Double-path engineering process intelligent question-answering system and method based on posterior probability dynamic routing

Abstract

The invention discloses a double-path engineering process intelligent question-answering system and method based on posterior probability dynamic routing, and relates to the field of computers, and the system comprises a data processing module which is used for preprocessing multi-source heterogeneous engineering data to obtain preprocessed data; the user interaction module is used for performing intention understanding and task decomposition on the received input problem through a large language model agent, selecting an adaptive information retrieval mode and generating a query request; and the information processing module is used for executing knowledge retrieval and answer generation by utilizing the preset symbolized knowledge graph and / or semantic vector database according to the query request, and the symbolized knowledge graph and the semantic vector database are constructed based on the preprocessed data. According to the method and device, it can be guaranteed that the finally generated answers are comprehensive and coherent, the key process parameters have map-level accuracy, numerical illusion is effectively eradicated, and the reliability and fidelity of the answers are improved.

Description

Technical Field

[0001] This application relates to the field of computer science, and in particular to a dual-path intelligent question-answering system and method for engineering processes based on posterior probability dynamic routing. Background Technology

[0002] Currently, intelligent knowledge management in the field of engineering processes has become a core link in the digital transformation of industry. However, existing technologies are unable to effectively cope with the complex retrieval and high-fidelity question-and-answer requirements of multi-source heterogeneous data in high-precision process scenarios such as welding and casting.

[0003] Existing Retrieval Augmentation (RAG) technology suffers from several significant bottlenecks: First, it lacks an automated and standardized processing mechanism for multi-source heterogeneous engineering data, failing to normalize terminology and unify units for semi-structured and unstructured data such as process cards, design standards, and inspection reports, leading to severe semantic ambiguity and noise at the knowledge base level. Second, retrieval path allocation relies on static classification models, making it impossible to make adaptive routing decisions based on the posterior probability confidence of query intent, which can easily lead to incorrect path selection and missed key information when facing mixed engineering problems involving "fuzzy descriptions + precise parameters." Third, the cross-modal retrieval result fusion method is superficial, merely inputting structured facts and unstructured context into the large model through text concatenation, lacking an active anchoring and calibration mechanism for precise numerical values, making the model susceptible to semantic interference and producing "numerical illusions," failing to meet the "zero-tolerance" accuracy requirements of process parameters. Therefore, overcoming these shortcomings is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] In view of this, the present invention provides a dual-path intelligent question-answering system and method for engineering processes based on posterior probability dynamic routing, which overcomes the above-mentioned defects.

[0005] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a dual-path intelligent question-answering system for engineering processes based on posterior probability dynamic routing, including: The data processing module is used to preprocess multi-source heterogeneous engineering data to obtain preprocessed data; The user interaction module is used to understand the intent and decompose the received input questions through a large language model agent, select an appropriate information retrieval method, and generate a query request. The information processing module is used to perform knowledge retrieval and answer generation based on the query request using the preset symbolic knowledge graph and / or semantic vector database, wherein the symbolic knowledge graph and the semantic vector database are constructed based on the preprocessed data.

[0006] Optionally, the data processing module includes: The data cleaning unit is used to perform multimodal analysis on the multi-source heterogeneous engineering data, extract text feature data, and remove noise and redundant information to obtain cleaned data. The data conversion unit is used to normalize the cleaned data into standard data and generate the preprocessed data.

[0007] Optionally, the system further includes a bimodal engineering knowledge base construction module, which includes: The knowledge graph construction unit is used to construct hub nodes and process parameter nodes based on structured data and / or semi-structured data in the preprocessed data, extract causal logical relationships based on unstructured data in the preprocessed data, and generate a symbolic knowledge graph based on the hub nodes, the process parameter nodes, and the causal logical relationships. The vector database construction unit is used to construct a semantic vector database based on the unstructured text data in the preprocessed data through semantic segmentation and vectorization methods.

[0008] Optionally, the knowledge graph construction unit includes: The hub node construction subunit is used to model core business objects in structured and / or semi-structured data as hub nodes, and to extract the attributes or associated entities of the hub nodes in the structured and / or semi-structured data. The process parameter node construction sub-unit is used to abstract process parameters in structured and / or semi-structured data into independent process parameter nodes; The causal chain extraction subunit is used to extract information from the unstructured data, identify causal entities, and establish causal logical relationships between the causal entities. The graph generation subunit is used to construct a symbolic knowledge graph with explicit logical associations based on the hub node, the process parameter node, and the causal logical relationship.

[0009] Optionally, the vector database construction unit includes: The semantic segmentation subunit is used to segment the unstructured text data into multiple text blocks using a semantically based sliding window strategy. A high-dimensional vectorization subunit is used to convert the text block into a high-dimensional dense vector; The metadata association subunit is used to bind the high-dimensional dense vector and the associated metadata, and store them in the semantic vector database.

[0010] Optionally, the user interaction module includes: The confidence calculation unit is used to calculate the posterior probability confidence scores of the input question for different query types. The intent understanding determination unit is used to determine the input question as an exact fact query, a fuzzy semantic query, or a complex hybrid query based on a preset threshold and the posterior probability confidence score. The retrieval path selection unit is used to select, based on the judgment result, the activation of the knowledge graph retrieval path, the vector retrieval path, or the hybrid retrieval path.

[0011] Optionally, the information processing module includes a hybrid retrieval unit, which includes: The task receiving subunit is used to receive the mixed path instructions issued by the user interaction module; The dual-path retrieval subunit is used to execute the knowledge graph retrieval path and the vector retrieval path in parallel to obtain the structured fact set and the relevant text fragment set; The structured injection subunit is used to inject the structured fact set into the relevant text fragment set to generate enhanced context; The numerical consistency verification subunit is used to perform numerical consistency verification and correction on the enhanced context.

[0012] Optionally, the information processing module further includes an answer generation unit, specifically: The answer generation and tracing subunit is used to receive the enhanced context and generate the final natural language answer. In the natural language answer generation, the data source identifier embedded in the enhanced context is converted into a preset format reference mark and appended to the corresponding parameter. The closed-loop evolution subunit is used to obtain feedback data of the natural language answers and trigger differentiated correction strategies based on the type of feedback data to update the symbolic knowledge graph, semantic vector database, or optimize the decision logic of the user interaction module.

[0013] Optionally, the closed-loop evolutionary subunit includes: A feedback collection component is used to obtain explicit and implicit feedback on the natural language answer; An exception log generation component is used to associate and store the explicit feedback, the implicit feedback, the input question, and the retrieved data as an exception log. A knowledge base correction component is used to update the symbolic knowledge graph that is confirmed as factually incorrect based on the anomaly log; An agent optimization component is used to determine an optimization dataset based on the anomaly logs and to make decision-making decisions for the user interaction module based on the optimization dataset.

[0014] Secondly, this application provides a dual-path intelligent question-answering method for engineering processes based on posterior probability dynamic routing, the specific steps of which are as follows: Preprocessing is performed on multi-source heterogeneous engineering data to obtain preprocessed data; A symbolic knowledge graph and a semantic vector database are constructed based on the preprocessed data; The large language model agent performs intent understanding and task decomposition on the received input question, selects the appropriate information retrieval method, and generates a query request. Based on the query request, knowledge retrieval and answer generation are performed using the symbolic knowledge graph and / or semantic vector database.

[0015] According to the specific embodiments provided in this application, this application has the following technical effects: 1. This application solves the problem of accurate routing for hybrid queries by employing a dynamic routing mechanism based on agent-based posterior probability calculation, which intelligently predicts query intent and adaptively allocates strategies. Furthermore, through an "injection-based" fusion generation technology, it uses precise structured facts from the knowledge graph as anchors to semantically reconstruct unstructured text, achieving data-level source calibration. The combination of these two approaches ensures that the final generated answer is both comprehensive and coherent, while maintaining graph-level accuracy in key process parameters, effectively eliminating numerical illusions and improving the reliability and fidelity of the response.

[0016] 2. By employing technologies such as a built-in terminology dictionary and unit conversion engine, this application can automatically identify, parse, and unify the content of documents from different sources. This not only reduces the reliance on human experience in building a high-quality engineering knowledge base and improves the efficiency of knowledge entry, but also fundamentally breaks down the "semantic silos" between heterogeneous data sources, laying a solid foundation for subsequent accurate retrieval and correlation analysis.

[0017] 3. This application adopts a closed-loop maintenance method with a three-level linkage of "user-expert-developer" to automatically capture anomalies, generate diagnostic reports, and submit them to domain experts for final review and decision-making. Subsequently, the knowledge source is corrected through specialized tools. This not only ensures the efficiency and traceability of the knowledge update process, but also firmly safeguards the lifeline of the accuracy of engineering knowledge, effectively preventing the introduction and spread of erroneous knowledge, and ensuring the professionalism and credibility of the entire intelligent question-answering system in the dynamic evolution process. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the system framework provided in one embodiment of this application; Figure 2 This is a schematic diagram of knowledge graph nodes and relationships provided in an embodiment of this application; Figure 3 This is a schematic diagram of a method flow provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] This embodiment discloses a dual-path intelligent question-answering system for engineering processes based on posterior probability dynamic routing, including: The data processing module is used to preprocess multi-source heterogeneous engineering data to obtain preprocessed data; The user interaction module is used to understand the intent and decompose the received input questions through a large language model agent, select an appropriate information retrieval method, and generate a query request. The information processing module is used to perform knowledge retrieval and answer generation based on the query request using a preset symbolic knowledge graph and / or semantic vector database. The symbolic knowledge graph and semantic vector database are constructed based on preprocessed data.

[0023] Furthermore, such as Figure 1 As shown, the system mainly includes a data processing module, a user interaction module, and an information processing module, which are as follows: The data processing module is used to preprocess multi-source heterogeneous engineering data. It includes a data cleaning unit and a data transformation unit. The data cleaning unit removes noise and redundant information, while the data transformation unit converts the cleaned data into a standardized format (such as TXT, JSON, etc.) for subsequent knowledge modeling and retrieval.

[0024] The User Interaction Module (User) receives user-input questions and decomposes them into tasks using a Large Language Model Agent (LLM Agent). This agent selects an appropriate information retrieval method based on the question type and sends the generated query request to the information processing module, while also providing feedback and interaction on the results.

[0025] The Information processing module is used for knowledge retrieval and answer generation. This module includes a Graph Query unit, a Vector Search unit, and an Answer Generation LLM unit. The Graph Query unit performs semantic association retrieval based on the knowledge graph; the Vector Search unit performs semantic similarity searches in a vector database; and finally, the Answer Generation unit integrates multi-source information to generate a user-understandable natural language answer.

[0026] In one embodiment, the data processing module includes: The data cleaning unit is used to perform multimodal analysis on multi-source heterogeneous engineering data, extract text feature data, and remove noise and redundant information to obtain cleaned data. The data transformation unit is used to normalize the cleaned data into standard data and generate preprocessed data.

[0027] Furthermore, the data processing module systematically preprocesses multi-source, heterogeneous engineering and technological knowledge. The aim is to clean the raw data and perform standardized transformations to generate a high-quality dataset suitable for subsequent knowledge extraction and model processing. The system accesses structured data (such as Excel process cards), semi-structured data (such as XML standard documents), and unstructured data (such as PDF / Word specifications) through preset interfaces.

[0028] Furthermore, taking welding process as an example, the data cleaning unit performs the following steps: For structured data (such as welding process cards in Excel format) and semi-structured data (such as XML standard documents): the Pandas library is used to read the data, and key-value pair mappings are established based on header keywords (such as "weld number", "current", "voltage") to retain the correspondence between parameters and their applicable working conditions; For unstructured data (such as process specifications and national standard documents in PDF / Word format): the PaddleOCR engine is integrated to extract the text layer, and regular expressions are used to remove headers, footers, and garbled noise while retaining the paragraph structure.

[0029] In one embodiment, taking welding process as an example, the data conversion unit replaces non-standard terms with standard terms through terminology normalization mapping, and converts non-standard units into international standard units through physical unit unification conversion. The specific execution steps are as follows: A synonym dictionary for the welding field is constructed, containing mappings between standard terms and common aliases. For example: {"TIG": "tungsten inert gas welding", "MAG": "gas metal arc welding", "argon arc welding": "tungsten inert gas welding"}. During text preprocessing, the Jieba word segmentation tool, combined with the synonym dictionary, is used, and a maximum positive matching algorithm is employed to automatically replace all non-standard terms in the text (i.e., text feature data) with national standard terms, eliminating semantic ambiguity during retrieval.

[0030] Furthermore, the data conversion unit has a built-in unit conversion engine that uses regular expressions to match the pattern {value} + {unit} (e.g., recognizing 150 bar or 2000 kgf / cm² in text). Based on preset conversion logic (e.g., 1 bar = 0.1 MPa), it uniformly converts all non-standard units to International Standard Units (SI) and updates the original values ​​synchronously. For example, it automatically rewrites "pressure is 150 bar" in text as "pressure is 15 MPa".

[0031] In one embodiment, the system further includes a bimodal engineering knowledge base construction module, which includes: The knowledge graph construction unit is used to construct hub nodes and process parameter nodes based on structured and / or semi-structured data in the preprocessed data, extract causal logical relationships based on unstructured data in the preprocessed data, and generate a symbolic knowledge graph based on hub nodes, process parameter nodes, and causal logical relationships. The vector database construction unit is used to construct a semantic vector database based on unstructured text data in preprocessed data through semantic segmentation and vectorization methods.

[0032] In one embodiment, the knowledge graph construction unit includes: The hub node construction subunit is used to model core business objects in structured and / or semi-structured data as hub nodes, and to extract the attributes or associated entities of hub nodes in structured and / or semi-structured data. The process parameter node construction sub-unit is used to abstract process parameters in structured and / or semi-structured data into independent process parameter nodes; The causal chain extraction subunit is used to extract information from unstructured data, identify causal entities, and establish causal logical relationships between causal entities. The graph generation subunit is used to construct a symbolic knowledge graph with explicit logical associations based on hub nodes, process parameter nodes, and causal logical relationships.

[0033] Furthermore, the parallel construction of the bimodal engineering knowledge base is the core of this embodiment. It aims to integrate a symbolic knowledge graph for precise fact retrieval with a vector database for capturing deep semantic context, providing a data foundation for downstream hybrid retrieval tasks.

[0034] Furthermore, the specific execution steps of the knowledge graph construction unit are as follows: Based on Neo4j, the system constructs symbolic knowledge graphs with explicit logical relationships for cleaned structured and semi-structured data. Core entity modeling: For structured data such as welding process cards, core business objects (such as "welding process" uniquely identified by "product code + weld sequence number") are modeled as hub nodes; and discrete information in the table (such as "welding method" and "groove form") are used as their attributes or related entities.

[0035] Structured abstraction of process parameters: Unlike treating parameters as ordinary text, key process parameters (such as welding current "140~160 A") are abstracted into "process parameter" nodes with independent numerical and unit attributes. This refined modeling enables the system to perform logical operations on numerical ranges, rather than simple character matching.

[0036] Unstructured Causal Chain Extraction: For unstructured text such as defect analysis reports, the system calls the DeepSeekAPI to perform information extraction driven by a large language model. Through zero-shot learning, it automatically identifies entities such as "defect type," "potential cause," and "solution" in the text and establishes causal logical relationships between them, thereby enhancing the reasoning capabilities of the graph.

[0037] Furthermore, such as Figure 2 As shown, the welding process knowledge graph it provides includes multiple knowledge nodes and their semantic relationships.

[0038] WeldingCard is a core node in the knowledge graph, used to record complete welding process information.

[0039] The Material node represents the original material being welded. It is associated with the welding process card by using the base material (using base material) to indicate the type of base material used in the process.

[0040] The WeldingMethod node describes the welding process (such as TIG, MIG, manual arc welding, etc.). It is connected to the process card through the welding method (or process used) and indicates the specific welding process used by the welding process card.

[0041] The FillerMaterial node indicates the welding wire or electrode used in the welding process. It is associated with the process card through the filler material and indicates the specific welding wire or electrode used in that welding process card.

[0042] The EvaluationReport node is used to store quality assessment and inspection data of welding results. It is linked to the process card through the evaluation report (or "have evaluation report"), indicating that the welding process card is associated with the corresponding quality inspection and evaluation results.

[0043] In one embodiment, the vector database construction unit includes: The semantic segmentation subunit is used to segment unstructured text data into multiple text blocks using a semantically based sliding window strategy. High-dimensional vectorization subunits are used to convert text blocks into high-dimensional dense vectors; The metadata association subunit is used to bind high-dimensional dense vectors and associated metadata and store them in the semantic vector database.

[0044] Furthermore, to compensate for the shortcomings of knowledge graphs in expressing implicit knowledge (such as descriptions of operational feel and explanations of complex principles), the system constructs a high-dimensional semantic vector database in parallel (semantic context layer) based on RAGFlow. Semantic chunking strategy: The data mainly comes from long texts in process evaluation reports and standard procedures. The system adopts a semantic-based sliding window strategy (e.g., segmentation by standard chapters or process step numbers) to ensure the logical integrity of each text chunk.

[0045] High-dimensional vectorization: The advanced BGE-M3 text embedding model is used to convert the segmented text blocks into 1024-dimensional high-dimensional dense vectors (Embedding).

[0046] Metadata binding: Vectorized data, along with the original text, source document ID, chapter index, and other metadata, is stored in a vector database built on the RAGFlow framework. This not only supports efficient semantic similarity retrieval but also provides precise document location data for subsequent "answer tracing" steps.

[0047] In one embodiment, the user interaction module includes: The confidence calculation unit is used to calculate the posterior probability confidence score of the input question for different query types. The intent understanding and determination unit is used to determine the input question as an exact fact query, a fuzzy semantic query, or a complex mixed query based on a preset threshold and a posterior probability confidence score. The retrieval path selection unit is used to select, based on the judgment result, the activation of the knowledge graph retrieval path, the vector retrieval path, or the hybrid retrieval path.

[0048] Furthermore, after the user inputs an engineering / process question in natural language form through the front-end interface, the system flow enters the core stage led by the task classification agent. This agent, driven by a large language model, is primarily responsible for performing intent analysis on the user query based on posterior probability confidence. After the user inputs question Q, it is not processed by a simple classifier, but rather by a task planning agent driven by a large model. The agent does not directly output a category, but instead simulates inference paths and calculates the posterior probability scores for different retrieval strategies.

[0049] The model is required to output a posterior probability confidence score (range 0-1) indicating whether the question belongs to the "fact query" or "semantic query" category. (Factual confidence score) This measures whether the question involves specific entity attributes (e.g., "What is the current of the ER50-6?"). If it contains explicit entity and attribute terms, the score tends towards 1. Semantic confidence. The score is determined by whether the question involves causal explanation or process description (e.g., "Why do pores appear?"). If it contains words like "cause," "how," or "steps," the score tends to be 1. The routing decision formula is: ; Dynamic routing strategy: setting a threshold for decision-making (For example ): like and : If determined to be an exact fact query, only path A (knowledge graph retrieval) will be activated to obtain accurate values.

[0050] like and The query is determined to be a fuzzy semantic query. Only path B (vector retrieval) is activated to obtain an explanation of the underlying principle.

[0051] like and (i.e., mixed-type problem): It is determined to be a complex mixed query, and both paths are activated at the same time, triggering the subsequent fusion process.

[0052] For example, suppose a frontline engineer inputs the question Q: "What causes porosity during ER50-6 welding, and what is the recommended welding current?". Faced with this complex instruction that contains both "principle explanation" and "parameter query" attributes, the agent will not hastily categorize it into a single label, but will instead calculate the applicability probability of the question under different search strategies: 1. The agent identifies that "the cause of porosity" is an unstructured description of the process principle and calculates its semantic confidence score. The score tends to be high (e.g., 0.92). 2. The agent identifies that "recommended welding current" is a precise entity attribute query and calculates its factual confidence level. The scores also tend to be high (e.g., 0.95).

[0053] Based on the above calculations, the system uses the routing decision formula. A judgment will be made. Because... and All exceeded the preset dynamic judgment threshold. The system automatically identifies this problem as a "complex mixed query" task and activates a dual-path retrieval strategy. Conversely, if the user only queries "the tensile strength of ER50-6", the agent will find after calculation that... extremely high The confidence level is extremely low, thus activating only knowledge graph paths to save computing power. This dynamic scheduling mechanism, based on scenario-specific confidence assessment, gives the system flexibility and accuracy in handling complex and ever-changing problems in the engineering field.

[0054] In one embodiment, the information processing module includes a hybrid retrieval unit, which includes: The task receiving subunit is used to receive mixed path instructions issued by the user interaction module; The dual-path retrieval subunit is used to execute the knowledge graph retrieval path and the vector retrieval path in parallel to obtain the structured fact set and the relevant text fragment set; The structured injection subunit is used to inject a set of structured facts into a set of relevant text fragments to generate enhanced context; The numerical consistency verification subunit is used to perform numerical consistency verification and correction on the enhanced context.

[0055] Furthermore, the hybrid retrieval unit executes the following steps: After receiving the hybrid path instruction assigned by the task classification agent, the system enters the core cross-modal retrieval and injection fusion stage, the goal of which is to use structured facts to perform data-level calibration of unstructured context. First, the system performs parallel dual-path retrieval: In the symbolic path, the DeepSeek 2.57B model is used to translate the natural language question into Cypher statements, extracting a precise set of structured facts from the Neo4j graph. (For example, {Entity: ER50-6, Attribute: Welding Current, Value: 140~160A}); In the semantic path, the BGE-M3 model is used to generate query vectors. Based on the cosine similarity formula: Retrieve the top-K relevant text fragments from Ragflow. Subsequently, the crucial structured injection operation is performed: the system utilizes named entity recognition technology to... Locating and in the text Matching entity anchor points and the precise values ​​in the graph Explicitly inject anchor points to generate the final enhanced context. This fusion logic can be formally expressed as: This involves mandatory numerical replacement or annotation. At this point, the system will initiate a crucial "numerical consistency verification" process: 1. If the text is only a vague description (such as "moderate current"), the system will directly perform "supplementary injection" to fill the anchor point with the precise value from the spectrum (such as "140-160A"), and then provide precise data support for the context. 2. If the text contains a specific numerical value that does not match the knowledge graph (e.g., the text displays "Recommended 180A", while the knowledge graph standard value is "140-160A"), the system will determine it as a "source data conflict" and execute the "authoritative correction" strategy. The system prioritizes knowledge graph data and forcibly overwrites conflicting values ​​in the text (e.g., rewriting the original sentence as: "Recommended 180A [Correction: Current standard limit is 140~160A]").

[0056] This corrects vague semantic descriptions into precise engineering parameters, thereby completely eliminating the risk of numerical illusions caused by large models at the data input level.

[0057] In one embodiment, the information processing module further includes an answer generation unit, specifically: The answer generation and tracing subunit is used to receive the enhanced context and generate the final natural language answer. In the natural language answer generation, the data source identifier embedded in the enhanced context is converted into a preset format reference mark and appended to the corresponding parameter. The closed-loop evolutionary subunit is used to obtain feedback data on natural language answers and trigger differentiated correction strategies based on the type of feedback data to update the symbolic knowledge graph, semantic vector database, or optimize the decision logic of the user interaction module.

[0058] Furthermore, the system submits the generated "enhanced context" to the DeepSeek 2.5 7B large-scale generative language model to synthesize the final response (i.e., the natural language answer). This model generates a logically clear and numerically accurate natural language answer based on high-quality, calibrated input, along with supporting evidence of key information sources (such as citations from a national standard section). In addition, this embodiment also designs a human-machine collaborative closed-loop evolution mechanism, specifically: Generation and Source Tracing: LLM generates answers based on the fusion context (i.e., enhanced context) and mandates that data sources be tagged in the output using [Ref: KG-Node-ID] or [Ref: Doc-ID]. To prevent large models from losing the source of critical data during language reorganization, the system incorporates a "reference retention directive" into the prompt.

[0059] Its input includes a context containing injection tags, such as "current should be moderate (data source ID_1024:140-160A)".

[0060] Instruction: When generating natural language responses, the model must include the data source ID (within parentheses) as a footnote after the corresponding parameter; this must not be deleted or modified.

[0061] Output: The final answer the user sees is "Recommended welding current is 140-160A [Ref: ID_1024]". Clicking this marker will display a pop-up window showing the original node attributes of this data in the Neo4j graph or the corresponding section of the national standard document.

[0062] In one embodiment, the closed-loop evolutionary subunit includes: A feedback collection component is used to obtain explicit and implicit feedback on natural language answers; An exception log generation component is used to associate and store explicit feedback with implicit feedback, input questions, and retrieved data as exception logs. A knowledge base correction component is used to update the symbolic knowledge graph and / or semantic vector database that are confirmed as factual errors based on anomaly logs; The proxy optimization component is used to determine the optimization dataset based on the anomaly logs, and then apply the decision logic of the user interaction module based on the optimization dataset.

[0063] Furthermore, the closed-loop evolutionary subunit executes the following steps: 1. Front-end Feedback: Users click "Report Error" or enter correction suggestions for unsatisfactory answers. 2. Automatic Recording: The system captures the current question, retrieved triples / text blocks, and user feedback to generate an "Abnormal Log." 3. Expert Review: Domain experts periodically review the abnormal log. 4. Knowledge Base / Library Correction: If an error is confirmed in the knowledge base, experts directly modify entity attributes or relationships in Neo4j using developer tools, or update the vector library documentation. 5. Model Fine-tuning: Accumulated high-quality correction pairs (Question-Correction Pairs) are added to the fine-tuning dataset to periodically optimize the agent's routing logic.

[0064] Furthermore, the system automatically mines outlier data: it uses a dual mechanism of "explicit + implicit" methods to identify potential errors. 1. Explicit feedback: Users click "Report an error" or "Suggestions for improvement".

[0065] 2. Implicit Feedback: Monitor users' "rejection behavior", including regenerating answers and modifying query terms multiple times in a short period of time. The system will automatically mark such conversations as "low-quality interaction samples" and avoid outputting similar answers in the future.

[0066] Furthermore, the differentiation correction strategy is as follows: Knowledge base correction (for factual errors): For errors such as "outdated parameters", the system automatically generates a "difference detection report" by adding documents through web crawling or manual addition, and pre-generates corrected Cypher statements. Experts only need to click "confirm" to complete the hot update of the map, which greatly reduces the cost of manual operation.

[0067] Model fine-tuning and optimization (for logic errors): For errors such as "Agent routing failures," the system constructs a "preference dataset" from the accumulated cleaned abnormal samples. Developers periodically use this dataset to perform lightweight fine-tuning on the task classification agent, allowing the model to automatically "remember" the correct logic for handling similar complex scenarios through data iteration, thus achieving self-evolution of its intelligence level.

[0068] This embodiment also discloses a dual-path engineering process intelligent question-answering method based on posterior probability dynamic routing, such as... Figure 3 As shown, the specific steps are as follows: Preprocessing is performed on multi-source heterogeneous engineering data to obtain preprocessed data; Construct a symbolic knowledge graph and semantic vector database based on preprocessed data; The large language model agent performs intent understanding and task decomposition on the received input question, selects the appropriate information retrieval method, and generates a query request. Based on the query request, knowledge retrieval and answer generation are performed using symbolic knowledge graphs and / or semantic vector databases.

[0069] The names of the relevant technologies used in this embodiment are explained as follows: Large Language Models (LLMs) are deep neural network models based on the Transformer architecture, possessing billions or even more parameters. These models are pre-trained using self-supervised learning on massive and diverse text corpora to capture deep syntactic, semantic, and contextual dependencies of language.

[0070] Agent: An agent is a computational entity that perceives its environment through sensors and takes actions through actuators to achieve a predefined goal. This architecture follows a "Perceive-Plan-Act" cycle. Essentially, it empowers LLMs to go beyond simple text generation, enabling them to interact with the external world and complete complex tasks.

[0071] Retrieval-Augmented Generation (RAG) is a large-scale language model application framework designed to improve the accuracy and timeliness of generated answers by leveraging external knowledge bases. It works by retrieving relevant information from a knowledge source based on the user's question before generating an answer. This retrieved information is then used as context, along with the original question, and provided to the language model. This reduces factual errors and allows the model to answer questions based on knowledge not included in its internal training data.

[0072] Dual-Path Retrieval: Dual-path retrieval is an information retrieval architecture designed to leverage two different types of knowledge sources simultaneously for complementary advantages. The first path is the symbolic path, which uses structured queries on knowledge graphs (such as Neo4j) to retrieve accurate and reliable factual knowledge such as entities, relationships, and parameters. The second path is the semantic path, which uses similarity calculations in vector databases (such as those managed by Ragflow) to recall rich, descriptive text fragments relevant to the question context. This invention overcomes the limitations of a single retrieval mode by dynamically selecting and collaboratively applying these two paths through an agent.

[0073] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0074] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A dual-path intelligent question-answering system for engineering processes based on posterior probability dynamic routing, characterized in that, include: The data processing module is used to preprocess multi-source heterogeneous engineering data to obtain preprocessed data; The user interaction module is used to understand the intent and decompose the received input questions through a large language model agent, select an appropriate information retrieval method, and generate a query request. The information processing module is used to perform knowledge retrieval and answer generation based on the query request using a preset symbolic knowledge graph and / or semantic vector database, wherein the symbolic knowledge graph and the semantic vector database are constructed based on the preprocessed data.

2. The dual-path intelligent question-answering system for engineering processes based on posterior probability dynamic routing according to claim 1, characterized in that, The data processing module includes: The data cleaning unit is used to perform multimodal analysis on the multi-source heterogeneous engineering data, extract text feature data, and remove noise and redundant information to obtain cleaned data. The data conversion unit is used to normalize the cleaned data into standard data and generate the preprocessed data.

3. The dual-path engineering process intelligent question-answering system based on posterior probability dynamic routing according to claim 1, characterized in that, The system also includes a bimodal engineering knowledge base construction module, which includes: The knowledge graph construction unit is used to construct hub nodes and process parameter nodes based on structured data and / or semi-structured data in the preprocessed data, extract causal logical relationships based on unstructured data in the preprocessed data, and generate a symbolic knowledge graph based on the hub nodes, the process parameter nodes, and the causal logical relationships. The vector database construction unit is used to construct a semantic vector database based on the unstructured text data in the preprocessed data through semantic segmentation and vectorization methods.

4. The dual-path engineering process intelligent question-answering system based on posterior probability dynamic routing according to claim 3, characterized in that, The knowledge graph construction unit includes: The hub node construction subunit is used to model core business objects in structured and / or semi-structured data as hub nodes, and to extract the attributes or associated entities of the hub nodes in the structured and / or semi-structured data. The process parameter node construction sub-unit is used to abstract process parameters in structured and / or semi-structured data into independent process parameter nodes; The causal chain extraction subunit is used to extract information from the unstructured data, identify causal entities, and establish causal logical relationships between the causal entities. The graph generation subunit is used to construct a symbolic knowledge graph with explicit logical associations based on the hub node, the process parameter node, and the causal logical relationship.

5. The dual-path engineering process intelligent question-answering system based on posterior probability dynamic routing according to claim 3, characterized in that, The vector database construction unit includes: The semantic segmentation subunit is used to segment the unstructured text data into multiple text blocks using a semantically based sliding window strategy. A high-dimensional vectorization subunit is used to convert the text block into a high-dimensional dense vector; The metadata association subunit is used to bind the high-dimensional dense vector and the associated metadata, and store them in the semantic vector database.

6. The dual-path engineering process intelligent question-answering system based on posterior probability dynamic routing according to claim 1, characterized in that, The user interaction module includes: The confidence calculation unit is used to calculate the posterior probability confidence scores of the input question for different query types. The intent understanding determination unit is used to determine the input question as an exact fact query, a fuzzy semantic query, or a complex hybrid query based on a preset threshold and the posterior probability confidence score. The retrieval path selection unit is used to select, based on the judgment result, the activation of the knowledge graph retrieval path, the vector retrieval path, or the hybrid retrieval path.

7. The dual-path engineering process intelligent question-answering system based on posterior probability dynamic routing according to claim 6, characterized in that, The information processing module includes a hybrid retrieval unit, which includes: The task receiving subunit is used to receive the mixed path instructions issued by the user interaction module; The dual-path retrieval subunit is used to execute the knowledge graph retrieval path and the vector retrieval path in parallel to obtain the structured fact set and the relevant text fragment set; The structured injection subunit is used to inject the structured fact set into the relevant text fragment set to generate enhanced context; The numerical consistency verification subunit is used to perform numerical consistency verification and correction on the enhanced context.

8. The dual-path engineering process intelligent question-answering system based on posterior probability dynamic routing according to claim 7, characterized in that, The information processing module also includes an answer generation unit, specifically: The answer generation and tracing subunit is used to receive the enhanced context and generate the final natural language answer. In the natural language answer generation, the data source identifier embedded in the enhanced context is converted into a preset format reference mark and appended to the corresponding parameter. The closed-loop evolution subunit is used to obtain feedback data of the natural language answers and trigger differentiated correction strategies based on the type of feedback data to update the symbolic knowledge graph, semantic vector database, or optimize the decision logic of the user interaction module.

9. A dual-path intelligent question-answering system for engineering processes based on posterior probability dynamic routing as described in claim 8, characterized in that, The closed-loop evolutionary subunit includes: A feedback collection component is used to obtain explicit and implicit feedback on the natural language answer; An exception log generation component is used to associate and store the explicit feedback, the implicit feedback, the input question, and the retrieved data as an exception log. A knowledge base correction component is used to update the symbolic knowledge graph and / or semantic vector database that are confirmed as factual errors based on the anomaly logs. An agent optimization component is used to determine an optimization dataset based on the anomaly logs and to make decision-making decisions for the user interaction module based on the optimization dataset.

10. A dual-path intelligent question-answering method for engineering processes based on posterior probability dynamic routing, characterized in that, The specific steps are as follows: Preprocessing is performed on multi-source heterogeneous engineering data to obtain preprocessed data; A symbolic knowledge graph and a semantic vector database are constructed based on the preprocessed data; The large language model agent performs intent understanding and task decomposition on the received input question, selects the appropriate information retrieval method, and generates a query request. Based on the query request, knowledge retrieval and answer generation are performed using the symbolic knowledge graph and / or semantic vector database.

Images