214 results about "Asynchronous processing" patented technology

The invention provides an agent memory management method and device based on asynchronous processing and a storage medium. The method comprises the following steps: after an interaction event is written into a first storage area, issuing a processing event associated with an event identifier to an event queue, and reading the corresponding interaction event from the first storage area according to the event identifier; when the interaction event passes the importance judgment, converting the interaction event into structured memory data; determining second storage areas based on the type identification field of the structured memory data, and respectively writing the structured memory data into at least one second storage area; and analyzing the query request to obtain a query condition, determining at least one target second storage area based on the query condition, reading structured memory data related to the query request from the target second storage area, performing aggregation processing, and outputting the aggregated memory data to the agent to generate a response result. According to the method, the interaction response delay can be reduced, the model calling cost is reduced, and the memory processing auditing performance is improved.

The invention relates to the technical field of data acquisition, in particular to an AOP-based user behavior data acquisition method. The method comprises the following steps: firstly, identifying user behavior related code features by statically analyzing Java source codes of an application program, and constructing a feature template library; during a compiling period, according to a template library identification target method, embedding a point burying instruction by utilizing a byte code enhancement technology, and constructing a monitoring section framework; and during running, dynamically adjusting an execution strategy according to the running state of the application program by virtue of a dynamic rule engine. Meanwhile, an annular buffer area is adopted to asynchronously process user behavior data streams, cache data are managed in a sub-generation mode, and asynchronous processing is triggered when a threshold value is exceeded. The method also constructs an adaptive thread pool model based on a hierarchical thread pool architecture, distributes tasks according to data value weights, monitors the task backlog rate of a high-priority sub-pool, adjusts the sampling frequency or acquisition dimension when the task backlog rate exceeds a threshold value, forms adaptive closed-loop control, and effectively acquires and processes user behavior data.

The invention discloses an asynchronous processing method and device for business process data, and belongs to a distributed system data processing technology, and the method comprises the following steps: S101, receiving a business request message, and persistently storing the business request message in a main message queue; s102, acquiring a to-be-processed message from the main message queue, and locking an associated target data record before executing processing; s103, verifying data consistency based on the version number recorded by the target data, if the version number is matched, executing processing and updating the version number, and otherwise, discarding the message; s104, after successful processing, sending a confirmation instruction deletion message to the main message queue; through persistent storage of the main message queue, a version number verification mechanism and management of the dead message queue, the problem of data inconsistency caused by concurrent operation is effectively prevented, reliable transmission of messages is ensured, loss of service data is avoided, dead message processing efficiency is optimized, and the method has the advantages of ensuring data consistency, preventing message loss and improving system reliability.

The invention provides a multi-strategy caching and automatic data generation method and system based on a Node environment, and relates to the technical field of data processing, and the method comprises the following steps: receiving a client request, obtaining a data generation template, constructing a data dependency graph to analyze a field dependency relationship, and determining a field generation sequence by adopting a recursive generation strategy and depth-first search; data validity verification is executed, a data generation thread pool is created by adopting an adaptive load balancing mechanism based on the data dependency graph, and asynchronous processing is realized through double buffer queues, so that the data generation efficiency is improved, the data logic integrity is ensured, and the system resource consumption is reduced.

The invention belongs to the technical field of computer graphic processing, and particularly relates to a computer device for chartlet loading, which comprises a central processing unit provided with a main thread for packaging a chartlet loading request into a task object and pushing the task object into a task queue of the central processing unit; the resource loading thread is used for sequentially pulling the to-be-processed task objects from the task queue in an asynchronous processing mode and reading the mapping files in the task objects through the asynchronous I / O interface; the resource decoding thread is used for decoding the mapping file and uploading texture data obtained by decoding to the graphics processor through an asynchronous interface; a plurality of threads are created in the graphics processor and at least comprise a resource processing thread used for receiving texture data uploaded by a resource decoding thread; and the rendering thread is used for carrying out texture binding on the texture data and carrying out rendering according to an instruction of the main thread after binding so as to draw and generate a corresponding map.

The invention discloses a volunteer service work order intelligent processing method and system based on multistage classification and similarity matching, and the method comprises the steps: carrying out the differential cleaning of a work order text and case data, constructing a structured template, and carrying out the label mapping coding; through coarse-grained classification of the second-level classification model, triggering fine-grained division of the third-level classification model when a large class is specific and the confidence coefficient is in a preset interval; retrieving associated historical cases based on a semantic similarity algorithm; and constructing a dynamic prompt template according to matched case elements, and calling a large model API to generate a mediation suggestion. Further, a paging mechanism is adopted to asynchronously process the work order, store the result and integrate an exception handling mechanism. According to the method, the work order processing intelligence level can be improved, and the classification and matching efficiency can be optimized. According to the method, field adaptive learning, hybrid index optimization and dynamic Prompt engineering are fused, an intelligent closed loop from work order analysis to mediation suggestion generation is constructed, and the matching efficiency of community service resources and the scientificity of decision making are remarkably improved.

The invention discloses a centralized to-do processing system and method constructed based on message driving, and aims to solve the problems of dispersed to-do entrances, low processing efficiency and the like of an enterprise multi-service system. The system takes a message drive as a core, acquires to-be-handled data of each service system through an external data acquisition module, encodes the to-be-handled data through a to-be-handled management module, generates a subscription relationship, and intelligently pushes the to-be-handled data to a terminal interaction module by a to-be-handled synchronization module based on user role preference, so as to realize'message searching for people '. The message driving model covers five algorithms of transmission, routing, asynchronous processing and the like, and ensures accurate transmission and efficient processing of messages. The system unifies message formats, supports intelligent reminding and dynamic authority management, and adopts an encryption technology to guarantee safety. An enterprise does not need to reform an original system, one-stop aggregation processing of the to-do tasks can be achieved through lightweight docking, and office efficiency and cooperation experience are remarkably improved.

The invention provides an asynchronous parallel-based service processing method, system and device and a storage medium, and the method comprises the steps: initializing a service context object according to a service type identifier in a request; a preprocessing rule script corresponding to the service type is loaded and executed through a dynamic rule engine, and parameter verification, service pre-check and context initialization are carried out; a plurality of asynchronous processing rule scripts corresponding to the service types are loaded and executed in parallel, each asynchronous processing rule script corresponds to one service task and is packaged into an asynchronous task, the asynchronous tasks are executed in parallel, and a result is written into a service context object; and loading and executing a result processing rule script corresponding to the service type through a dynamic rule engine, aggregating results of the different step tasks, and generating and returning a service response after applying a final service logic. Parallel processing can be realized through a three-stage asynchronous processing architecture, multi-core CPU resources are fully utilized, and the throughput and response speed of the system are improved.

The invention provides an Internet of Things event center distributed alarm processing method based on Redis Stream, belongs to the field of Internet of Things (IoT) data processing, and realizes millisecond-level event processing by utilizing the high-performance streaming consumption capability of the Redis Stream and combining a distributed rule matching engine and a cross-device data dynamic association mechanism. Comprising the following steps that: Internet of Things terminal equipment writes data into Redis Stream according to a stream: Model: DeviceId mode to form an independent data stream; an event consumer reads the data from the stream, performs matching based on a preloaded rule, and supports cross-device linkage calculation through a metadata relation table and an SQL template; real-time duplicate removal and full-life-cycle management are carried out on alarm events through a dual convergence mechanism based on a database and a cache, and repeated alarms are eliminated; and finally, through multi-channel push and push state write-back, a closed-loop alarm system is constructed. The system adopts a layered architecture, comprises an event acquisition layer, a rule processing layer, an event management layer, a push service layer and an asynchronous processing layer, has the characteristics of high throughput, strong orderliness and high reliability, and can effectively improve the alarm processing efficiency and accuracy in a large-scale Internet of Things scene.

The invention discloses an API (Application Program Interface) paging query method based on a streaming data request, and relates to the technical field of data processing. Comprising the following steps: 1, generating a data identifier and a vernier; 2, performing paging request and data acquisition; 3, data updating and consistency processing are carried out, wherein when the data of the data source is updated, the server maintains a data change log; when a paging request is processed, the server performs consistency check and adjustment on returned data according to the data change log and the vernier; 4, streaming data processing and real-time guarantee are carried out, specifically, for continuous streaming data requests, the server adopts an asynchronous processing mechanism, the requests are put into an asynchronous task queue to be processed, and meanwhile a temporary response is returned; the server asynchronously obtains the data in the background, and pushes the paging data to the client after the data is asynchronously obtained; 5, a cache mechanism is optimized, specifically, the server is provided with multiple levels of caches, and paging data with frequent requests and small data changes are preferentially obtained from the caches; and when the data of the data source changes, the cache is updated in time.

The invention relates to the technical field of artificial intelligence and distributed computing, in particular to a multi-agent-based asynchronous collaboration method and system, and the method comprises the following steps: S1, user input and intention analysis; s2, task decomposition and asynchronous distribution; s3, asynchronous processing of the sub-agents and generation of an intermediate result; s4, result aggregation and standardized output; s5, result output interacts with user feedback; according to the invention, a decoupling cooperation mechanism of the main control agent and the plurality of functional sub-agents is introduced at the architecture level, and an asynchronous concurrence and cache control strategy is combined, so that efficient cooperation and smooth data transmission among the multiple agents are realized, the result accuracy is ensured, the overall throughput and the interaction response speed of the system are remarkably improved, and the system performance is improved. Through a multi-agent asynchronous cooperation architecture, the processing efficiency and reliability of complex interaction tasks are effectively improved, and high universality and commercial application value are achieved.

The invention provides an OFD archive file retrieval method and system, and the method comprises the steps: receiving file data, and carrying out the authentication verification of the file data; splitting the file into file units and storing the file units into a message queue for asynchronous processing in sequence; extracting text information, table information and picture information in the OFD file to generate comprehensive text information; preprocessing the comprehensive text information, and recording the occurrence frequency of each segmented word and the corresponding page number; creating an index in a search engine cluster, wherein an index field comprises a file ID (Identity), word segmentation content, occurrence frequency and a page number; searching according to an input keyword, and arranging results in a descending order according to the occurrence frequency of word segmentation; and marking the page number of the keyword for retrieval and positioning. The method has the beneficial effects that rapid analysis and preprocessing of large-scale OFD data are realized through asynchronous processing of the message queue and an OCR (Optical Character Recognition) technology; redundancy is reduced through coarse-grained word segmentation during indexing, and recall is enhanced through fine-grained word segmentation during retrieval; the keyword granularity input by the user is adapted, and the retrieval scene adaptability is improved.

The invention discloses a multi-type XML (Extensible Markup Language) equipment data processing method and device based on Netty. The method comprises the following steps: constructing a TCP long connection server through a Netty framework to receive a device coding data message; decoding and decompressing the message to obtain XML format data; analyzing the XML data to extract a type identification field; routing the data to a corresponding service processing strategy according to the type identification field; packaging the data into a processing task, and delivering the processing task to a corresponding thread pool for asynchronous processing based on the priority; and finally, converting the processed data into a unified data structure and storing the unified data structure in a database. High-concurrency receiving and processing of equipment data are achieved, multi-type XML message analysis is supported through dynamic routing, key data processing timeliness is guaranteed based on priority scheduling, the data processing throughput is improved by adopting an asynchronous mechanism, and high-concurrency receiving and processing of equipment data are achieved through a unified data conversion standard storage structure. And the efficiency, the real-time performance and the expandability of data processing are remarkably improved.

The invention provides a firmware management method and system, a storage medium and a program product. The method comprises the following steps: marking a firmware exclusive analysis class in advance through a (at) Firmware Type custom annotation, and defining a multi-product line verification rule by adopting a configuration file in advance; receiving a firmware uploading request, creating a corresponding asynchronous processing task, and temporarily storing a corresponding firmware file to the temporary storage path; obtaining a matched firmware analysis rule in the central rule base, and instantiating a corresponding (at) Firmware Type firmware exclusive analysis class to obtain analyzed firmware data; sequentially executing version preprocessing, dynamic rule matching and version verification on the analyzed firmware data according to a multi-product-line version verification rule; the firmware passing the version verification is intelligently stored; and cleaning the temporary file, and executing physical deletion on the firmware file with the reference return to zero stored in the formal storage path. By means of the technical scheme, the problems that an existing analysis rule is rigid, rule modification is tedious, and the storage utilization rate is low are solved.

The invention discloses a medical data increment synchronization method based on an HL7 protocol, and the method comprises the following steps: S1, a hospital internal system sends an HL7 message to a data receiving and processing module, and the data receiving and processing module carries out the following processing. According to the method, data transmission integrity is guaranteed through MLLP packaging verification and CRC verification, multi-message type dynamic analysis and heterogeneous data standardization are achieved through a high-performance HL7 analysis engine, data safety and compliance are guaranteed through TLS / SSL encryption, bidirectional authentication and the like, efficient increment synchronization of ten millions of data is achieved by means of a unique identifier and Redis cache, resource overhead is reduced, and data transmission efficiency is improved. Meanwhile, batch asynchronous processing and multi-thread parallelism are supported, high-concurrency scenes are adapted, analysis rule configuration management is matched, new requirements can be quickly adapted, the maintenance cost is reduced, and the requirements of medical informatization for data synchronization safety, real-time performance, high efficiency and extensibility are comprehensively met.

The invention discloses an electronic archive management method for a business document generation layout file, and belongs to the technical field of electronic archive management. Comprising the following steps: constructing a multi-dimensional feature vector by collecting business document scene features, screening candidate templates from a template pool by adopting a multi-level similarity matching algorithm, and adjusting and determining a target template in combination with a dynamic structure; the basic metadata and the associated system data are integrated, and qualified metadata is output through multi-level rule verification and abnormal data correction; based on the target template, executing data filling typesetting and format resource preprocessing by adopting a double-thread asynchronous processing mechanism, and generating a layout file conforming to PDF / OFD double standards through an adaptive algorithm; calculating a hash value of the layout file and writing the hash value into a block chain to realize tamper-proof evidence storage, and realizing secure storage and quick retrieval by adopting a distributed architecture in which a main file is separated from an index file. According to the method, the problems of low template matching precision, incomplete data verification, low processing efficiency, insufficient file safety guarantee and the like in the prior art are solved.

The invention provides a non-intrusive asynchronous processing method based on transaction event monitoring. An asynchronous event is triggered through a transaction event monitor after a database transaction is submitted, it is ensured that asynchronous operation strictly follows a transaction life cycle, and the risk of data inconsistency caused by dislocation of a thread pool task and a transaction time sequence in a traditional scheme is fundamentally avoided; by intercepting a business method carrying asynchronous execution annotations and automatically extracting parameters to generate an asynchronous request event, physical isolation of business logic and an asynchronous mechanism is realized, and architecture pollution is eliminated; and further dynamically routing to an instant execution or persistent execution module according to the policy type annotated and declared, so as to provide elastic execution capability for different service scenes. Therefore, on the premise of guaranteeing transaction atomicity, a zero-intrusive and extensible asynchronous processing framework is constructed, and the reliability and maintainability of the system are remarkably improved.

The invention discloses an interface calling method and device, equipment, a storage medium and a product, which can be applied to the field of financial science and technology, and the method comprises the following steps: obtaining an interface calling request sent by a target user; determining an interface processing mechanism according to the interface calling request; if the interface processing mechanism is an asynchronous interface processing mechanism, feeding back confirmation information to the target user, processing the interface calling request based on a pre-established asynchronous processing process, and generating an asynchronous processing identifier; and feeding back the asynchronous processing identifier to the target user, so that the target user inquires a request processing result of the asynchronous processing process on the interface calling request based on the asynchronous processing identifier. According to the technical scheme, dynamic interface calling under different interface calling scenes is achieved, and the interface calling efficiency and the interface calling flexibility are improved.

The application discloses a JSON data processing system implementation method based on FastAPI, which comprises the following steps: establishing a working directory based on FastAPI, creating a request processor in the working directory, and receiving original JSON data through the request processor; constructing a database object model in the working directory, generating a database table, and storing the original JSON data into the database table; constructing a business verification model in the working directory, preprocessing the original JSON data through the business verification model, and converting the processed JSON data into a Python object; creating a business processing tool in the working directory, processing the business logic of the Python object through the business processing tool, converting the processed Python object into a JSON data format; and creating a request agent and a middleware client agent of an external system in the working directory, and interacting the converted JSON data with the external system and the middleware. The application realizes an asynchronous processing kernel, supports high-concurrency tasks, improves the execution efficiency, and reduces the response time.

The application provides a data interaction method and device, electronic equipment, storage medium and program product; the method is applied to a data receiving end, the data receiving end comprises a first buffer for asynchronously processing and caching data, and a second buffer for controlling the sending behavior of a data sending end, the method comprises the following steps: receiving a data packet sent by the data sending end, and storing the data packet in the first buffer; reading to-be-processed data from the first buffer, processing the to-be-processed data, and deleting the to-be-processed data from the first buffer in response to the completion of the data processing; in response to the deletion of the to-be-processed data from the first buffer, generating a credit identifier between the data sending end and the data receiving end, and updating the state of the second buffer based on the credit identifier, the state being used for the data sending end to regulate the sending time of the data packet. Through the application, the stability of data interaction can be effectively improved.

The invention provides a data interaction method and device, electronic equipment, a storage medium and a program product. The method is applied to a data receiving end, the data receiving end comprises a first buffer area used for asynchronously processing and caching data and a second buffer area used for controlling the sending behavior of a data sending end, and the method comprises the following steps: receiving a data packet sent by the data sending end, and storing the data packet in the first buffer area; the method comprises the steps of reading to-be-processed data from a first buffer area, performing data processing on the to-be-processed data, and deleting the to-be-processed data from the first buffer area in response to completion of data processing; in response to the fact that the to-be-processed data is deleted from the first buffer area, a credit identifier between the data sending end and the data receiving end is generated, the state of the second buffer area is updated based on the credit identifier, and the state is used for the data sending end to regulate and control the sending opportunity of the data packet. According to the invention, the stability of data interaction can be effectively improved.

The invention discloses an IO cross-Die multi-core asynchronous processing method based on NVMeoverFabrics, and the method comprises the steps: building a CPU core resource pool under an NUMA architecture, carrying out the NUMA affiliation judgment in combination with a PCIe slot where a port is located, carrying out the sorting according to the number of CPU core binding queues during the queue creation, and selecting a target core, thereby achieving the balanced scheduling of the CPU cores and the reasonable allocation of the queues. Meanwhile, the NVMeIO processing process is divided into three stages of receiving, data transceiving and state sending which are asynchronously executed by different CPU cores respectively, so that the situation that a single core is occupied for a long time is avoided, and a plurality of queues are supported to share the same core so as to improve the resource utilization rate; and the threads in each stage are bound to the core in the NUMADie to which the port belongs, so that the performance loss caused by cross-Die memory access and PCIeDMA cross-node transmission is reduced.

The application provides a graph computing method and device, electronic equipment and storage medium, and relates to the technical field of data processing. The graph computing method comprises the following steps: determining a node task corresponding to a graph computing task and a thread pool corresponding to the node task according to real-time system resources and task information of the graph computing task, so that the thread pool acquires exchange data of other node tasks and node cache data in a cache module; in the case where the node cache data is not acquired, a preset database is searched through an asynchronous mode, and the searched node cache data is written into the cache module, so that the thread pool is re-acquired based on a synchronous mode; in the case where the node cache data is acquired, multi-thread synchronous calculation is performed according to the node cache data and the exchange data, so as to determine a graph computing result. The application combines synchronous processing, asynchronous processing and the mode of using a preset database when a cache is not hit, thereby improving the overall efficiency in large-scale data processing.

The present application relates to the field of integrated circuit automatic test equipment, and discloses an asynchronous execution method and system for ATE test machines.The asynchronous execution method comprises the following steps: an SDK module forms a plurality of executable tasks from test item interface instructions issued by an upper computer and stores the executable tasks in a task stack; the SDK module polls tasks in the task stack, judges the executability of the tasks and processing modules; if the tasks are executable, the SDK module further judges whether lower modules participate; according to the judgment result, the SDK module processes the tasks by itself or distributes the tasks to an ATOS module for processing or distributes the tasks to an FPGA module through the ATOS module for processing; the SDK module, the ATOS module and the FPGA module asynchronously process the assigned tasks at the same time; and after the tasks are completed, the processing results are returned to the task stack; after all subtasks of the current test item are completed, the SDK module loads tasks of the next test item until all test items are completed.

The embodiment of the application discloses a kind of LDPC decoding device and electronic equipment applied to chip, it is related to circuit field.The LDPC decoding device of the application includes: demodulation unit, first buffer unit, second buffer unit, LDPC decoder and controller;Wherein, the output of demodulation unit is connected with the input of first buffer unit, the output of first buffer unit is connected with the input of second buffer unit, the output of second buffer unit is connected with LDPC decoding unit;Controller is connected with the control end of first buffer unit and the control end of second buffer unit respectively;First data queue is arranged in first buffer unit, and second data queue is arranged in second buffer unit.The queue buffering mechanism is introduced in the application, the order and integrity of data are guaranteed, decoding requirements are flexibly adapted, and the rate matching problem between the LDPC decoder of rear end and the demodulation unit of front stage is effectively solved by using asynchronous processing mode.

The application provides an image processing method, device and electronic equipment, the method comprising: dividing a to-be-processed image into a plurality of sub-blocks; processing the plurality of sub-blocks using a plurality of CUDA (Compute Unified Device Architecture) streams, each CUDA stream in the plurality of CUDA streams comprising: a calculation operation and a memory operation, and a sub-block in the plurality of sub-blocks being processed asynchronously between the calculation operation in a first CUDA stream and the memory operation in a second CUDA stream; and splicing and fusing the processing results of the plurality of sub-blocks to obtain a processing result of the to-be-processed image. In the implementation process of the above scheme, the multi-core resources of the GPU are fully utilized through the asynchronous processing between the calculation operation and the memory operation in different CUDA streams, the idle time of the processing unit is reduced, the resource utilization rate of the CUDA stream is maximized, and therefore the overall processing efficiency of the high-resolution image is improved.

This application relates to an image display method, apparatus, computer device, chip, and chip module, and pertains to the field of image technology, capable of improving the smoothness and stability of image display. The method includes: obtaining a first target time for the drawing stage based on a first actual time consumption and a first historical time consumption for the application to draw historical image frames; determining a second target time for the compositing and display stage based on a second actual time consumption, a second historical time consumption for the image compositing service to process historical image frames, and a third historical time consumption for the hardware compositor to synthesize and display historical image frames; after the image compositing service submits the compositing and display task corresponding to the current image frame to the asynchronous thread of the hardware compositor, sending the second target time consumption to the hardware compositor and the first target time consumption to the application, to instruct the asynchronous thread of the hardware compositor to asynchronously process the compositing and display task according to the second target time consumption, and instructing the application to process subsequent image frames according to the first target time consumption.

The present application relates to the field of computer communication technology and provides a distributed ledger asynchronous data processing method and system based on a MEMO state machine, which is used to improve the intelligent adaptability of a distributed ledger network to dynamic business scenarios. The method comprises: obtaining a transaction request set in a distributed ledger network, extracting asynchronous processing trigger events of the transaction request set, parsing the asynchronous processing trigger events, generating a state transition instruction sequence including M state migration paths associated with K transaction data blocks, asynchronously processing the M state migration paths using a parallel prediction module based on a MEMO state machine, generating N predicted state vectors and corresponding conflict detection tags; performing feature fusion on the N predicted state vectors based on the conflict detection tags to generate a global state update matrix; writing the global state update matrix into a consensus node set of the distributed ledger network, and generating an asynchronous processing state log matching the transaction request set.

A method for generating a mark containing a moving target tracking image belongs to the technical field of image processing, and comprises the following steps: collecting a scene image through a multi-platform device, carrying out transparency processing, and generating training data containing a coordinate mark in combination with background standardization and an intelligent insertion mechanism; random noise points are used for simulating interference factors to enhance data robustness, and zero manual annotation is realized through structured filenames. A lightweight neural network constructed based on PyTorch integrates an attention mechanism, optimizes small target detection performance, and then deploys to an edge device. The system adopts a multi-thread asynchronous processing technology, realizes video stream real-time reasoning and visual coordinate labeling, and adapts to various edge hardware platforms. According to the method, generative data enhancement and edge calculation optimization are fused, the technical bottlenecks of manual annotation dependence, high model redundancy and large edge end delay of a traditional method are broken through, the target detection precision, the data generation efficiency and the system real-time performance are remarkably improved, and the method is suitable for motion tracking application of complex dynamic scenes.

The invention belongs to the technical field of natural language processing and artificial intelligence, and discloses a low-cost training and multi-round dialogue memorability method based on a large model technology. S1, transfer learning; s2, a data enhancement technology; s3, automatically adjusting parameters; s4, performing staged training; s5, model pruning and compression; S6, risk control and detection; s7, a streaming transmission technology; and S8, storing a chat algorithm. According to the method, fine adjustment is performed by utilizing knowledge of the pre-trained model, and the model is trained in stages, so that the training cost is reduced; the model answering efficiency is improved by adopting streaming transmission and asynchronous processing technologies; secure storage of user chat records is ensured, and model replies are adjusted individually according to historical dialogue records so as to maintain coherent interaction; the model size is reduced through pruning, quantification, compression and other technologies, and the deployment cost is reduced; the dependence on a large amount of data is reduced by using a data enhancement and regularization technology, and a good effect can be obtained on a small amount of data.