will be marked as "insufficient information" and added to the list of information to be supplemented. For hard parameters such as thresholds, levels, sizes, or performance upper and lower bounds, only when the normalized constraint set is considered... Deterministic values ​​are only allowed to be output when there are determinate fields that can point back to the source of evidence; otherwise, they are treated as insufficient information.

[0077] In the compliance module, for the set of hard anchor points To generate a draft Perform a line-by-line consistency check:

[0078] Where: 1 indicates satisfaction, 0 indicates non-satisfaction, and 2 indicates that a determination cannot be made due to insufficient information. The determination result and its corresponding evidence pointer must be explicitly given in the output; when there are anchor points that are not satisfied or have insufficient information and affect key conclusions, they are recorded as violations and handed over to subsequent consistency verification and local correction closed-loop processing.

[0079] Output scheme The organization is structured into fixed modules, including: a solution skeleton (component composition, structural relationships, and functional implementation path), a key parameter table (parameter items, recommended range / level, unit, applicable conditions, and corresponding evidence pointers), a materials and process route (material selection, process steps, key control points, compatibility / mutual exclusion descriptions, and corresponding evidence pointers), a testing and judgment plan (testing items, methods / steps, judgment thresholds / levels, and corresponding clause pointers), compliance conclusions (providing "satisfied / unsatisfied / insufficient information" and evidence pointers for each hard anchor point), and a list of information to be supplemented.

[0080] After generation, the present invention runs a consistency checker and verifier, which verifies the following: anchor point consistency (threshold, level, prohibited items, mandatory inspection process), unit consistency (consistency of the same parameter across paragraphs and correct conversion), parameter self-consistency (reasonable upper and lower bounds of the range and matching with the working conditions), material-process compatibility / mutual exclusion (surface treatment and material mutual exclusion, process window conflict, etc.), and evidence binding integrity (whether there are traceable pointers for key conclusions).

[0081] If a conflict or missing element is found, this invention only performs "targeted supplementary retrieval + local rewriting" on the conflict point, rather than regenerating the entire system, in order to reduce drift and improve controllability. This closed loop can be represented as:

[0082] in: Output the set of non-compliant anchor points, their corresponding paragraph positions, missing parameters, reasons for conflicts, and query anchor points that need to be added. This indicates a partial repair operation, which only updates the evidence package, subgraph nodes, and corresponding modules in the solution related to the violation anchor point, thereby obtaining... With the updated evidence subgraph .

[0083] To ensure the process is reproducible and converges, this invention sets a termination condition: when all hard anchor points satisfy (i.e.) The iteration will terminate when: (1) there are still insufficient information items, but a clear list of information to be supplemented has been output and no modeling parameters have been compiled, and the maximum number of iterations has been reached. The iteration will terminate when there are no changes in the set of key fields and no new violations are added in two consecutive iterations.

[0084] To verify the effectiveness of the method of this invention, we used a multimodal industrial design question-and-answer dataset for innovative design of molds and home appliances, provided by a regional manufacturing industry alliance. This dataset was used to evaluate the invention's ability to retrieve industry evidence in real-world industrial scenarios, as well as its compliance performance in generating multimodal R&D service questions and answers constrained by evidence and standards. The dataset consists of two parts: an industry evidence corpus and a multimodal design question-and-answer evaluation set, covering the complete design process from requirement understanding and implicit constraint completion to specification alignment and result verification.

[0085] The industry evidence corpus compiles real engineering data from SMEs in the collaborative R&D of home appliances and molds between 2020 and 2024, including standards and regulations, mold processes and manufacturing specifications, testing and verification documents, and typical failure and engineering improvement cases. To support traceable retrieval and generation, all original documents were segmented into fine-grained evidence fragments and uniformly modeled as a multimodal knowledge graph of "functional objectives—components—materials—processes—parameters—testing—standard clauses—failure cases." Determinable standard anchors were also abstracted from standards and specifications for subsequent consistency judgment and compliance verification. The corpus contains 1908 documents, segmented into 378,540 evidence fragments, and is accompanied by 24,300 multimodal image resources. Statistical data for the dataset are shown in Table 1. Table 1

[0086] This multimodal design question-and-answer evaluation dataset addresses common innovation and mold manufacturability issues in home appliance products faced by SMEs. Each example consists of requirement text, reference images, and a multi-turn dialogue history, accompanied by structured annotations including functions and component slots, critical engineering constraints, gold standard evidence pointers, and anchor point fulfillment labels. The evaluation dataset contains 3840 examples, 74.6% of which include image input. The average number of dialogue turns is 3.1. Data is grouped by project and time, with a training / validation / test ratio of 8:1:1.

[0087] Experimental parameter settings: The experiment was conducted on a server equipped with Ubuntu 22.04, with hardware configuration including two Intel(R) Xeon(R) Platinum 8488C CPUs, four NVIDIA RTX A6000 GPUs, and 256 GB of memory; the software environment was Python 3.10 and PyTorch 2.2. In the industry multimodal knowledge graph construction phase, TransE was used for representation learning, with the entity / relation vector dimension set to 256. Training employed the AdamW optimizer with a learning rate of 0.001, a batch size of 4096, and 100 epochs (using an early stopping strategy). In the design intent slot parsing, the confidence threshold was set to 0.6; in complementary association retrieval, the explicit relevance mixed scoring weights were set to λ1=0.65 and λ2=0.35, and the image bridging retrieval fusion coefficient was... β =0.6; Recall the top-20 evidence fragments for each slot. The source credibility Trust(e) in the evidence aggregation stage is assigned a graded value (National / Industry Standards = 1.0, Enterprise Specifications = 0.8, Testing Procedures = 0.7, Patents and Engineering Cases = 0.6).

[0088] Comparative experiment: To comprehensively evaluate the effectiveness and practical value of the method of this invention in real industrial design scenarios, this invention conducts comparative experiments on two levels: industry evidence retrieval capability and the quality of multimodal R&D service question-answer generation. The evidence retrieval capability comparison experiment uses an industry evidence corpus as the evaluation object, focusing on assessing the model's recall and coverage capabilities for key engineering evidence, standard clauses, and implicit constraint information under real industrial requirements. The multimodal R&D service question-answer generation comparison experiment, based on a multimodal design question-answer evaluation set, systematically compares the accuracy, traceability, and compliance performance of the model's generated results under standard and engineering constraints.

[0089] Evidence retrieval capability - evaluation indicators: To evaluate the effectiveness of the method in industry evidence retrieval tasks, we use recall (Recall@K) and normalized depreciation cumulative gain (NDCG@K) as the main evaluation metrics. Recall@K represents the proportion of gold-standard evidence relevant to the current industrial design problem label among the top K evidence fragments in the retrieval results, used to measure the model's overall recall capability for key engineering evidence and standard clauses. NDCG@K further considers the ranking position of relevant evidence in the retrieval results list, assigning higher weights to relevant evidence ranked higher, thereby reflecting the model's performance in evidence ranking and importance differentiation. To comprehensively analyze the model's performance under different retrieval scales, K was set to 10, 20, and 50 in the experiments. The values ​​of the above metrics range from 0 to 100, with higher values ​​indicating better model performance in industry evidence retrieval tasks.

[0090] Evidence retrieval capability - comparative experiment: To demonstrate the effectiveness of the method (model) of this invention, several cutting-edge methods (models) were selected in the experiment for comparison with the method of this invention. Existing methods include: Text-Image Residual Gated Model (TIRG): Combines reference images and modified text into query vectors through "residual update + gating fusion" for combined image retrieval; Modality-Independent Attention Fusion Network (MAAF): Treats spatial tokens of images and word tokens of text as modality-independent sequences, concatenates them and fuses them using dot product attention / self-attention, thereby learning the common embedding representation of combined queries; Attention-Based Retrieval (ARTEMIS): Combines text explicit matching and visual implicit similarity through attention mechanism to achieve more effective combined image and text retrieval; CLIP-Based Combined Image and Text Retrieval Model (CLIP4CIR): Based on CLIP dual encoder, it models the combination of image and text conditions in a shared embedding space to achieve efficient combined image and text retrieval; BLIP-Based Combined Image and Text Retrieval Model (BLIP4CIR): Uses BLIP visual-language representation as the backbone network to fuse image and text conditions, improving the combined retrieval effect under stronger cross-modal semantic understanding.

[0091] Evidence retrieval capability - Experimental results: Table 2 shows the experimental results of the model of this invention (hereinafter referred to as this model) and various baseline models on the industry evidence corpus.

[0092] Table 2

[0093] As can be seen from Table 2, this model achieves the best performance across all metrics, indicating that the design of this invention in evidence modeling and retrieval ranking optimization for industrial constraints can steadily improve the "hit rate" and "top-rank quality", thereby effectively supporting the subsequent question-and-answer generation constrained by evidence and standards.

[0094] Compared to other methods, traditional combined image-text retrieval models (such as TIRG, MAAF, and ARTEMIS), while achieving a basic fusion of "reference image + modified text," rely more on general semantic matching. They struggle to explicitly characterize the importance of strong constraints in industrial scenarios, such as "standard thresholds, prohibited items, and mandatory inspection processes." This leads to relevant evidence being diluted by descriptive but not constraining fragments, and makes it harder to prioritize key terms in the ranking. Consequently, these methods generally have lower recall and NDCG scores, reflecting their insufficient ability to quickly pinpoint key evidence in real-world engineering retrieval.

[0095] Strong baselines, represented by CLIP4CIR and BLIP4CIR, have significantly improved matching performance in industrial corpora due to their stronger cross-modal representation capabilities, with overall Recall and NDCG being significantly higher than the aforementioned methods. However, as shown in Table 2, our model continues to lead in Recall@10, Recall@20, and corresponding NDCG@10 and NDCG@20, indicating that this invention not only improves the ability to "find relevant evidence" but also further strengthens the ability to "prioritize the most critical and determinate clause evidence." This demonstrates that the model is better at distinguishing the engineering importance and constraint strength of evidence, and can prioritize presenting evidence that truly affects compliance judgments and solution feasibility to the generation module, thereby improving traceability and reducing the risk of clause conflicts.

[0096] Multimodal R&D Service Question and Answer Generation - Evaluation Metrics: To evaluate the effectiveness of the method proposed in this invention for large-scale model-driven product design question-answering tasks, the model performance is assessed from three aspects: constraint satisfaction, parameter accuracy, and overall solution quality. We introduce the Constraint Satisfaction Rate (CSR) to measure the degree to which the design parameters and solutions generated by the model meet user requirements and engineering constraints. CSR is defined as the ratio of the number of constraints satisfied in the generated results to the total number of constraints. This indicator directly reflects the controllability and reliability of the model under complex design constraints. For the generation quality of structured design parameters, Parameter Accuracy is used to evaluate the numerical parameters output by the model, characterizing the numerical accuracy of the model in generating design parameters. To comprehensively evaluate the rationality and engineering feasibility of the model-generated design solutions, a Human Overall Quality Score is introduced. Multiple reviewers with relevant field backgrounds score the generated results from dimensions such as solution feasibility, requirement matching, and overall design quality. The values ​​of the above indicators range from 0 to 100, with higher values ​​indicating better quality of the model's generated results.

[0097] Multimodal R&D Service Question and Answer Generation - Comparative Experiment: To demonstrate the effectiveness of this model, several cutting-edge methods (models) were compared with the method of this invention in the experiment. Existing methods include GPT-4O: OpenAI's flagship multimodal large model, supporting multimodal understanding and generation of text and images; Retrieval Enhanced Generation (RAGNLP): retrieves evidence from external corpora through dense retrieval and jointly models it with the generative model to improve the factuality and traceability of the answer; Decoder Fusion Retrieval Generation (FiD): encodes multiple retrieval evidences separately and fuses them during the decoding stage to more effectively integrate multi-evidence information; Chained Validation (CoVe): first generates a draft, then plans verification questions and answers them independently to avoid self-bias, and finally outputs the corrected final answer based on the verification results, thereby reducing illusions and inconsistencies; Adaptive Multimodal Retrieval Enhanced Generation (SAM-RAG): adaptively selects and filters retrieval evidence in multimodal scenarios, and introduces image descriptions when necessary to improve the reliability of generation.

[0098] Multimodal R&D Service Question and Answer Generation - Experimental Results: Table 3 shows the experimental results of this model and the baseline models on the multimodal design question-and-answer evaluation set.

[0099] Table 3

[0100] As can be seen from Table 3, the model achieves the best performance in all three indicators: CSR, Parameter Accuracy, and Human Overall Quality. This shows that the method of the present invention can simultaneously improve constraint satisfaction, parameter generation accuracy, and overall solution usability in the R&D design question-and-answer scenario.

[0101] Regarding constraint satisfaction rate, while end-to-end generative models such as GPT-4O possess strong expressive and cross-modal understanding capabilities, they are still prone to omissions or inconsistencies when faced with strong engineering constraints such as standard thresholds, prohibited items, and mandatory inspection processes. After introducing retrieval enhancement and verification mechanisms (RAGNLP, FiD, CoVe, SAM-RAG), the CSR gradually improved, indicating that external evidence and posterior verification play a positive role in reducing constraint violations. Furthermore, this model further improves constraint satisfaction stability by explicitly modeling constraints and performing anchor point consistency verification.

[0102] Regarding parameter accuracy, traditional generation methods and standard RAG methods still suffer from issues such as numerical offset and unit inconsistency. This model, by binding key parameters to standard anchor points and performing consistency checks, makes parameter generation more accurate and reliable, outperforming all compared methods. In terms of comprehensive human quality scoring, this model received the highest rating, indicating that its generated results better meet the actual needs of engineers in terms of scheme structural integrity, engineering feasibility, and clarity of compliance documentation.

[0103] Ablation experiment: To evaluate the contribution of each key component in the method of this invention to the performance improvement, this invention removes or degrades several core modules based on the full model for comparative experiments. The definitions of each ablation model are as follows: w / o Bridge (Image-free bridging retrieval): Removes the "bridging query" formed by the engineered structural description generated based on the reference image. The retrieval stage only uses direct text-image fusion queries and no longer introduces bridging pathways to enhance evidence recall and ranking.

[0104] w / o Comp (No Complementary Association Retrieval): The complementary association retrieval process driven by "anchor gaps" is removed. That is, gap detection and map expansion and completion are not performed on the missing standard anchors in the initial retrieval results. The final evidence set consists only of direct retrieval (including basic fusion) results.

[0105] w / o AnchorRerank: Preserves retrieval recall but removes the explicit gain term of standard anchor relevance on ranking score. The retrieval ranking degenerates into a ranking mechanism that only depends on basic similarity, thereby verifying the role of "anchor-guided front-row presentation".

[0106] w / o Agg&Proj (no evidence aggregation and anchor projection): Removes source credibility weighting, anchor relevance weighting, and unit / version / caliber normalization from the evidence aggregation stage. At the same time, it does not perform the "feasible set projection / alignment" process for design parameters. The generation stage directly uses the original retrieved evidence to splice the input.

[0107] w / o VerifyLoop (no consistency check closed loop): Removes the consistency check based on anchor points and the "targeted supplementary retrieval + local rewriting" closed loop mechanism after generation, and only performs one-time generation output, no longer triggering iterative correction and compliance completion.

[0108] like Figure 2 As shown, removing any key component will lead to varying degrees of performance degradation in the multimodal R&D service question-answering generation, indicating that each module of the present invention plays a necessary role in the overall chain of "evidence traceability retrieval - standard constraint alignment - compliance generation". Figure 2 The three sub-images from left to right correspond to the three indicators: CSR, Parameter Accuracy, and Human Overall Quality. After removing image bridging retrieval (w / o Bridge) and complementary association retrieval (w / o Comp), the model's ability to explicitly model the engineering constraints such as structure, materials, and processes implicit in the reference images decreases. The coverage and completeness of key standard clauses and mandatory inspection processes are affected, which in turn manifests as a decrease in constraint satisfaction rate and overall solution quality at the generation end. After removing anchor-guided ranking (w / o AnchorRerank), it is more difficult for key judgment clauses to be consistently ranked at the top of the search results, weakening the constraining and guiding role of evidence in the generation process.

[0109] Furthermore, removing the evidence aggregation and anchor projection modules (w / o Agg&Proj) significantly reduced parameter accuracy, indicating that source credibility weighting, unit and version normalization, and feasible set projection play crucial roles in ensuring numerical parameter consistency and engineering feasibility. Removing the generated consistency verification loop (w / o VerifyLoop) led to a decrease in both CSR and human comprehensive scores, demonstrating that the anchor consistency check combined with the iterative correction mechanism of targeted supplementary retrieval and local rewriting can effectively reduce the risk of constraint violation generation and improve compliance stability under complex engineering constraints. In summary, the ablation experiments verified that the key designs of this invention have a synergistic and significant contribution to improving the question-answering generation performance of multimodal R&D services constrained by evidence and standards.

[0110] The above description of the embodiments is provided to enable those skilled in the art to understand and apply the present invention. Those skilled in the art can readily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without creative effort. Therefore, the present invention is not limited to the above embodiments, and any improvements and modifications made to the present invention by those skilled in the art based on the disclosure thereof should be within the scope of protection of the present invention.

Claims

1. A smart question-answering method for R&D services based on industry evidence and standard anchors, characterized in that, The steps include the following: (1) Perform evidence fragment analysis on multi-source data, including standards, regulations, process specifications, testing methods, and engineering cases, and construct an industry multimodal knowledge graph with evidence as the core and traceability. The specific implementation method is as follows: S11: Perform layout-aware analysis and segmentation on multi-source materials including standards, regulations, process specifications, testing methods and engineering cases. Abstract multimodal content, including text paragraphs, table units, process items and illustrations, into a minimum set of citationable evidence fragments, and generate a unique evidence pointer for each evidence fragment, including source type, document identifier, version number, location information, applicable domain and key parameters. S12: For each evidence fragment and its evidence pointer, construct text semantic representation and visual / structural semantic representation respectively, and generate a unified embedding representation through a controlled fusion function based on meta-features; S13: Identify entity mentions from evidence fragments using embedded representations and evidence pointers, map entity mentions to normalized entity nodes based on industry-controlled vocabularies and synonym mapping rules, construct candidate sets for entity mentions that cannot be uniquely determined, and generate inter-entity relationships supported by at least one evidence fragment under the constraints of a predefined relationship template. S14: Integrate entity nodes and relationships between entities into an industry multimodal knowledge graph, and establish a traceable link between graph conclusions and original evidence fragments through evidence pointers; at the same time, maintain time-sliced ​​subgraphs according to document identifiers and version numbers, and only perform incremental updates on the affected parts when the version is updated to ensure the long-term consistency and traceability of the knowledge graph. (2) Abstract a set of definable standard anchor points from standards and specifications, and parse user multimodal input into structured design intent elements, explicitly labeling constraints, conflicts and missing information. The specific implementation method is as follows: S21: Based on the knowledge graph, locate clauses containing judgment statements including thresholds, levels, prohibitions and mandatory inspection processes from standards, regulations and enterprise specifications, and generate a standard anchor point set containing applicable domains, constraint types, constraint expressions and anchor point evidence pointers through field extraction and constraint template mapping. S22: Parse the user-input text, images, and historical dialogue context respectively; extract candidate values ​​and confidence scores for each design intent slot from the text; generate an engineered structural description containing structural objects, connection relationships, size cues, and material characteristics from the image and extract corresponding candidate values; extract candidate values ​​with time markers from the historical context to form a multimodal candidate set for each slot. S23: Merge the multimodal candidate sets of each slot, calculate a comprehensive score for each candidate value, the comprehensive score is based on modal confidence weighting, historical time decay and consistency function, wherein the consistency function comprehensively evaluates the degree of matching between the candidate value and the text semantics, image structure description and standard anchor point identifiable alignment, and imposes a penalty on mutual exclusion conflicts with other candidate values; S24: Filter candidate values ​​based on slot threshold gating. When there are candidates that meet the scoring threshold, minimum interval, consistency threshold and evidence sufficiency constraints, determine the principal value of the slot and retain multimodal information in the form of principal value plus supplementary value; otherwise, mark the slot as unknown and output the supplementary information items and recommended acquisition methods. Finally, output the structured design intent elements aligned with knowledge graph fields and standard anchor points. (3) Explicit relevance retrieval, bridging retrieval, and complementary association retrieval driven by anchor gaps are performed with design intent and standard anchor points as constraints to construct an evidence subgraph supporting the judgment. The specific implementation method is as follows: S31: Based on the structured design intent slots and image engineering structure description, construct explicit relevant queries and bridging queries, execute the retrieval to obtain a preliminary evidence set, and calculate the basic score of each piece of evidence; S32: Align the preliminary evidence set with the standard anchor set, calculate the coverage of the fields required for anchor determination by the current evidence, and when the coverage is lower than the preset threshold, identify the set of missing fields necessary to trigger anchor determination as input for complementary retrieval; S33: For the identified missing fields, perform a breadth-first search on the industry multimodal knowledge graph with limited relation types to obtain evidence that can fill in the missing fields, and merge the newly added evidence to update the evidence set, while retaining the scoring information of each piece of evidence; S34: Repeat steps S32~S33 until the coverage of the anchor decision field of the evidence set meets the threshold or the maximum number of iterations is reached, and output the final evidence set and retrieval status marker. S35: Extract evidence subgraphs from the knowledge graph based on the final evidence set, rearrange them by combining the basic score and anchor point hit items, and mark the nodes that are aligned with the standard anchor point pointers for subsequent evidence aggregation and hard anchor point filtering. (4) Generate an executable design scheme under the constraints of evidence aggregation and anchor point projection, and form a final output design scheme that satisfies or clearly indicates insufficient annotation information through anchor point consistency verification and orientation correction. The specific implementation method is as follows: S41: Construct an evidence package for the relevant evidence for each design intent slot, determine the evidence weight based on the source credibility, retrieval score and anchor point association strength, and normalize the unit, scope and version to generate a two-layer constraint representation for each slot. S42: After filtering based on the rules of the applicable domain of anchor points, the prior distribution of anchor points is formed by combining the proximity of evidence subgraphs and the authority of the source. The neighborhood of anchor points is screened and aggregated according to weight to obtain the normalized constraint representation, and the set of hard anchor points is determined at the same time. S43: Encapsulate the design intent slot, evidence set, evidence subgraph, normalized constraint representation, and hard anchor point set into a generation constraint package. Use this constraint package as a controlled input to generate a structured draft, requiring key fields to be bound to evidence pointers or anchor point pointers. S44: Render the structured draft into a human-readable solution text, and perform consistency checks on each hard anchor point set, marking the state of being satisfied, not satisfied, or lacking information, and output the design solution with traceability pointers. S45: Perform consistency verification on the design scheme, detect violations and missing information, perform targeted supplementary retrieval and local rewriting for conflict points, iteratively update the scheme and evidence subgraphs until the termination conditions are met, and output the final design scheme and a list of information to be supplemented.

2. A computer device, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: The processor is used to execute the computer program to implement the intelligent question-answering method for R&D services based on industry evidence and standard anchors as described in claim 1.

3. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by the processor, it implements the intelligent question-answering method for R&D services based on industry evidence and standard anchors as described in claim 1.

Images