CN-122020163-A - Industry special system construction method based on large model and intelligent agent cooperative optimization

CN122020163ACN 122020163 ACN122020163 ACN 122020163ACN-122020163-A

Abstract

The invention discloses an industry special system building method based on large model and agent collaborative optimization, which comprises two core steps of industry large model generation and agent collaborative optimization, and concretely comprises the following steps of S1, the industry large model generation sequentially completes industry adaptation model building through a data set generation module, a model fine adjustment module, a model format conversion and quantization module and a model reconstruction and loading module, S11, the data set generation module adopts an open source general large model as an initial semantic generation engine, and the method realizes quick injection of industry knowledge and efficient migration of model capability by constructing a full-link system which covers data set automatic generation, parameter efficient fine adjustment, model format conversion and quantization, model reconstruction and quick loading and agent building and collaborative optimization control.

Inventors

LI QIAN
YIN CHENGLIANG
WU YUEPENG

Assignees

上海智能网联汽车技术中心有限公司

Dates

Publication Date: 20260512
Application Date: 20260120

Claims (10)

1. The method for constructing the industry special system based on the cooperative optimization of the large model and the intelligent agent is characterized by comprising two main core steps of the generation of the large model of the industry and the cooperative optimization of the intelligent agent, and specifically comprises the following steps: S1, generating an industry large model, namely completing industry adaptation model construction through a data set generation module, a model fine adjustment module, a model format conversion and quantization module and a model reconstruction and loading module in sequence; s11, a data set generation module adopts an open-source general large model as an initial semantic generation engine, carries out structural expansion on industry task description based on self-supervision learning and semantic generation mechanisms, automatically generates diversified sample texts and question-answer data, calculates semantic similarity, logical continuity and anomaly detection indexes through a semantic consistency screening algorithm, and filters low-quality samples to form a high-quality available data set; S12, a model fine adjustment module adopts a parameter efficient fine adjustment (PEFT) method, a low-rank parameter matrix is introduced into a model part weight structure to carry out incremental training, and industry task adaptation and generalization capacity optimization are realized; S13, loading the fine-tuned model by the model format conversion and quantization module, reading structural configuration parameters and weight tensors, mapping the structural configuration parameters and the weight tensors to a unified weight storage format file, adopting fixed-point quantization to replace floating point calculation, and combining tensor rearrangement, block loading, mixing precision and cache optimization strategies to reduce the occupation of a video memory and the complexity of operation; S14, defining meta information by a model reconstruction and loading module through a model description file, loading model weights in blocks according to a weight index and a memory mapping mechanism, realizing dynamic adjustment of reasoning parameters through a parameter mapping table, and finishing input structuring pretreatment through a template engine to form an operable model instance; S2, the intelligent agent collaborative optimization is realized, namely end-to-end task execution and closed-loop optimization are realized through an intelligent agent building module and a collaborative optimization controller module; S21, the agent building module calls a model reasoning service and a third-party tool API based on task arrangement and chained task control logic to complete user intention recognition and task planning, chained tool calling and execution, result generation and multi-round interaction; S22, the collaborative optimization controller module collects interaction logs, reasoning results, tool call records and execution feedback in real time, performs abnormal attribution analysis through the multi-classifier, and triggers an automatic response strategy aiming at different root causes to form a closed-loop optimization flow.
2. The method for building the industry-specific system based on the collaborative optimization of the large model and the agent according to claim 1, wherein in S11, the open-source general large model is a LLaMA series model, the industry task description comprises a core task instruction of a target industry typical scene, the generated data type covers question-answer data, task instruction data and a plurality of dialogue samples, a semantic consistency screening algorithm calculates the comprehensive score of semantic similarity, logical consistency and abnormal detection indexes through weighting, and the sample with the comprehensive score lower than a preset threshold is filtered.
3. The method for building the industry special system based on the collaborative optimization of the large model and the intelligent agent according to claim 1, wherein in the step S12, the efficient fine adjustment of parameters is realized based on LLaMAFactory frames, a low-rank parameter matrix adopts a low-rank adaptation (LoRA) structure, an incremental training data set is the industry special data set generated in the step S11, the trunk weight of the model is fixed in the training process, only the low-rank parameter matrix and associated bias parameters are updated, the performance index of the model on an industry task is monitored in real time through a verification set, and training is stopped when the performance index is continuously preset for a round without improvement.
4. The industry special system building method based on the collaborative optimization of the big model and the intelligent agent according to claim 1, wherein in the S13, the structural configuration parameters comprise model layer number, hidden layer dimension, word list size and tokenizer configuration, the unified weight storage format file adopts a custom binary format, the fixed point quantization adopts an FP16 quantization strategy, the 32-bit floating point weight is weighted into 16-bit fixed point weight, the tensor rearrangement performs dimension recombination according to a layer sequence of a Transformer network and an attention head dimension, and the partition loading splits the weight tensor according to a preset memory threshold and loads in batches.
5. The method for building the industry-specific system based on the collaborative optimization of the large model and the intelligent agent according to claim 1, wherein in the S14, the model description file adopts a JSON format, the meta-information comprises a weight file storage path, model parameter configuration, reasoning super-parameters and input template description, the reasoning super-parameters comprise batch processing size, reasoning temperature coefficient and maximum generation length, the memory mapping mechanism is realized through Ollama model management tools, and the parameter mapping table establishes a corresponding relation between the reasoning parameters and model weight tensor and supports dynamic adjustment of the reasoning super-parameters in operation.
6. The industry special system building method based on the collaborative optimization of the large model and the agent according to claim 1, wherein in S21, the agent building realizes task arrangement based on LANGCHAIN frames, tool interface definition adopts standardized JSONSchema format, user intention identifies that the industry large model generated by calling S14 completes deep semantic analysis, task arrangement generates a structured execution plan, defines subtask splitting, execution sequence and dependency relationship, chain tool calling sequentially calls a third party service API according to the execution plan, and multi-round interaction supports context association response based on dialogue memory.
7. The method for building the industry-specific system based on the collaborative optimization of the large model and the intelligent agent according to claim 1, wherein in the S22, training data of the multi-classifier comprises a manual labeling sample and an automatic synthesis negative sample of a historical fault case, the negative sample is generated by adding noise to normal interaction data, modifying key parameters, simulating tool call failure and the like, and due analysis covers five root cause types of insufficient model knowledge, missing of third-party tool call, interface mismatch, parameter configuration errors and user instruction ambiguity, and an automatic response strategy comprises automatic fine tuning trigger, API template generation and work order report, diagnosis report generation and operation and maintenance notification.
8. The method for building the industry-specific system based on the collaborative optimization of the large model and the intelligent agent according to claim 1 is characterized in that when the result is that the model knowledge is insufficient, the system automatically extracts a 'user query-standard response' sample pair from a failed dialogue, supplements the sample pair to a training data set and submits the training data set to a fine adjustment pipeline of S12, starts automatic incremental fine adjustment, the fine adjustment process multiplexes the high-efficiency fine adjustment method of the parameters of S12, only updates a low-rank parameter matrix, and when the result is that a third party tool call is absent or an interface is not matched, the system generates a function template or an API description document containing a function description, a parameter definition and a return value format based on an industry API specification and pushes the function template or the API description document to a developer for auditing in a worksheet mode.
9. The method for building the industry special system based on the collaborative optimization of the large model and the intelligent agent according to claim 1, wherein the industry special system is suitable for the intelligent travel field, during the generation process of the industry large model, S11 generates corpus data aiming at scenes such as man-machine interaction, travel planning, parking lot guidance, path avoidance and the like, S12 carries out model fine adjustment aiming at travel tasks such as multi-mode scene understanding, instruction generation, path optimization and the like, S13 model format conversion and quantization are suitable for vehicle or edge node deployment environments, and S21 intelligent agent supports travel related end-to-end tasks such as path navigation, parking service, charging pile query, scenic spot coupon query and the like.
10. The method for building the industry special system based on the collaborative optimization of the large model and the intelligent agent according to claim 1, wherein in a closed-loop optimization flow of the collaborative optimization controller module, an execution result of automatic response and subsequent interaction data are used as feedback data to flow back to an analysis model for updating a training sample set of the multi-classifier, attribution accuracy of the multi-classifier is continuously optimized through incremental training, execution effects of different response strategies are recorded, a strategy effect evaluation model is built, and priority and execution parameters of the response strategies are dynamically adjusted.

Description

Industry special system construction method based on large model and intelligent agent cooperative optimization Technical Field The invention relates to the technical field of automatic driving control, in particular to an industry special system building method based on cooperative optimization of a large model and an intelligent agent. Background With the rapid development of the generated artificial intelligence and natural language processing technology, an intelligent agent system based on a large-scale pre-training model has become a core engine for promoting the intelligent upgrading of industries. The general large model shows excellent performance in general scenes by virtue of strong language understanding and generating capability, but faces a plurality of problems to be solved urgently when the general large model is applied to specific industry scenes. The development path of the intelligent agent commonly adopted in the current industry has obvious limitations that the data layer depends on manual acquisition and labeling, the time and the labor are consumed, the cost is high, the consistency of the data quality and the industry suitability are difficult to ensure, the model layer needs manual fine adjustment of parameters, the parameter updating range is difficult to accurately control, the consumption of computing resources is large, the training period is long, the professional requirements on technicians are extremely high, a higher technical threshold is formed, the deployment layer has low migration efficiency of the model due to format differences of different hardware and reasoning environments, the lightweight operation is difficult to realize, various terminal devices are adapted, the intelligent agent construction layer lacks a unified scheduling mechanism, the model capability and external tools interact smoothly, a large number of repeated developments are needed, and the function expansibility and the cooperativity are insufficient. More importantly, in the prior art system, the large model update and the intelligent agent function promotion are mutually fractured, and an effective cooperative mechanism is lacked, so that an automatic closed-loop optimization flow cannot be formed. Meanwhile, the existing system lacks a systematic attribution and automation repair link for abnormality in the operation process, so that the system is difficult to continuously iterate and optimize, the quick landing and large-scale application of industrial intelligent agents are severely restricted, and an industrial special system building method based on cooperative optimization of a large model and the intelligent agents is needed to solve the problems. Disclosure of Invention The invention aims to provide an industry special system building method based on cooperative optimization of a large model and an intelligent agent, so as to solve the problems in the background technology. In order to achieve the aim, the invention provides the technical scheme that the method for constructing the industry special system based on the cooperative optimization of the large model and the intelligent agent comprises two main core steps of the generation of the industry large model and the cooperative optimization of the intelligent agent, and the method comprises the following steps of: S1, generating an industry large model, namely completing industry adaptation model construction through a data set generation module, a model fine adjustment module, a model format conversion and quantization module and a model reconstruction and loading module in sequence; s11, a data set generation module adopts an open-source general large model as an initial semantic generation engine, carries out structural expansion on industry task description based on self-supervision learning and semantic generation mechanisms, automatically generates diversified sample texts and question-answer data, calculates semantic similarity, logical continuity and anomaly detection indexes through a semantic consistency screening algorithm, and filters low-quality samples to form a high-quality available data set; S12, a model fine adjustment module adopts a parameter efficient fine adjustment (PEFT) method, a low-rank parameter matrix is introduced into a model part weight structure to carry out incremental training, and industry task adaptation and generalization capacity optimization are realized; S13, loading the fine-tuned model by the model format conversion and quantization module, reading structural configuration parameters and weight tensors, mapping the structural configuration parameters and the weight tensors to a unified weight storage format file, adopting fixed-point quantization to replace floating point calculation, and combining tensor rearrangement, block loading, mixing precision and cache optimization strategies to reduce the occupation of a video memory and the complexity of operation; S14, defining meta information by