Aircraft manufacturing process instruction intelligent generation method based on hybrid retrieval enhancement generation

Abstract

The invention provides an aircraft manufacturing process instruction intelligent generation method based on hybrid retrieval enhanced generation, and the method guarantees that a generated instruction has professional sentence patterns with historical experience and compliance indexes with standard specifications through a hybrid retrieval mechanism. A large model is limited in a specific process generation link through the agent workflow, strict placeholder filling logic is matched, the illusion problem that key parameters are pinched by the model is thoroughly eradicated, meanwhile, the parallel generation architecture enables the generation time of a long-flow process instruction to be greatly shortened, and unification of high accuracy, high controllability and high efficiency is achieved.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing, and in particular to an intelligent generation method for aircraft manufacturing process instructions based on hybrid retrieval enhancement. Background Technology

[0002] In high-end industrial fields such as aerospace and precision manufacturing, which require high complexity and precision, process instructions are key technical documents guiding on-site production operations and serve as the core link between product design and actual manufacturing. Their quality directly determines the standardization of production operations, the consistency of product manufacturing, and the controllability of final quality. With increasing product complexity, growing customization demands, and the development of multi-variety, small-batch production models, unprecedented challenges have been placed on the accuracy, standardization, executability, and efficiency of process instructions.

[0003] For a long time, the formulation of process instructions has relied heavily on the experience of senior process engineers or technicians. The drafter needs to deeply interpret product design drawings, technical standards, and process specifications, and, in conjunction with the equipment, tooling, and material resources of a specific production line, break down the manufacturing process into a series of operable and verifiable steps, accurately setting various process parameters. This process is not only time-consuming and labor-intensive (especially for complex parts, the formulation cycle can last for several weeks), but also highly dependent on personal experience and a sense of responsibility. It is prone to errors due to human negligence, differences in experience, or inconsistent understanding of standards, leading to problems such as incorrect instructions, missing parameters, vague descriptions, or non-standard formats. These problems can directly cause processing errors, quality defects, or even safety accidents on the production floor, and the subsequent proofreading, revision, and version management also bring significant additional costs.

[0004] In recent years, the global manufacturing industry has been accelerating its transformation towards digitalization, networking, and intelligence. Artificial intelligence technology, especially large language models, has brought new possibilities for the automation and intelligence of process design. The industry has begun to explore using general-purpose large language models to assist in generating natural language descriptions of process steps or for text polishing. However, practice shows that directly applying general-purpose large language models to process instruction generation faces severe challenges: On the one hand, process instructions are highly domain-specific, and their content must strictly adhere to industry standards, corporate specifications, specific equipment operating procedures, and historically validated best practices. General-purpose models lack a deep understanding and constraint of this structured, high-precision domain knowledge, easily generating a "content illusion" that seems reasonable but is actually erroneous. On the other hand, process parameters (such as processing speed, feed rate, temperature, tolerance, etc.) require extremely high numerical accuracy. General-purpose models struggle to accurately extract and apply these specific values ​​from massive amounts of unstructured training data, resulting in a very high risk of parameter errors. Therefore, the instructions generated by general models often fail to meet the stringent requirements of high-end industrial fields in terms of standardization, accuracy, and reliability. They cannot be directly applied to actual production and still require extensive and meticulous verification and correction by process experts, which greatly diminishes the value of intelligentization.

[0005] To address the aforementioned bottlenecks, this invention achieves a breakthrough innovation in the field of intelligent process instruction generation. This method does not rely on direct generation from a single, black-box large model. Instead, it constructs a hybrid retrieval mechanism integrating semantic retrieval and keyword retrieval. Leveraging intelligent agent workflow technology, it systematically breaks down the complex process instruction generation task into a set of logically clear, collaborative, and deterministic functional nodes. The hybrid retrieval mechanism can efficiently and accurately retrieve the most relevant historical experience, standard clauses, and examples from structured process knowledge bases, historical high-quality process archives, and standard specification documents, based on semantic similarity and precise keywords. The intelligent agent workflow then deeply integrates these retrieval results with product design inputs (such as 3D model features and BOM information). Through a series of sequential or parallel intelligent nodes (such as process route reasoning nodes, detailed process design nodes, parameter calculation and filling nodes, and specification review nodes), under strong domain knowledge constraints, it gradually generates process instructions that are structurally complete, parameter-accurate, and formatted correctly. This method effectively integrates the reuse of historical experience, rigid constraints of standard specifications, and precise filling of structured parameters, curbing content illusion and parameter errors at the source, significantly improving the accuracy, standardization, and consistency of process instruction generation, and greatly enhancing the automation level and efficiency of compilation, providing a reliable and feasible solution for intelligent process design in high-end manufacturing. Summary of the Invention

[0006] The present invention relates to an intelligent generation method for aircraft manufacturing process instructions based on hybrid retrieval enhancement, comprising the following steps: Construct a multi-source heterogeneous process knowledge base. In view of the complex data sources in the process field, we will build a structured parameter library to store the precise process numerical information of parts, a historical experience vector library to store fragmented and quantified historical high-quality process document slices, and establish a full-text index library. The system designs a context enhancement mechanism based on hybrid retrieval. When generating each process instruction, the system performs dual-path retrieval in parallel. On the one hand, it retrieves the most semantically similar historical process fragments from the vector library based on the structural features of the current part and the process name to obtain mature sentence templates and parameter placeholder structures. On the other hand, it retrieves relevant technical specification clauses from the standard library based on the process name to obtain hard technical requirements. The dual-path retrieval results are then dynamically assembled into prompt words. A state-graph-based intelligent agent workflow is constructed. This workflow includes an initialization node, a routing decision mechanism, a parallel generation node, and an intelligent fill-in node. The generation task is broken down into a finely detailed pipeline. The initialization node parses requests and loads templates. The routing decision mechanism dynamically selects the "full reuse" or "step-by-step generation" path based on the retrieval results. The parallel generation node drives the large model to independently generate instruction templates for each process under strict constraints. Finally, the intelligent fill-in node uses real data from a structured parameter library to automatically replace placeholders in the template with actual parameters through regular expression matching and logical reasoning.

[0007] In one implementation, the fragmentation and vectorization of unstructured historical process documents includes the following steps: Step 111: Use the Python-docx tool to read historical process documents, identify the table structure in the documents, and extract metadata including "process name", "process content" and the corresponding "part drawing number" and "part name".

[0008] Step 112: Using the "semantic integrity segmentation" strategy, each independent "process" is taken as the smallest segmentation unit, and the entire operation description text of the process is taken as a slice; at the same time, the part structure information to which the process belongs is attached as metadata to the slice. Step 113: Load the pre-trained Embedding model and vectorize the "process name" and "process content" of each process slice to generate a high-dimensional dense vector; Step 114: Store the generated vector and its corresponding metadata into the Qdrant vector database; during storage, create a Payload index for the "Part Structure" and "Process Name" fields.

[0009] In one implementation, building a full-text index database includes the following steps: Step 121: Parse the manufacturing process standard document, identify and separate the narrative text area and the tabular data area in the document; for the narrative text area, use a sliding window of preset length to segment it, and retain the overlap of the preset length between adjacent segmented paragraphs; for the tabular data area, extract it completely in units of tables, and convert it into a structured data format that retains the logical relationship between rows and columns. Step 122: Import the segmented text into the search engine and build an inverted index; set relevance ranking rules to ensure the accuracy of search results.

[0010] In one implementation, the multi-source heterogeneous process knowledge base includes a structured part parameter library, and the cleaning and storage of the structured part parameter library includes the following steps: Step 131: Read the Excel file containing part information, and merge the attribute fields scattered in different worksheets using "part drawing number" as the unique identifier Key. Step 132: Convert the merged part data into JSON objects and store them in a MongoDB database; each document corresponds to a part and contains all its attribute information, forming a flat "data dictionary".

[0011] In one implementation, the intelligent agent workflow initialization based on the state diagram, global state definition, and request parsing include the following steps: Step 211: Define a globally shared dictionary structure as the workflow's "state context"; the "state context" contains the following key fields: input fields, intermediate fields, and output fields; the "state context" will flow and be updated between all nodes in the workflow. Step 212, Request Parsing: The system receives the JSON request sent by the user via API, extracts the "part drawing number" and the "process list" to be generated; if the user does not provide a process list, the system will automatically load the default template based on the subsequently identified part structure.

[0012] In one implementation, the agent workflow initialization based on the state diagram includes the following steps when executing the initialization node: Step 221: Automatically query the MongoDB database based on the "part drawing number"; if the query is successful, load all process attribute information of the part into the "part details" field of the status context; if the query fails, mark the error and terminate the process. Step 222: Extract the "part structure" field from the loaded part information; if the user does not specify a list of processes in the request, the system will load the corresponding standard process template from the preset configuration center according to the structure type and update it in the status context.

[0013] In one implementation, executing the routing decision logic includes the following steps: Step 231: The system calls the counting interface of the Qdrant vector database to check if there is a historical process record that is completely consistent with the current "part drawing number". This check aims to identify whether the current task belongs to the "instruction upgrade" or "revision" scenario, that is, the part has been programmed before, and this time only needs to be fine-tuned or reissued on the original basis.

[0014] Step 232, Dynamic route distribution: Path A: If a completely identical historical record exists (i.e., count > 0), the system determines that the current task is an instruction upgrade scenario, marks the status as is_exact_match=True, and the workflow will point to the "full reuse node". In this mode, the system directly retrieves all the process content of the historical version as a benchmark, avoiding the inefficiency caused by AI regeneration and ensuring the inheritance and stability between versions.

[0015] Path B: If no completely identical historical records exist, the system determines that the current task is a new part creation scenario, marks the status as is_exact_match=False, and the workflow will point to the "parallel step-by-step generation node". In this mode, the system will launch a hybrid retrieval and large model generation engine to synthesize new process instructions based on the experience of similar parts.

[0016] In one implementation, performing parallel step-by-step node generation includes the following steps: Step 31: The system reads the "process list" in the state context and breaks it down into N independent subtasks; each subtask contains the ID, name and part structure type of the current process; the system initializes a thread pool and distributes these N subtasks to different threads for parallel execution. Step 32: For each subtask, before generating the instruction content, the system first performs parallel dual-path knowledge retrieval to obtain accurate contextual references. Step 33: The system dynamically assembles the retrieved multi-source information into a structured prompt word and inputs it into the central large language model; Step 34: The large model inference generates text and performs post-processing.

[0017] The beneficial effects of this invention are: This invention provides an intelligent process instruction generation method suitable for high-reliability process scenarios. Through a hybrid retrieval mechanism, it ensures that the generated instructions have both professional sentence structures based on historical experience and compliant indicators of standard specifications. By using an intelligent agent workflow to restrict the large model to specific process generation stages and with strict placeholder filling logic, it completely eliminates the illusion problem of the model fabricating key parameters. At the same time, the parallel generation architecture significantly shortens the generation time of long-process instructions, achieving a balance of high accuracy, strong controllability, and high efficiency. Attached Figure Description

[0018] To more clearly illustrate the technical solutions implemented in this invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this invention. For those skilled in the art, other ideas can be obtained from these drawings without creative effort.

[0019] Figure 1 This is the overall system architecture diagram; Figure 2 This is a flowchart of the construction process of a multi-source heterogeneous process knowledge base; Figure 3 It is a workflow state transition diagram of the intelligent agent; Figure 4 It is a step-by-step logic diagram generated based on hybrid retrieval instructions; Figure 5 This is a schematic diagram of the structured parameter intelligent filling principle. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0021] The compilation of process instructions mainly relies on manual editing, and the use of large language models to assist intelligent process design is still in the exploratory stage.

[0022] Step 1: Construct a multi-source, heterogeneous process knowledge base. Addressing the characteristics of dispersed data sources and diverse formats in the process field, this invention designs a complete data preprocessing and data entry process, such as... Figure 1 , Figure 2 As shown, unstructured historical documents, semi-structured standard specifications, and structured part parameters are transformed into machine-understandable and searchable knowledge forms.

[0023] Step 11: Fragmentation and vectorization of unstructured historical process documents. Historical process documents (usually in Word or PDF format) contain a wealth of proven and mature process experience, which forms the basis for generating high-quality instructions.

[0024] Step 111: Use tools such as Python-docx to read historical process documents and remove irrelevant information such as headers, footers, and watermarks. Identify the table structure in the document and extract metadata such as "process name," "process content," and the corresponding "part drawing number" and "part name."

[0025] Step 112 abandons the traditional method of segmenting by a fixed number of characters and adopts a "semantic integrity segmentation" strategy. Each independent "process" is used as the smallest segmentation unit, and the entire operational description text of that process is treated as a chunk. Simultaneously, the structural information of the part to which that process belongs is appended as metadata to the chunk for subsequent precise filtering.

[0026] Step 113: Load the pre-trained Chinese Embedding model, and vectorize the "process name" and "process content" of each process slice to generate a high-dimensional dense vector.

[0027] Step 114: Store the generated vector and its corresponding metadata into the Qdrant vector database. During storage, create payload indexes for the "Part Structure" and "Process Name" fields to support subsequent mixed filtering queries.

[0028] Step 12: Constructing a full-text index of standard and specification documents. Process standards and technical specifications typically contain stringent parameter indicators, making them suitable for precise keyword searches.

[0029] Step 121: Parse the manufacturing process standard document, identify and separate the narrative text area and the tabular data area in the document; for the narrative text area, use a sliding window of preset length to segment it, and retain the overlap of the preset length between adjacent segmented paragraphs; for the tabular data area, extract it completely in units of tables, and convert it into a structured data format that retains the logical relationship between rows and columns.

[0030] Step 122: Import the segmented text into the MeiliSearch search engine and build an inverted index. Set relevance ranking rules to ensure the accuracy of search results.

[0031] Step 13: Cleaning and importing the structured parts parameter library. The geometric dimensions, material grades, and other data of the parts are usually stored in Excel spreadsheets or ERP systems, which serve as the basis for subsequent parameter filling.

[0032] Step 131: Read the Excel file containing part information, and merge the attribute fields scattered across different worksheets using the "part drawing number" as the unique identifier (Key). Clean the data, removing blank values ​​and correcting formatting errors.

[0033] Step 132: Convert the cleaned part data into JSON objects and store them in a MongoDB database. Each document corresponds to one part and contains all its attribute information, forming a flat "data dictionary" for quick retrieval via key-value pairs later.

[0034] Step 2, Initialization of the agent workflow based on the state graph. This invention adopts an orchestration approach based on graph theory to define an agent workflow that includes states, nodes, and edges, as follows: Figure 3 As shown. When the system receives a generation request, it first initiates the workflow initialization phase.

[0035] Step 21, Global State Definition and Request Resolution.

[0036] Step 211: Define a globally shared dictionary structure as the workflow's "state context." This state contains the following key fields: input field (part drawing number), intermediate fields (part details, part structure type, operation list, and whether it is a complete match), and output fields (instruction template and final instruction). This state will flow and be updated across all nodes of the workflow.

[0037] Step 212, Request Parsing: The system receives the JSON request sent by the user via API, extracts the "part drawing number" and the "process list" to be generated. If the user does not provide a process list, the system will automatically load a default template based on the subsequently identified part structure.

[0038] Step 22: Execute the initialization node. This node is the starting point of the workflow, responsible for data preparation and environment awareness, and does not involve calling large models.

[0039] Step 221: Automatically query the MongoDB database based on the "Part Drawing Number". If the query is successful, load all process attribute information of the part into the "Part Details" field of the status context; if the query fails, mark an error and terminate the process.

[0040] Step 222: Extract the "Part Structure" field from the loaded part information. If the user does not specify a list of processes in the request, the system will load the corresponding standard process flow template from the preset configuration center based on the structure type and update it in the status context.

[0041] Step 23: Execute the routing decision logic. After initializing the nodes, the workflow enters the conditional branch judgment stage, determining the subsequent generation strategy based on the business scenario.

[0042] Step 231: The system calls the counting interface of the Qdrant vector database to check if there is a historical process record that is completely consistent with the current "part drawing number". This check aims to identify whether the current task belongs to the "instruction upgrade" or "revision" scenario, that is, the part has been programmed before, and this time only needs to be fine-tuned or reissued on the original basis.

[0043] Step 232, Dynamic route distribution: Path A: If a completely identical historical record exists (i.e., count > 0), the system determines that the current task is an instruction upgrade scenario, marks the status as is_exact_match=True, and the workflow will point to the "full reuse node". In this mode, the system directly retrieves all the process content of the historical version as a benchmark, avoiding the inefficiency caused by AI regeneration and ensuring the inheritance and stability between versions.

[0044] Path B: If no completely identical historical records exist, the system determines that the current task is a new part creation scenario, marks the status as is_exact_match=False, and the workflow will point to the "parallel step-by-step generation node". In this mode, the system will launch a hybrid retrieval and large model generation engine to synthesize new process instructions based on the experience of similar parts.

[0045] Step 3: Execute the parallel step-by-step generation node. When the routing decision enters the "new part instruction compilation mode," the system starts this node. Multi-threading technology is used to process each process in parallel, significantly improving generation efficiency.

[0046] Step 31, Task Decomposition and Parallel Distribution. The system reads the "process list" from the state context and decomposes it into N independent subtasks. Each subtask contains the ID, name, and part structure type of the current process. The system initializes a thread pool and distributes these N subtasks to different threads for parallel execution.

[0047] Step 32: Perform dual-path hybrid retrieval. For each subtask, before generating the instruction content, the system first performs parallel dual-path knowledge retrieval to obtain accurate contextual references, such as... Figure 4 As shown.

[0048] Step 321: The system invokes the Embedding model to convert the current "process name" into a query vector. Subsequently, a similarity search is performed in the Qdrant vector database. During the search, filtering conditions are applied, searching only in historical data with similar "part structures," and sorting by cosine similarity to return the top-3 most similar historical process segments. This step aims to obtain mature sentence templates and parameter placeholder structures.

[0049] Step 322: The system extracts "process name" as a keyword and performs a full-text search in the MeiliSearch standard library index. The system returns the most relevant standard specification clauses or technical requirement fragments. This step aims to obtain rigid process indicators and operating procedures.

[0050] Step 33: Construct dual-constraint prompt words. The system dynamically assembles the retrieved multi-source information into a structured prompt word, which is then input into the central large language model.

[0051] Step 34, Large Model Generation and Post-processing.

[0052] Step 341: Invoke the large language model to generate the specific instruction content for the current process based on the above Prompt.

[0053] Step 342: The system receives the model's text output and uses regular expressions to remove any possible thought chain markers, such as... <think>Remove tags or redundant Markdown formatting to extract the clean instruction text.

[0054] Step 343, Template Aggregation: After all subtasks are completed, the system reassembles the generated results of N processes in their original order to form an instruction template with placeholders and updates it to the state context.

[0055] Step 4: Execute the intelligent fill-in node. This node, as the final step in the workflow, is responsible for converting the instruction template with abstract placeholders generated in Step 3 into the final process instruction containing specific values. This invention designs a two-level fill-in mechanism combining "regular expression direct fill" and "reasoning fill" (e.g., Figure 5 As shown in the figure, ensure the accuracy and coverage of parameter filling.

[0056] Step 41, Placeholder Extraction and Classification. The system iterates through each process in the instruction template, automatically extracting all placeholders to be filled using regular expressions, such as {part drawing number}, {material grade}, etc. The system compares the extracted placeholders with the keys of the "part details" dictionary loaded in MongoDB, classifying them into two categories: direct matching: the placeholder name exists directly in the part attribute dictionary; inference required: the placeholder name cannot be found directly in the dictionary and needs to be inferred from the context.

[0057] Step 42, Level 1 Filling: Rule-based Deterministic Replacement. For "Direct Match Class" placeholders, the system performs a deterministic string replacement operation.

[0058] Step 421, exact match: directly replace the placeholder with the Value from the dictionary.

[0059] Step 422, Fuzzy matching of suffixes: If an exact match fails, the system attempts to match the suffix of the dictionary key. This mechanism enhances the system's compatibility with different naming conventions.

[0060] Step 43, Second-level filling: Logical reasoning based on the large model. For the remaining "required reasoning" placeholders after the first-level filling, the system initiates the second-level filling process.

[0061] Step 431, Constructing the Inference Prompt: The system takes the "List of Remaining Placeholders" and the "Part Details Dictionary" as input to construct a dedicated inference prompt. This prompt instructs the large model: "Based on the provided data dictionary, infer what values ​​should be filled in the following placeholders." Step 432, JSON Mapping Generation: The large model analyzes the logical relationships between the data and outputs a JSON-formatted key-value pair mapping table.

[0062] Step 433, Secondary Replacement: The system parses the JSON returned by the model, fills the corresponding values ​​back into the instruction template, and completes the final parameter instantiation.

[0063] Step 44: Result Encapsulation and Streaming Output. After two levels of data filling, the system obtains a complete and logically coherent process instruction. The system encapsulates this instruction into a standard JSON response format and pushes it to the front-end interface via the Server-Sent Events (SSE) protocol. Users can view the generated process instructions line by line in real time on the interface and perform final manual confirmation and release.< / think>

Claims

1. A method for intelligently generating aircraft manufacturing process instructions based on hybrid retrieval and enhanced generation, characterized in that, Includes the following steps: Construct a multi-source heterogeneous process knowledge base. In view of the complex data sources in the process field, we will build a structured parameter library to store the precise process numerical information of parts, a historical experience vector library to store fragmented and quantified historical high-quality process document slices, and establish a full-text index library. The system designs a context enhancement mechanism based on hybrid retrieval. When generating each process instruction, the system performs dual-path retrieval in parallel. On the one hand, it retrieves the most semantically similar historical process fragments from the vector library based on the structural features of the current part and the process name to obtain mature sentence templates and parameter placeholder structures. On the other hand, it obtains the hard technical requirements based on the retrieval results of the process name in the full-text index library and dynamically assembles the dual-path retrieval results into the prompt words. A state-graph-based intelligent agent workflow is constructed. This workflow includes an initialization node, a routing decision mechanism, a parallel generation node, and an intelligent fill-in node. The generation task is broken down into a fine-grained pipeline. The initialization node parses requests and loads templates. The routing decision mechanism dynamically selects the "full reuse" or "step-by-step generation" path based on the search results. The parallel generation node drives the large model to independently generate instruction templates for each process under strict constraints. Finally, the intelligent fill-in node uses real data from a structured parameter library to automatically replace placeholders in the template with actual parameters through regular expression matching and logical reasoning.

2. The method according to claim 1, characterized in that, The fragmentation and vectorization of unstructured historical process documents includes the following steps: Step 111: Use the Python-docx tool to read historical process documents, identify the table structure in the documents, and extract metadata including "process name", "process content" and the corresponding "part drawing number" and "part name"; Step 112: Using the "semantic integrity segmentation" strategy, each independent "process" is taken as the smallest segmentation unit, and the entire operation description text of the process is taken as a slice; at the same time, the part structure information to which the process belongs is attached as metadata to the slice. Step 113: Load the pre-trained Embedding model and vectorize the "process name" and "process content" of each process slice to generate a high-dimensional dense vector; Step 114: Store the generated vector and its corresponding metadata into the Qdrant vector database; during storage, create a Payload index for the "Part Structure" and "Process Name" fields.

3. The method according to claim 1, characterized in that, Building a full-text index includes the following steps: Step 121: Parse the manufacturing process standard document, identify and separate the narrative text area and the tabular data area in the document; for the narrative text area, use a sliding window of preset length to segment it, and retain the overlap of the preset length between adjacent segmented paragraphs; for the tabular data area, extract it completely in units of tables, and convert it into a structured data format that retains the logical relationship between rows and columns. Step 122: Import the segmented text into the search engine and build an inverted index; set relevance ranking rules to ensure the accuracy of search results.

4. The method according to claim 1, characterized in that, The multi-source heterogeneous process knowledge base includes a structured part parameter library. The cleaning and storage of the structured part parameter library includes the following steps: Step 131: Read the Excel file containing part information, and merge the attribute fields scattered in different worksheets using "part drawing number" as the unique identifier key. Step 132: Convert the merged part data into JSON objects and store them in a MongoDB database; each document corresponds to a part and contains all its attribute information, forming a flat "data dictionary".

5. The method according to claim 1, characterized in that, The initialization of a state-graph-based agent workflow, including global state definition and request parsing, includes the following steps: Step 211: Define a globally shared dictionary structure as the workflow's "state context"; the "state context" contains the following key fields: input fields, intermediate fields, and output fields; the "state context" will flow and be updated between all nodes in the workflow. Step 212, Request parsing: The system receives the JSON request sent by the user via API, extracts the "part drawing number" and the "process list" to be generated; if the user does not provide a process list, the system will automatically load the default template based on the subsequently identified part structure.

6. The method according to claim 1, characterized in that, The initialization of a state-graph-based agent workflow includes the following steps in executing the initialization node: Step 221: Automatically query the MongoDB database based on the "part drawing number"; if the query is successful, load all process attribute information of the part into the "part details" field of the status context; If the query fails, mark the error and terminate the process; Step 222: Extract the "part structure" field from the loaded part information; If the user does not specify a list of procedures in the request, the system will load the corresponding standard procedure template from the preset configuration center according to the structure type and update the status context.

7. The method according to claim 1, characterized in that, Executing routing decision logic includes the following steps: Step 231: The system calls the counting interface of the Qdrant vector database to check if there is a historical process record that is completely consistent with the current "part drawing number". This check aims to identify whether the current task belongs to the "instruction upgrade" or "revision" scenario, that is, the part has been programmed before, and this time only needs to be fine-tuned or reissued on the original basis; Step 232, Dynamic route distribution: Path A: If a completely identical historical record exists (i.e., count > 0), the system determines that the current task is an instruction upgrade scenario, marks the status as is_exact_match=True, and the workflow will point to the "full reuse node". In this mode, the system directly retrieves all the process content of the historical version as the benchmark, avoiding the inefficiency caused by AI regeneration and ensuring the inheritance and stability between versions; Path B: If no completely identical historical records exist, the system determines that the current task is a new part creation scenario, marks the status as is_exact_match=False, and the workflow will point to the "parallel step-by-step generation node". In this mode, the system will launch a hybrid retrieval and large model generation engine to synthesize new process instructions based on the experience of similar parts.

8. The method according to claim 1, characterized in that, Performing parallel step-by-step node generation includes the following steps: Step 31: The system reads the "process list" in the state context and breaks it down into N independent subtasks; each subtask contains the ID, name and part structure type of the current process; the system initializes a thread pool and distributes these N subtasks to different threads for parallel execution. Step 32: For each subtask, before generating the instruction content, the system first performs parallel dual-path knowledge retrieval to obtain accurate contextual references. Step 33: The system dynamically assembles the retrieved multi-source information into a structured prompt word and inputs it into the central large language model; Step 34: The large model inference generates text and performs post-processing.

Images