Collaborative bom data analysis and verification system and method based on ai model

Abstract

The application relates to the technical field of BOM data analysis, and discloses a collaborative BOM data analysis and checking system and method based on an AI model, wherein the collaborative BOM data analysis and checking system based on the AI model comprises an intelligent document perception and structure analysis module, the intelligent document perception and structure analysis module performs listening and identification on a list file, and outputs a unified format document, and an AI assistant dialogue engine module adopts a language model to interact with an engineer to analyze the unified format document field by field. The application introduces a field-level approval mechanism, each suggestion needs to be confirmed by a user before being modified, the user can perform operations such as rejection, remarks and pending, instead of only reviewing in the last stage, the user's right, safety and traceability are increased, the responsibility boundary of AI-assisted design is clear, AI evolves from a tool into an engineering partner with learning ability, the interactive intelligent experience is significantly enhanced, and the application has strong adaptability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of BOM data analysis and verification, in particular to a collaborative BOM data analysis and verification system and method based on an AI model. BACKGROUND

[0002] A bill of materials, commonly abbreviated BOM, is a technical document that describes the composition of a product. In the processing capital industry, it indicates the structural relationship between the total assembly, subassembly, component, part, and raw material of the product, as well as the required quantity. In the chemical, pharmaceutical, and food industries, it describes the main raw materials, intermediates, auxiliary materials, and their formulations, as well as the required quantities. BOM is a data table representation of product composition instead of a graphical representation. It is an important control file in the MRP process of the MRPII system.

[0003] In the development process of electronic products, design engineers need to prepare single-board BOM lists and product hardware lists for subsequent procurement, manufacturing, assembly, and testing. However, due to the large number of hardware BOM lists, large quantities, non-uniform formats, and human error in filling out forms, BOM lists often have various problems. Traditional manual checking methods are inefficient and prone to errors, which is not conducive to batch rapid review. Existing tools on the market have fixed logic and cannot handle complex exceptions. AI-based large language models (such as CN119004372A) are mostly limited to automatic analysis and lack interactive mechanisms with engineers, modification and approval processes, user preference memory mechanisms, and other key designs, making it difficult to cover the entire engineering decision-making process. SUMMARY

[0004] The present application provides a collaborative BOM data analysis and verification system and method based on an AI model, which has the beneficial effect of interacting with engineers and covering the entire engineering decision-making process. It solves the problem of not being conducive to batch rapid review, being limited to automatic analysis, lacking interactive mechanisms with engineers, modification and approval processes, user preference memory mechanisms, and other key designs, and being difficult to cover the entire engineering decision-making process.

[0005] The present application provides the following technical solution: a collaborative BOM data analysis and verification system based on an AI model, comprising:

[0006] An intelligent document perception and structure analysis module, which listens to and identifies list files and outputs them as unified format documents;

[0007] An AI assistant dialogue engine module that uses a language model to interact with engineers to analyze unified format documents field by field to assist engineers in understanding the content of list files;

[0008] A semantic classification and field specification module, through which a semantically ambiguous field in the uniform format document is marked, and the AI assistant dialogue engine module is triggered to issue suggestion information for confirmation by an engineer;

[0009] An intelligent verification and exception diagnosis module, which interfaces with a field in a device database, verifies and diagnoses the uniform format document according to the field in the device database, and generates a corresponding suggestion record for an engineer to process the suggestion record;

[0010] A deployment and resource scheduling optimization module, which integrates the AI assistant dialogue engine module and a language model embedding, and the integrated AI assistant dialogue engine module analyzes the uniform format document in an off-network and low-power-consumption environment.

[0011] As an optional solution of the AI model-based collaborative BOM data analysis and verification system, the AI assistant dialogue engine module includes a user interaction layer, a semantic intermediate processing layer, and an inference and response layer;

[0012] The user interaction layer provides a natural language interaction window for users by using PySide2;

[0013] The semantic intermediate processing layer analyzes the uniform format document by using a session state manager, a Prompt builder, and an instruction intent recognizer;

[0014] The inference and response layer dynamically adds labels to the analyzed uniform format document in real time by using a model service interface and a streaming output manager.

[0015] As an optional solution of the AI model-based collaborative BOM data analysis and verification system, the model service interface uses the mistral.cpp framework to load the language model in the GGUF format to the local, and uses INT4 quantization to optimize the inference efficiency;

[0016] The model service interface automatically switches and hot loads the language model based on the context or task type of the uniform format document.

[0017] As an optional solution of the AI model-based collaborative BOM data analysis and verification system, the streaming output manager pushes the token results returned by the language model to the AI assistant dialogue engine module through SSE based on the Streaming Token inference mechanism;

[0018] The AI assistant dialogue engine module dynamically adds labels in real time according to the token results.

[0019] As an optional solution of the collaborative BOM data analysis and verification system based on the AI model, the deployment and resource scheduling optimization module comprises a model running interface layer and an inference scheduling and cache layer.

[0020] The model running interface layer comprises a model loader for loading a GGUF format weight file of a local language model into a memory, a memory mapping mechanism for realizing efficient access and on-demand calling of the weight file, and a local inference engine of the language model compiled as a static link library and embedded into a desktop application front-end program.

[0021] The inference scheduling and cache layer comprises a thread pool scheduling mechanism, encapsulating each inference request as an asynchronous task object and submitting it to a thread pool managed by ThreadPoolExecutor for concurrent execution, and scheduling and processing all request entries according to task priority.

[0022] As an optional solution of the collaborative BOM data analysis and verification system based on the AI model, the deployment and resource scheduling optimization module further comprises a model management and upgrading subsystem layer and a local log and permission control subsystem layer.

[0023] The model management and upgrading subsystem layer comprises automatically finding a language model file conforming to a format through a specified model folder when a user switches multiple language models, unloading a current model weight and clearing a related memory mapping through a resource release interface when dynamically replacing the current loaded language model, subsequently reloading a new GGUF model file, and refreshing an inference context cache and a user session record.

[0024] The local log and permission control subsystem layer is used for writing each round of dialogue record into a local database.

[0025] As an optional solution of the collaborative BOM data analysis and verification system based on the AI model, the system further comprises an approval mechanism and a state tracking module, and the approval mechanism and the state tracking module establish an auxiliary decision-making mechanism based on combining suggestion information issued by the AI assistant dialogue engine module with user engineering behaviors.

[0026] The approval mechanism and the state tracking module comprise a suggestion state data structure submodule, an approval interaction interface submodule, and a state marking visualization submodule.

[0027] The suggestion state data structure submodule records and controls a whole life cycle state change process of AI generated suggestions, records all changes into a local log database, and forms a suggestion state flow history.

[0028] The approval interaction interface submodule allows the user to trigger state change behavior through button clicks or natural language instruction triggers by providing button operation or dialogue instruction trigger mechanisms.

[0029] The state marker visualization submodule is used to visually present the current processing state of each suggestion in the human-computer interaction interface.

[0030] As an optional solution of the AI model-based collaborative BOM data analysis and verification system, the memory mechanism and individualized adaptation module are further included, and the memory mechanism and individualized adaptation module are used to train the AI assistant in the AI assistant dialogue engine module.

[0031] The memory mechanism and individualized adaptation module are used to track and record the feedback behavior of the user to the system generated suggestions, form a data basis for preference learning and behavior modeling, automatically inject user preferences and historical context into large model prompts, sort results, sort and optimize multiple suggestions or alternative devices or explanation content returned by the large model, and build a long-term preference modeling and behavior memory storage for the user, so that the AI assistant in the AI assistant dialogue engine module has intelligent behavior with memory.

[0032] As an optional solution of the AI model-based collaborative BOM data analysis and verification system, the memory mechanism and individualized adaptation module are further included, and the memory mechanism and individualized adaptation module are used to train the AI assistant in the AI assistant dialogue engine module.

[0033] The main table view, dialogue window, approval window and graphic window data are interlinked, and clicking a field automatically links other windows to update.

[0034] The processing method of the AI model-based collaborative BOM data analysis and verification system includes the following steps:

[0035] S1: When opening the hardware BOM list file, automatically activate the AI assistant window and read and parse the file, initialize the user session, read the user memory configured field preferences, operation habits and project context, and generate a unified data structure for the interaction module to call;

[0036] S2: Process the markers of the file through three triggering methods, including user questioning, clicking fields and automatic checking;

[0037] S3: Generate model input through the Prompt builder, dynamically construct customized prompts according to the combined field content, user preferences and project memory;

[0038] S4: Local large language model running and adjusted model parameter running;

[0039] S5: Generate field explanation, check judgment, modification suggestion and alternative device recommendation answers through AI assistant, which are output in streaming token form, supporting user "interrupt generation" and "reask";

[0040] S6: User approval mechanism is triggered, and "confirm / reject" operation is performed through a button or a dialogue, and the user operation triggering behavior collection module is recorded at the same time, the approval result is marked to the hardware BOM file and written into the audit log record, and the user's operation behavior and preference trend are recorded, and the long-term structure preference is extracted;

[0041] S7: Semantic optimization and response sorting are performed on the reference memory of the subsequent dialogue, the response result order is optimized according to the historical behavior, the user portrait is constructed, and the long-term collaborative efficiency optimization is formed;

[0042] S8: Export the application approval log, the personalized preference log, and the result file model record file.

[0043] The application has the following beneficial effects:

[0044] 1. The collaborative BOM data analysis and verification system and method based on the AI model introduce a field-level approval mechanism, each suggestion needs to be confirmed by the user before being modified, and the user can perform operations such as rejection, note, and pending, instead of only reviewing in the last stage, thereby increasing user sovereignty, security, and traceability, realizing clear responsibility boundary of AI-assisted design, evolving AI from a tool to an engineering partner with learning ability, significantly enhancing interactive intelligent experience, and being highly adaptable.

[0045] 2. The collaborative BOM data analysis and verification system and method based on the AI model support local off-network operation, can run on an enterprise intranet or a local terminal, ensure data security, compared with the existing system which relies on cloud API calling and has a data leakage risk, solve the enterprise landing problem of AI in BOM processing, have industry deployability, and are particularly suitable for confidential research and development environment, create a model for engineering-level system design, and are different from general natural language analysis tools.

[0046] 3. The collaborative BOM data analysis and verification system and method based on the AI model use a front-end embedded assistant architecture, have independent workflows and operation states, break through traditional command-type AI interaction, actively perceive the files opened by the user in the assistant mode, and follow in real time, and simulate the dialogue review behavior of real engineers. DETAILED DESCRIPTION

[0047] Figure 1 It is a flowchart of an AI collaborative BOM data analysis and auxiliary verification system.

[0048] Figure 2Flowchart for AI assistant dialogue engine module.

[0049] Figure 3 Flowchart for local deployment and resource scheduling optimization module.

[0050] Figure 4 Flowchart for approval mechanism and state tracking module.

[0051] Figure 5 Flowchart for memory mechanism and personalized adaptation module.

[0052] Figure 6 Flowchart for overall AI model-based collaborative BOM data analysis and auxiliary verification method. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0054] Embodiment 1

[0055] Please refer to Figures 1-6 The AI model-based collaborative BOM data analysis and verification system and method includes:

[0056] The intelligent document perception and structure analysis module listens to and identifies the list file, and outputs a unified format document;

[0057] The BOM and hardware list files are listened to. If a user opens a BOM or hardware list file, the file is automatically identified (based on field keywords, format characteristics, etc.). The openpyxl or pandas is used to analyze the file content, supports merging cells, and restores the structure of the transformed table, outputs a standard two-dimensional data structure, and unifies the output format to a DataFrame structure for downstream module calling;

[0058] It should be particularly noted that pandas is a tool based on NumPy, which is created to solve data analysis tasks. Openpyxl is a Python library that focuses on processing modern Excel files, mainly used for automated operation of Excel documents;

[0059] The AI assistant dialogue engine module uses a language model to interact with engineers to analyze unified format documents field by field to assist engineers in understanding the content of the list file;

[0060] The AI assistant dialogue engine module is embedded in the front-end dialogue window, supports mixed Chinese and English questions, field clicking to trigger context explanation, and fuzzy query instructions such as "What is abnormal about this line?";

[0061] It should be particularly noted that the language model is the open source Mistral 7B model, which is run in a local PC off-network deployment mode, and the dialogue interaction process is realized based on a context memory buffer mechanism, supporting multi-round semantic interaction and state continuous tracking.

[0062] Specifically, the AI assistant dialogue engine module includes a user interaction layer, a semantic intermediate processing layer, and an inference and response layer.

[0063] The user interaction layer provides a natural language interaction window for users using PySide2.

[0064] The semantic intermediate processing layer parses the unified format document through a session state manager, a Prompt builder, and an instruction intent recognizer.

[0065] The inference and response layer uses a model service interface and a streaming output manager to add labels to the parsed unified format document in real time and dynamically.

[0066] The model service interface uses the mistral.cpp framework to load the language model in GGUF format to the local, and uses INT4 quantization to optimize the inference efficiency.

[0067] The model service interface automatically switches and hot loads the language model based on the context or task type of the unified format document.

[0068] The streaming output manager pushes the token results returned by the language model to the AI assistant dialogue engine module through SSE based on the Streaming Token inference mechanism.

[0069] SSE stands for Server-Sent Events, which is a one-way communication mechanism based on HTTP protocol, used for the server to actively push real-time data streams to the client (usually a web browser). SSE only supports the server pushing data to the client (one-way).

[0070] The AI assistant dialogue engine module dynamically adds labels in real time based on the token results.

[0071] Among them, the AI assistant dialogue engine module is responsible for the core subsystem of human-computer natural language interaction, and its goal is to build a floating window type intelligent assistant embedded in the desktop, which assists engineers in processing BOM and hardware list files, and performs interactive operations such as field interpretation, abnormal diagnosis, suggestion approval, etc. through dialogue. Each sub-module adopts asynchronous calling and event queue mechanism decoupling, and the UI interaction, semantic processing and model reasoning three layers can be independently tested and replaced, with good engineering modularization characteristics. The implementation of the local offline semantic assistant engineering system integration has the ability of semantic driven behavior, dialogue memory state mechanism and closed loop ability of engineering linkage operation, which includes the following functions:

[0072] Field interpretation: the user clicks on the table field, and the assistant automatically interprets its meaning and typical format;

[0073] Abnormal diagnosis: the user asks "is this row a problem?", and the assistant analyzes and diagnoses the abnormality according to the database;

[0074] Device replacement recommendation: input "replace a suitable capacitor", and the assistant queries the database and returns the suggestion;

[0075] Operation instruction triggering: the user inputs "confirm the suggestion", and the assistant drives the approval process to update;

[0076] Multi-round memory question and answer: continuous context questions, no need to input device information repeatedly;

[0077] Session log tracking: all questions and answers and operations are automatically recorded, which can be traced back and exported;

[0078] Linkage interaction: when the user clicks on a field or selects a device row in the main table, the assistant window can automatically insert information;

[0079] The AI assistant dialogue engine module includes the user interaction layer (UI interface), the semantic intermediate processing layer, and the reasoning and response layer;

[0080] Among them, the user interaction layer (UI interface) provides a natural language interaction window for users, mainly including:

[0081] The floating assistant interface is implemented using PySide2;

[0082] The interface includes input box, send button, scroll response area, linkage interaction area, etc;

[0083] The response area supports Markdown rendering (bold / underline / code block);

[0084] The user clicks on the main BOM table field, which is automatically transmitted to the assistant as context injection;

[0085] Among them, the semantic intermediate processing layer includes a session state manager, a Prompt builder and an instruction intention recognizer;

[0086] Session state manager: maintain the session state of the current BOM document, generate a separate session_id for each document opening, save the log using local.json / SQLite file, provide session loading, clearing, exporting capabilities;

[0087] Prompt builder: generate standardized prompts that can be used for model understanding, automatically identify field semantics and construct templates, such as encapsulation, voltage, description field dynamic generation, and structure data templates into natural language input, for example, the current device is: capacitor, 0603 package, 10uF / 16V, please judge whether it meets the requirements. Through the introduction of dynamic Prompt selection after user preference memory, field naming deviation as context prompt, the final input is sent to Mistral7B model;

[0088] Instruction intention recognizer: determine whether the user input is a certain type of system instruction, such as approval, recommendation, and verification. Use MiniLM or BERT to vectorize the input sentence, and perform Top-K similarity matching with the instruction corpus ("confirm" "change one" "replace recommended capacitor"). The instruction corpus supports dynamic learning and user customization, and supports embedding field context into instruction vector calculation, such as "click the capacitor value field and say 'change one'";

[0089] Among them, the reasoning and response layer includes a model service interface and a streaming output manager;

[0090] Model service interface: use mistral.cpp framework to load Mistral7B model in GGUF format to local, use INT4 quantization to optimize inference efficiency. This interface supports automatic switching and hot loading of models based on user context or task type;

[0091] Streaming output manager: the streaming output manager is implemented based on StreamingToken inference mechanism, which pushes the token-by-token results returned by the language model to the front-end assistant interface through SSE. The interface uses QTextBrowser control for batch rendering and real-time updating, supports user interaction interruption (such as "stop generating" button), and dynamically inserts field highlighting, text underlining, jumpable hyperlinks and other enhanced visual feedback during display to improve human-computer interaction efficiency;

[0092] Semantic classification and field specification module, through the semantic classification and field specification module, mark the semantic fuzzy fields in the document of uniform format, and trigger the AI assistant dialogue engine module to issue suggestion information for engineers to confirm;

[0093] The semantic classification and field specification module uses a rule engine and regular identification to identify common field aliases, such as "PN / Part No. / Number → Material Code". When the field semantics are ambiguous, an AI assistant pop-up is triggered to confirm the suggestion, such as "Is the Value field a resistance value or a device description?". The semantic classification and field specification module supports automatic construction of a "field semantic mapping dictionary" for reuse by the verification and export module. Unstructured fields are vectorized for subsequent search using the BGE-m3 model. The semantic classification and field specification module is used to explicitly mark primary key fields, ambiguous fields, and ambiguous fields in the unified format document;

[0094] The intelligent verification and exception diagnosis module interfaces with the fields in the device database, verifies and diagnoses the unified format document according to the fields in the device database, and generates corresponding suggestion records for engineers to process;

[0095] The intelligent verification and exception diagnosis module interfaces with the internal device database (MySQL / SQLite) to perform field legality matching, uses a rule library for duplicate position number checking, format illegal checking, voltage / tolerance / package recommendation, and redundant material number identification, uses a large model to analyze the context and provide "natural language suggestions", all abnormal items automatically generate suggestion records, assign suggestion IDs (such as SUG-00001), and push to the approval module, the diagnosis result has a state label and a manual intervention interface, each suggestion has a number, source, and context prompt information;

[0096] The approval mechanism and state tracking module includes a state label for each suggestion item, including pending, pass, reject, and skip, users can click to confirm, reject, and comment on the reason for rejection, all operations are recorded in the "approval log", including suggestion ID, original value, new value, timestamp, user, and decision state, providing an approval UI interface that can filter all "items to be approved" for batch processing;

[0097] The deployment and resource scheduling optimization module integrates the AI assistant dialogue engine module and the language model, and the embedded AI assistant dialogue engine module analyzes the unified format document in off-network and low-power consumption environments;

[0098] The deployment and resource scheduling optimization module includes a model running interface layer and a reasoning scheduling and caching layer;

[0099] The model running interface layer includes a model loader that loads the GGUF format weight file of the local language model into memory, uses a memory mapping mechanism to achieve efficient access and on-demand retrieval of the weight file, and builds a static link library for the compiled local reasoning engine of the language model and embeds it into the desktop application front-end program;

[0100] The inference scheduling and caching layer includes using a thread pool scheduling mechanism to encapsulate each inference request as an asynchronous task object and submit it to the thread pool managed by ThreadPoolExecutor for concurrent execution. All requests are queued and processed according to task priority.

[0101] The deployment and resource scheduling optimization module also includes a model management and upgrade subsystem layer and a local log and permission control subsystem layer.

[0102] The deployment and resource scheduling optimization module is responsible for the key basic subsystems of large language model semantic interaction capabilities for full localization and off-network operation. The goal is to enable AI assistants to implement model local deployment without calling any external API, ensuring that the model and system front end are tightly integrated and embedded, ready for use, data remains local, and model hot swapping and upgrade interfaces are preserved, supporting subsequent model switching and performance expansion. The specific functions include:

[0103] Local deployment capability: supports model and front-end assistant process embedding integration, off-network operation of large language models on personal PCs, ensuring data privacy and local computing needs;

[0104] Inference process embedding front end: model calls are directly executed by the assistant program without RESTAPI, servers, or background service processes, improving deployment simplicity and security;

[0105] Model management and upgrade: supports multiple model files coexistence, hot loading switching through configuration or interface, facilitating subsequent model capability expansion (e.g., from 7B to 13B);

[0106] Streaming inference control: uses Streaming Token push mechanism during inference, generating and outputting in real-time, reducing waiting time and improving interaction smoothness;

[0107] Cache mechanism support: provides local caching and quick return mechanism for repetitive prompts, common device explanations, and other content, avoiding repeated model calls and reducing resource waste;

[0108] Data closure and permission control: inference, caching, and log data are all stored locally and do not rely on external networks, allowing for user operation permission and sensitive log access control strategies;

[0109] The model management and upgrade subsystem layer includes automatically finding language model files that meet the format when users switch between multiple language models by specifying the model folder, dynamically replacing the currently loaded language model by unloading the current model weights and clearing related memory mappings through resource release interfaces, subsequently reloading the new GGUF model file, and refreshing the inference context cache and user session records;

[0110] The local log and permission control subsystem layer is used to write each round of conversation records into a local database;

[0111] Specifically, the model management and upgrade subsystem layer includes that when the system is started, a model loader is called to load the GGUF format weight file of the local Mistral7B model into the memory, and a memory mapping mechanism such as mmap() is used to realize efficient access and on-demand calling of the weight file. In this way, the user terminal can realize the local readiness and initialization of the large language model without the need for an external server or REST interface, significantly improving the inference response efficiency and ensuring that data does not leave the local area. The local inference module supports embedded integration deployment, that is, the compiled local inference engine of the Mistral7B model is built as a static link library (such as.a or.lib file) and embedded into the desktop application front-end program for direct calling, eliminating the traditional RESTAPI or server service process and improving the runtime response speed and overall system architecture compactness. The reasoning interface encapsulation shields the complexity of the underlying model inference process through a standardized calling method, and unifies details such as multi-parameter, asynchronous execution, and resource scheduling into a function interface that can be called by the front end. The front-end program only needs to pass in user input and model parameters to complete the inference task, improving the system's modularization ability and calling convenience, and facilitating integration into desktop assistant interfaces or scripted processes. The interface encapsulation mechanism supports automatic splicing of structured field input prompts (fields + user preferences). Parameter control supports flexible configuration of multiple core parameters of the language model generation strategy, including generation length (max_tokens), randomness control coefficient (temperature), and sampling coverage parameter (top_p), which can be dynamically set through the front-end interface or user configuration file when calling. This mechanism avoids parameter hardcoding, allowing dynamic selection of generation strategies based on task context, user preferences, or predefined templates, improving system controllability and adaptability. It is used to build a trusted local runtime environment, shield external model calling capabilities, and ensure that all inference behaviors only come from local input.

[0112] Specifically, the inference scheduling and caching layer includes an asynchronous task scheduling mechanism suitable for desktop local inference deployment scenarios. The system uses a thread pool scheduling mechanism to encapsulate each inference request as an asynchronous task object and submit it to the thread pool managed by ThreadPoolExecutor for concurrent execution, thereby avoiding blocking the main thread. The task scheduling system supports dynamically adjusting thread priority based on the type of inference task, and the scheduling optimization of "dialog type request > batch generation type request" can be combined with the caching mechanism to form a "cache-driven task preemption scheduling" to realize the linkage of asynchronous scheduling and streaming output (the task can be interrupted when it is not completed), improve the overall response efficiency and resource utilization of the system, and be suitable for running scenarios of multiple windows and multiple sessions calling local language models in parallel. A thread-safe task queue with a first-in-first-out (FIFO) structure is used to manage all inference requests. All requests are scheduled and processed according to the task priority after being entered into the queue, in which the system distinguishes between critical tasks and non-critical tasks, such as "explanation type requests triggered by field clicks" have higher scheduling priority than "user natural language free questions". This mechanism ensures that high-priority tasks are quickly responded to, improving user interaction experience and system task concurrency processing efficiency. By performing a hash digest operation on the user input prompt, a unique identifier (Key) is generated, and a quick matching search is performed in the local cache database. If the cache hits, the last generated response result is returned directly, skipping the inference process, and if it does not hit, the regular inference is executed and the new result is written into the cache. This mechanism effectively reduces repeated calculations, improves response efficiency, and reduces system resource occupation.

[0113] Specifically, the model management and upgrade subsystem layer includes supporting users to quickly switch between multiple model versions, monitoring a specified model folder, such as the models / directory, automatically finding all model files that meet the format, and displaying a dynamically generated switchable model list on the front end. It is used to dynamically replace the currently loaded language model during running to avoid context residue or reasoning state conflicts. In the model switching process, the resource release interface is first called to unload the current model weights and clear the related memory mapping, and then the new GGUF model file is reloaded, the inference context cache and user session record are refreshed, and the running isolation and environment cleanliness of the multi-model switching process are ensured. This process can provide a progress bar and switching state prompt in combination with the user interface to improve the controllability and visual experience. It is used to track the call history of the language model during running to support performance comparison analysis, running statistics, and behavior tracing of multiple model versions. The calling time, model name, version number, generated task type, and calling result are automatically recorded at each model call, and the data is written into a local log database such as SQLite. This module can assist in subsequent model switching strategy optimization, user behavior analysis, and calling effect statistics, improving maintainability and intelligent level. To prevent misoperation, security restrictions are added, and only administrators can switch or import new models (authorized password required) to ensure that the AI service runs safely within a trusted range and locally.

[0114] Specifically, the local log and permission control subsystem includes writing each round of dialogue record (prompt, response (response is the answer generated by the AI model based on prompt), time consumption, model name, timestamp) into a local database for subsequent tracing and analyzing output. A log export capability is designed for compliance, and administrators can export.csv / .md / .json format audit reports for security audits or version comparisons. Automatic networking is prohibited, remote model download is prohibited, and the entire running process is closed locally (no public network socket / request). A "log cleaning" function is provided to periodically delete historical session records 30 days ago, support keyword / field / user filtering and cleaning, and save space.

[0115] It also includes an approval mechanism and state tracking module that establishes an auxiliary decision-making mechanism based on the combination of the AI assistant dialogue engine module's suggestion information and user engineering behavior;

[0116] The approval mechanism and state tracking module includes a suggestion state data structure submodule, an approval interaction interface submodule, and a state marking visualization submodule.

[0117] The suggestion state data structure submodule records and controls the entire life cycle state change process of AI-generated suggestions, records all changes to a local log database, and forms a suggestion state flow history.

[0118] The approval interaction interface submodule allows users to trigger state change behavior through button clicks or natural language instruction triggers by providing button operations or dialogue instruction trigger mechanisms;

[0119] The state marker visualization submodule is used to visually present the current processing state of each suggestion in the human-computer interaction interface.

[0120] The approval mechanism and state tracking module is used to effectively combine AI-generated suggestions with user engineering behavior, establishing a set of "structured state tracking, manual confirmation link, operation log recording" combined intelligent auxiliary decision-making mechanism, giving users the "approval right" of AI output results, and ensuring process compliance, safety and traceability. Specifically, the following functions are included:

[0121] Suggestion state management: Each suggestion is bound to a unique state: pending confirmation, confirmed, rejected, skipped, and withdrawn.

[0122] User approval interaction: Users perform approval actions on suggestions through dialogue instructions or button operations.

[0123] State visualization markers: Mark the "confirmed / rejected" field in the main table (such as color, icon).

[0124] Approval log tracking: Automatically records approval behavior, operator, time, suggestion ID, decision result, and other information.

[0125] Batch approval function: Supports one-key confirmation / rejection of similar suggestions to improve efficiency.

[0126] Approval result export: Approval records can be exported to files for archiving or external review.

[0127] Specifically, the suggestion state data structure submodule is used to record and control the full life cycle state change process of AI-generated suggestions, ensuring that all suggestion content is "controllable, traceable, and withdrawable." The system assigns a unique identifier (such as SUG-0001) to each suggestion and provides a unified state update interface that receives suggestion ID, new state type, and operator identity as parameters to complete the suggestion state update process. All changes are recorded in the local log database, forming a suggestion state flow history that can be used for approval traceability, version comparison, or user behavior analysis.

[0128] Specifically, the approval interaction interface submodule provides button operation or dialogue instruction triggering mechanism, allowing users to trigger state change behavior through button click or natural language instruction. Graphically: each suggestion shows "confirm / reject / skip" buttons on the right side, and dialogically: supports semantic instructions such as "confirm the second suggestion", "reject the alternative suggestion", "pass all encapsulated modification suggestions", etc. The intent recognition module is responsible for converting natural language into standard commands. After successful operation, the assistant returns the feedback: "the modification suggestion has been confirmed, the system has updated the state".

[0129] The state marker visualization submodule is used to visually present the current processing state of each suggestion in the human-computer interaction interface. The system embeds the suggestion state into the corresponding cell of the main table interface through color identification (such as green for confirmed, red for rejected, and yellow for pending) and icon prompts. When the user hovers the mouse over the suggestion field, a pop-up prompt is displayed, including the state change time and the operating user information. After the suggestion state is updated, the system automatically refreshes the visualization marker area without the need for page reloading or manual operation.

[0130] The approval log recording submodule provides interfaces for audit writing, exporting, and backtracking, records complete operation traces for each approval behavior, supports subsequent review, export, and reuse. By storing the approval record format in SQLite, and supporting single export, condition filtering export (such as exporting all "rejected" suggestions), etc., the approval operation record is only increased and not modified, ensuring data auditability.

[0131] It also includes a memory mechanism and a personalized adaptation module, which is used to train the AI assistant in the AI assistant dialogue engine module;

[0132] The memory mechanism and personalized adaptation module are used to track and record user feedback behavior on system-generated suggestions, forming a data basis for preference learning and behavior modeling, automatically injecting user preferences and historical context into large model prompts, sorting results, and sorting and optimizing multiple suggestions or alternative devices or explanation content returned by the large model, and building user long-term preference modeling and behavior memory storage, so that the AI assistant in the AI assistant dialogue engine module has intelligent behavior with memory.

[0133] The memory mechanism and personalized adaptation module upgrade the AI assistant from a "general large model" to an "engineer assistant familiar with user habits", realizing true intelligent human-computer collaboration. Its goal is to record user behavior, identify usage preferences, gradually establish a "user portrait" and "semantic preference database", support project-level memory and user-level memory hierarchical management, all memory content is saved locally and can be transparently viewed or cleared, including the following functions:

[0134] User Behavior Memory: Record user's operation behavior on suggestions (confirmation, rejection, modification, etc.);

[0135] Preference Learning: Automatically learn preferences (such as packaging priority, brand inclination) based on user historical behavior;

[0136] Prompt Adaptation: Automatically embed user habits and context features when constructing prompts;

[0137] Response Ranking Optimization: Personalized ranking and filtering of suggestion results;

[0138] Multi-dimensional Memory Structure: Support independent memory cache in "user / item / device" dimensions;

[0139] Visual Management: Provide memory item viewing, editing, and clearing functions to ensure controllability;

[0140] Specifically, the user behavior acquisition submodule is used to track and record user feedback behavior on system generated suggestions, forming a data basis for preference learning and behavior modeling. Each user operation on a suggestion (such as confirmation, rejection, modification, and ignore) will be converted into a structured behavior log item, including suggestion number, operation user, operation time, suggestion field, and final accepted value, and will be classified and written into the local user behavior log database. The system supports triggering log collection through front-end interaction (such as clicking the confirmation button) or assistant natural language interaction instruction parsing, building a unified behavior collection channel. This module can provide behavior data support for subsequent suggestion ranking optimization, user preference customization, and system recommendation strategy.

[0141] Specifically, the personalized Prompt constructor automatically injects user preferences and historical context into large model prompts, improving reasoning accuracy and constructing personalized templates, for example: you are a BOM assistant familiar with user "X" preferences, this user prefers to use domestic components, 1% resistors, and prefers certain brands. The current component is a 10uF capacitor of certain brand, please judge whether to recommend, whether to replace it with a domestic component. All personalized injection logic is completed by the Prompt splicer, embedding a generated historical preference summary.

[0142] Specifically, the response ranking optimizer re-ranks the results and optimizes the ranking of multiple suggestions / alternative components / explanation content returned by the large model. If the model output is multiple candidate results (such as recommending three component models), the following priority is preferred:

[0143] High: Match user preferences (brand / packaging / voltage);

[0144] Medium: Higher frequency of historical confirmation options;

[0145] Low: Recommend the option with a high logic score (assisted by the logit probability returned by the model), which is the "confidence score" output by the model, not yet converted into the final probability. The higher the logit, the more the model is "biased" towards this result;

[0146] The ranking result is displayed in the assistant window and can be marked with the words "in line with preferences";

[0147] Specifically, the memory database manager is used to build a user long-term preference modeling and behavior memory storage, so that the local AI assistant has intelligent behavior with memory, provides the ability to locally persistently store memory content, and provides viewing, clearing, and exporting interfaces, all memory data is uniformly saved through a SQLite database file, different data tables record the static identity information or setting preferences of the user, support clearing memory and exporting, similar to restoring default settings and exporting preference summary, and some preference strategies (such as naming encapsulation) can also be locked by an administrator.

[0148] It also includes a multi-window parallel view module, and the multi-window parallel view module is provided with a main table view, a dialogue window, an approval window and a graph window;

[0149] The main table view, the dialogue window, the approval window and the graph window data are interacted with each other, and clicking a field automatically updates other windows.

[0150] The application also provides a collaborative BOM data analysis and verification method based on an AI model, including the following steps:

[0151] S1: When opening a hardware BOM list file, automatically activate the AI assistant window and read and parse the file, initialize the user session, read the field preferences, operation habits and project context configured by the user memory, and generate a unified data structure for calling by the interaction module;

[0152] S2: Mark the file through three trigger modes, including user questioning, clicking a field and automatic checking;

[0153] S3: Generate model input through a Prompt builder, and dynamically construct a customized Prompt according to the combined field content, user preferences and project memory;

[0154] S4: Local large language model running and adjusted model parameter running;

[0155] S5: Generate field explanation, verification judgment, modification suggestion and alternative device recommendation answers through the AI assistant, wherein the response is output in the form of Streaming Token, supporting user "interrupt generation" and "requestioning";

[0156] S6: User approval mechanism triggers, through the button or the dialogue to perform "confirmation / rejection" operation and at the same time record the user operation trigger behavior acquisition module, the approval result is marked to the hardware BOM file and written in the audit log record, and the user's operation behavior and preference trend are recorded, and the long-term structure preference is extracted;

[0157] S7: The semantic optimization and response sorting of the subsequent dialogue reference memory are carried out, the response result order is optimized according to the historical behavior, the user portrait is constructed, and the long-term collaborative efficiency optimization is formed;

[0158] S8: The application export approval log, the personalized preference log and the result file model record file are exported.

[0159] It should be noted that, in this text, the relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the term "include", "contain" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment.

[0160] The above is only the preferred embodiment of the present application, and it should be pointed out that for ordinary skilled in the art, without departing from the technical principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should be considered as the protection scope of the present application.

Claims

1. The collaborative BOM data analysis and verification system based on AI model is characterized by: include: An intelligent document perception and structure analysis module, which monitors and identifies the manifest file and converts the output into a document in a unified format; An AI assistant dialogue engine module, which uses a language model to interact with engineers to analyze documents in a unified format field by field, thereby assisting engineers in understanding the contents of the manifest file; A semantic classification and field specification module, which marks semantically ambiguous fields in a unified format document and triggers the AI ​​assistant dialogue engine module to issue a suggestion message for the engineer to confirm; An intelligent verification and abnormality diagnosis module, which interfaces with fields in a device database, verifies and diagnoses a unified format document based on the fields in the device database, and generates corresponding suggestion records for engineers to process; A deployment and resource scheduling optimization module, wherein the deployment and resource scheduling optimization module embeds and integrates the AI ​​assistant dialogue engine module with the language model, and the embedded integrated AI assistant dialogue engine module analyzes unified format documents in an offline and low-power environment; It also includes an approval mechanism and a status tracking module, which establishes an auxiliary decision-making mechanism based on combining the suggestion information issued by the AI ​​assistant dialogue engine module with the user's engineering behavior; The approval mechanism and status tracking module includes a suggested status data structure submodule, an approval interaction interface submodule, and a status mark visualization submodule; The suggestion status data structure submodule records and controls the entire life cycle of AI-generated suggestions, records all changes in the local log database, and forms a history of suggestion status changes. The approval interaction interface submodule provides a button operation or dialogue instruction triggering mechanism, allowing users to trigger state change behavior by clicking a button or using natural language instructions; The status mark visualization submodule is used to intuitively present the current processing status of each suggestion in the human-computer interaction interface.

2. The collaborative BOM data analysis and verification system based on the AI ​​model according to claim 1 is characterized by: The AI ​​assistant dialogue engine module includes a user interaction layer, a semantic intermediate processing layer, and a reasoning and response layer; The user interaction layer uses PySide2 to provide users with a natural language interaction window; The semantic intermediate processing layer parses the unified format document through the session state manager, prompt builder and instruction intent recognizer; The reasoning and response layer uses the model service interface and the streaming output manager to dynamically add tags to the parsed unified format documents in real time.

3. The collaborative BOM data analysis and verification system based on the AI ​​model according to claim 2 is characterized by: The model service interface uses the mistral.cpp framework to load the language model into the local machine in GGUF format, and uses INT4 quantization to optimize inference efficiency. The model service interface automatically switches and hot loads language models based on the context or task type of the unified format document.

4. The collaborative BOM data analysis and verification system based on the AI ​​model according to claim 2 is characterized by: The streaming output manager pushes the step-by-step token results returned by the language model to the AI ​​assistant dialogue engine module through SSE based on the Streaming Token inference mechanism; The AI ​​assistant dialogue engine module dynamically adds tags in real time based on the level-by-level token results.

5. The collaborative BOM data analysis and verification system based on the AI ​​model according to claim 1 is characterized by: The deployment and resource scheduling optimization module includes a model operation interface layer and an inference scheduling and caching layer; The model operation interface layer includes using a model loader to load the GGUF format weight file of the local language model into the memory, using a memory mapping mechanism to achieve efficient access and on-demand retrieval of the weight file, and building the local inference engine compiled from the language model into a static link library and embedding it into the desktop application front-end program; The inference scheduling and caching layer includes a thread pool scheduling mechanism, which encapsulates each inference request as an asynchronous task object and submits it to the thread pool managed by ThreadPoolExecutor for concurrent execution. After all requests are queued, they are scheduled and processed according to task priority.

6. The collaborative BOM data analysis and verification system based on the AI ​​model according to claim 5 is characterized by: The deployment and resource scheduling optimization module also includes a model management and upgrade subsystem layer and a local log and authority control subsystem layer; The model management and upgrade subsystem layer includes automatically searching for language model files that meet the format through the specified model folder when the user switches multiple language models. When dynamically replacing the currently loaded language model, it unloads the current model weights and clears the related memory mapping through the resource release interface, then reloads the new GGUF model file, and refreshes the inference context cache and user session records. The local log and authority control subsystem layer is used to write each round of conversation records into the local database.

7. The collaborative BOM data analysis and verification system based on the AI ​​model according to claim 6 is characterized by: It also includes a memory mechanism and a personalized adaptation module, which are used to train the AI ​​assistant in the AI ​​assistant dialogue engine module; The memory mechanism and personalized adaptation module are used to track and record the user's feedback behavior on the system-generated suggestions, form a data basis for preference learning and behavior modeling, automatically inject user preferences and historical context into the large model prompt words, sort the results, and optimize the sorting of multiple suggestions or alternative devices or instructions returned by the large model, and build a long-term preference modeling and behavior memory storage for users, so that the AI ​​assistant in the training AI assistant dialogue engine module has intelligent behavior of memory.

8. The collaborative BOM data analysis and verification system based on the AI ​​model according to claim 7 is characterized by: It also includes a multi-window parallel view module, wherein the multi-window parallel view module is provided with a main table view, a dialogue window, an approval window and a graphic window; The data in the main table view, dialogue window, approval window and graphic window are linked to each other, and clicking on a field automatically links other windows to update.

9. A method for applying the collaborative BOM data analysis and verification system based on an AI model according to any one of claims 1 to 8, characterized in that: The following steps are involved: S1: When a hardware BOM file is opened, the AI ​​assistant window is automatically activated and the parsed file is read to initialize the user session. The user's memory configuration field preferences, operation habits, and project context are read to generate a unified data structure for the interactive module to call. S2: Marking of documents through three triggering methods: user questions, field clicks, and automatic checks; S3: Generate model inputs through the Prompt Builder and dynamically construct customized prompts based on the combined field content, user preferences, and project memory; S4: local large language model operation and adjusted model parameters operation; S5: The AI ​​assistant generates answers including field explanations, verification judgments, modification suggestions, and recommended alternative components. Responses are streamed out in the form of Streaming Tokens, allowing users to "interrupt generation" and "ask again." S6: The user approval mechanism is triggered, and the "confirm / reject" operation is executed through buttons or dialogues. The user operation triggers the behavior collection module to record the approval result. The approval result is marked in the hardware BOM file and written into the audit log. The user's operation behavior and preference trends are recorded to extract long-term structural preferences. S7: Subsequent conversations refer to the memory for semantic optimization and response sorting. The response order is optimized based on historical behavior, user profiles are constructed, and long-term collaborative efficiency optimization is achieved. S8: Application export approval log, personalized preference log, result file model record file.

Images