Intelligent reminder method, device and equipment based on large model

Abstract

The present application discloses a large-scale model-based intelligent reminder method, device, and equipment, which relates to the field of natural language processing. The method includes identifying the first conversation content input by the user through a dedicated model intelligent agent to obtain a first conditional execution rule corresponding to the reminder task. The dedicated model intelligent agent is obtained by fine-tuning the parameters of the pre-trained large model; converting the first conditional execution rule corresponding to the reminder task into a corresponding target CRON expression; establishing a connection channel between multiple user terminals of the user; when the triggering timing of the target CRON expression is met, based on the pre-configured execution mode and target CRON expression of the reminder task, actively triggering the reminder task. The reminder task is triggered in sequence according to the multi-level reminder matrix to actively send reminder information to the user through the connection channel of the user terminal corresponding to each level of reminder. The present application improves the cross-device collaboration capability by constructing a reminder matrix.

Description

Technical Field

[0001] The present application relates to the field of natural language processing technology, and in particular to a large model-based intelligent reminder method, device and equipment. Background Art

[0002] The intelligent systems in related technologies have problems such as shallow semantic understanding, mechanized task management, and insufficient multi-device collaboration. Summary of the Invention

[0003] The purpose of this application is to provide a large model-based intelligent reminder method, device, equipment, medium and product.

[0004] To achieve the above objectives, this application provides the following solutions:

[0005] In a first aspect, the present application provides an intelligent reminder method based on a large model, comprising:

[0006] Recognizing the first conversation content input by the user through a dedicated model agent to obtain a first conditional execution rule corresponding to the reminder task, wherein the dedicated model agent is obtained by fine-tuning parameters of a pre-trained large model;

[0007] Convert the first conditional execution rule corresponding to the reminder task into a corresponding target CRON expression;

[0008] establishing connection channels with multiple user terminals of the user;

[0009] When the triggering timing of the target CRON expression is met, the reminder task is actively triggered based on the pre-configured execution mode of the reminder task and the target CRON expression. The reminder task is triggered in sequence according to the multi-level reminder matrix to actively send reminder information to the user through the connection channel of the user terminal corresponding to each level of reminder.

[0010] In a second aspect, the present application provides a large model-based intelligent reminder device, comprising:

[0011] a recognition module configured to recognize the first conversation content input by the user through a dedicated model agent, and obtain a first conditional execution rule corresponding to the reminder task, wherein the dedicated model agent is obtained by fine-tuning parameters of a pre-trained large model;

[0012] A conversion module, configured to convert the first conditional execution rule corresponding to the reminder task into a corresponding target CRON expression;

[0013] An establishing module, configured to establish connection channels with multiple user terminals of the user;

[0014] An execution module is configured to actively trigger a reminder task based on a pre-configured execution mode of the reminder task and the target CRON expression when the triggering timing of the target CRON expression is met. The reminder task is triggered sequentially according to a multi-level reminder matrix to actively send reminder information to the user through a connection channel of a user terminal corresponding to each level of reminder.

[0015] In a third aspect, the present application provides a computer device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any one of the above-mentioned large-model-based intelligent reminder methods.

[0016] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any one of the above-mentioned large-model-based intelligent reminder methods.

[0017] In a fifth aspect, the present application provides a computer program product, comprising a computer program, which, when executed by a processor, implements the steps of any one of the above-mentioned large-model-based intelligent reminder methods.

[0018] According to the specific embodiments provided in this application, this application discloses the following technical effects:

[0019] The present application provides an intelligent reminder method, apparatus, equipment, medium and product based on a large model. The present application obtains a first conditional execution rule corresponding to a reminder task by identifying the content of a user's first conversation, and converts the first conditional execution rule into a target CRON expression, thereby more accurately identifying the user's intention; and actively triggering the reminder task when the triggering timing of the target CRON expression is met, so as to actively send a reminder message to the user, thereby realizing the automation and active management of the reminder task without the need for manual configuration; by constructing a cross-device reminder matrix, real-time synchronous reminders across devices can be realized, thereby improving the cross-device collaboration capability, and improving the accuracy, efficiency, real-timeness and reliability of the reminders. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0021] Figure 1 A flowchart of a large model-based intelligent reminder method provided in one embodiment of the present application;

[0022] Figure 2 A flowchart of a large model-based intelligent reminder method provided in another embodiment of the present application;

[0023] Figure 3 A flowchart of a large model-based intelligent reminder method provided in another embodiment of the present application;

[0024] Figure 4 A flowchart of a large model-based intelligent reminder method provided in yet another embodiment of the present application;

[0025] Figure 5 A flowchart of an intelligent reminder method based on a large model provided in another embodiment of the present application;

[0026] Figure 6 A flowchart of an intelligent reminder method based on a large model provided in another embodiment of the present application;

[0027] Figure 7 A flowchart of an intelligent reminder method based on a large model provided in another embodiment of the present application;

[0028] Figure 8 A flowchart of an intelligent reminder method based on a large model provided in another embodiment of the present application;

[0029] Figure 9 A schematic diagram of the functional modules of a large-scale model-based intelligent reminder device provided in one embodiment of the present application;

[0030] Figure 10 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0031] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0032] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0033] In an exemplary embodiment, Figure 1As shown, a large model-based intelligent reminder method is provided, which is executed by a computer device, specifically a computer device such as a terminal or a server, or a terminal and a server. In the embodiment of the present application, the method includes the following steps 102 to 108. Among them:

[0034] Step 102: Recognizing the first conversation content input by the user through a dedicated model agent to obtain a first conditional execution rule corresponding to the reminder task, wherein the dedicated model agent is obtained by fine-tuning parameters of a pre-trained large model;

[0035] In this approach, question-answer pair data can be first fine-tuned into the foundation of a large model to obtain a trained dedicated model. A dedicated model agent is then generated based on the trained dedicated model. A large model is a deep learning model with a large number of parameters, extensive training data, and high computational resource consumption. It is typically based on architectures such as the Transformer and is capable of handling complex natural language understanding (NLU), natural language generation (NLG), and multimodal tasks (such as image and text generation). A dedicated model agent can be generated based on the trained dedicated model and is capable of perceiving the environment and taking actions to achieve specific goals. It possesses autonomy, adaptability, and interactive capabilities. The agent perceives changes in the environment (e.g., through sensors or data input), makes judgments and decisions based on its learned knowledge and algorithms, and then executes actions to influence the environment or achieve a predetermined goal. The dedicated model agent can dynamically monitor the environment, invoke multimodal tools and trigger tasks, and adjust strategies based on real-time feedback. For example, if a user modifies a reminder time or changes their location, the conditional execution rules generated by the dedicated model agent will also change accordingly.

[0036] During real-time user conversations, the dedicated model agent recognizes the text content entered by the user in real time to extract reminder intentions and conditional execution rules. The reminder intentions include TRUE and FALSE. When the reminder intention is TRUE, it indicates that the user needs to be reminded. When the reminder intention is FALSE, it indicates that the user does not need to be reminded.

[0037] The first conversation content may be the original intention expressed by the user in real time. For example, the first conversation content may be "Remind me to drink water in 35 minutes.", "If it rains after school, tell my mom to pick me up.", or "Send an alarm when the temperature exceeds 50 degrees."

[0038] Conditional execution rules, also known as IF-THEN rules or trigger-action rules, are an automated execution mechanism based on logical judgment. Their core structure is: IF (condition satisfied) → THEN (action executed). This means that when a specific condition is detected, a preset action or process is automatically triggered. Each conditional execution rule can correspond to a reminder task. For example, if the first dialogue input by the user is "If it rains after school, tell Mom to pick me up," the corresponding conditional execution rule output by the dedicated model agent might be "IF school_time == trueAND weather == "rainy" THEN send_messag(mom, "Please pick up my child")." If the first dialogue input by the user is "Alarm when the temperature exceeds 50 degrees," the corresponding conditional execution rule output by the dedicated model agent might be "IF sensor_temp>50 THEN trigger_alarm("High temperature warning").

[0039] Step 104: converting the first conditional execution rule corresponding to the reminder task into a corresponding target CRON expression;

[0040] A CRON expression is a string format used to define the triggering time of periodic tasks. It is used for system scheduled tasks (such as backups, reminders, and data synchronization). Like a "time code," it uses concise code to describe complex time rules. A standard CRON expression contains 5 to 7 fields (this number may vary depending on the system). Each field represents a time unit, separated by spaces. The following example uses a 5-field CRON expression as an example: 1 2 3 4 5

[0042] │││││

[0043] * * * * *

[0044] The first field represents minutes (0-59), the second field represents hours (0-23), the third field represents the date (1-31), the fourth field represents the month (1-12 or JAN-DEC), and the fifth field represents the day of the week (0-7, where 0 and 7 both represent Sunday, or SUN-SAT). The CRON expression includes basic time parameters, date modifiers, and special marks. The basic time parameters are numeric or alphabetic parameters for each field. For example, 12 in the second field represents 12 o'clock, and JAN in the fourth field represents January. The date modifiers can include asterisks (*), question marks (?), minus signs (-), commas (,), slashes ( / ), L, W, #, LW combinations, and C. Symbols appearing in different fields have different meanings. For example, the # character usually appears only in the day of the week field, indicating a weekday of the month, and 5#3 indicates the third Friday of the month. Special marks include holiday marks, adjusted holiday marks, and seasonal marks.

[0045] Step 106: Establishing connection channels with multiple user terminals of the user;

[0046] Among them, the connection channel refers to a two-way communication link established between a server or other terminal device and multiple terminal devices of the same user. Each user terminal has an independent connection channel, and the user terminals can be devices of different levels.

[0047] Step 108: When the triggering timing of the target CRON expression is met, the reminder task is actively triggered based on the pre-configured execution mode of the reminder task and the target CRON expression. The reminder task is triggered in sequence according to the multi-level reminder matrix to actively send reminder information to the user through the connection channel of the user terminal corresponding to each level of reminder.

[0048] Among them, the trigger timing may include trigger time, trigger location and device status, etc. The trigger time may be a preset time, the trigger location may include home or company address, etc., and the device status may be device power, device online or offline, etc.; the confirmation of the trigger timing may be based on a dynamic adjustment task (such as "arrival at a designated location to trigger a reminder") based on geographic location, device status (mobile phone power), etc. For example, if the current time is the preset time, the user location is "company" and the mobile phone power is greater than 20%, a reminder is triggered.

[0049] Flexible task scheduling and lifecycle management can be achieved by configuring the execution mode of the reminder task; the reminder task can be dynamically scheduled, and by configuring the set execution mode, the execution mode of the reminder task can include automatic destruction after immediate execution (the corresponding first dialogue content can be "Turn off the lights in the living room now"), maintaining the task until the exit condition is met (the corresponding first dialogue content can be "Check the temperature every 10 minutes until it is below 25℃"), cumulative triggering, triggering after reaching a threshold number of times (the corresponding first dialogue content can be "Alarm after detecting 3 abnormal vibrations or remind me until I have exercised three times").

[0050] The multi-level reminder matrix achieves different degrees of reminder effects through the coordinated response of different device levels. The differences between different levels of reminders can be reflected in four dimensions: reach intensity, sensory intrusion, coverage reliability, and response urgency.

[0051] By implementing the above steps 102 to 108, the user's first conversation content is identified, and the first conditional execution rule corresponding to the reminder task is obtained, and the first conditional execution rule is converted into a target CRON expression, so that the user's intention can be identified more accurately; and when the triggering timing of the target CRON expression is met, the reminder task is actively triggered to actively send a reminder message to the user, thereby realizing the automation and active management of the reminder task without the need for manual configuration; by constructing a cross-device reminder matrix, real-time synchronous reminders across devices can be realized, thereby improving the cross-device collaboration capability and improving the accuracy, efficiency, real-timeness and reliability of the reminders.

[0052] In an exemplary embodiment of the present application, step 106 of "establishing connection channels with multiple user terminals of the user" includes: establishing a low-latency link for WebSocket duplex communication with multiple user terminals of the user, wherein the WebSocket duplex communication includes:

[0053] Step S1: Establish a connection. Through the HTTP Upgrade handshake (including the Upgrade: websocket header), the HTTP protocol is switched to the WebSocket protocol. The binary frame format (including masking and fragmentation control) is used for data transmission. The frame header is only 2-14 bytes, which significantly reduces the transmission overhead. The TCP long connection is kept alive to avoid repeated handshakes.

[0054] Step S2: Establish a latency optimization strategy, use CompositeByteBuf to merge multiple buffers, set the number of WorkerGroup threads to the number of CPU (Central Processing Unit) cores × 2 to avoid context switching, enable the permessage-deflate extension for protocol compression (reducing bandwidth usage), and use a cumulative window (50ms / 32KB) to aggregate small data packets for batch merging (effectively reducing the number of small packets in the network and reducing transmission overhead).

[0055] 50ms / 32KB represents the time window and space window. Small data packets within 50ms are sent with a delay, waiting to be merged, and are sent immediately when they reach 32KB.

[0056] In the embodiment of the present application, by using CompositeByteBuf to merge multiple buffers, the number of memory copies and system call overhead can be reduced. By setting the number of WorkerGroup threads = the number of CPU cores × 2, the multi-core CPU performance can be fully utilized to avoid context switching losses caused by too many threads. By enabling the permessage-deflate extension for protocol compression, bandwidth occupancy can be reduced. By aggregating small data packets through the accumulation window for batch merging, the number of small packets in the network can be effectively reduced, reducing transmission overhead.

[0057] The embodiment of the present application can compress the cross-device synchronization delay to 300ms, and the arrival rate of key scene reminders is 99.99%.

[0058] In another exemplary embodiment of the present application, Figure 2 As shown, step 104 can be replaced by the following steps 1041 to 1043:

[0059] Step 1041: using a parameter parser, parsing the structure of the first conditional execution rule corresponding to the reminder task to obtain a component structure;

[0060] Among them, a parameter parser is used to parse a structure similar to "every weekday [time] [excluding holidays]". The structure may include regular conditions, actions, exception conditions, etc. The regular conditions may include time conditions (minutes, hours, days, months, weeks, etc.), location conditions, device status conditions, weather conditions, etc. The actions may include closing, opening, checking, alarming, etc. The exception conditions may include time exceptions (excluding specific dates or time periods), status exceptions (skipping when the device is turned on, or not executing when the user is at home), or environmental exceptions (cancel watering if it rains), etc.

[0061] Step 1042: Assembling the component structure into a first CRON expression through a hierarchical replacement strategy based on a pre-established rule mapping table, wherein the rule mapping table is used to associate the first conditional execution rule with the CRON expression;

[0062] Among them, the hierarchical replacement strategy can be a regularized method for processing parameters at different levels in stages and according to priority. The hierarchical replacement strategy can include basic time parameter filling, date modifier processing and special tag injection. The following Table 1 is a rule mapping table. Based on the rule mapping table, the component structure in the first conditional rule can be assembled into the first CRON expression through the hierarchical replacement strategy.

[0063] Table 1 Rule mapping table

[0064]

[0065] Step 1043: Standardize the heterogeneous event source data in the first CRON expression to obtain the target CRON expression.

[0066] Heterogeneous event source data can be a collection of event data collected from different devices using different protocols, formats, structures, or semantics. This data is inherently heterogeneous due to its origins and requires standardization before it can be analyzed and responded to by a unified system. Standardization can involve converting heterogeneous event source data into a unified format. For example, heterogeneous event source data collected from IoT (Internet of Things) devices using MQTT (Message Queuing Telemetry Transport) persistent connections, HTTP WebSocket mobile apps, and third-party systems can be converted into a unified format.

[0067] In the embodiments of the present application, accurate conversion from business rules to executable tasks is achieved through rule parsing, hierarchical replacement and standardized processing.

[0068] In another exemplary embodiment of the present application, the user terminal includes at least one intelligent terminal layer device, a near field device layer device and a remote service layer device, such as Figure 3 As shown, the step 108 of "when the triggering timing of the target CRON expression is met, executing the reminder task based on the pre-configured execution mode of the reminder task and the target CRON expression" can be replaced by the following steps 1081 to 1083:

[0069] Based on the pre-configured execution mode of the reminder task, the following three phases are performed in sequence:

[0070] Step 1081: In the first stage, at least one of a pop-up reminder, a vibration reminder, and a light-emitting diode (LED) reminder is executed by at least one of the smart terminal layer devices;

[0071] Among them, the smart terminal layer device can be a user's personal device such as a mobile phone and / or a watch; the first stage can be within 300 milliseconds (unit: ms), and the first level reminder can be triggered within 300ms, and the mobile phone and / or watch can issue a pop-up reminder (for example, "It's time to drink water"), a vibration reminder, and an LED reminder.

[0072] Step 1082: In the second stage, at least one of a voice announcement and a screen flashing reminder is performed by at least one of the near-field device layer devices;

[0073] Among them, the near-field device layer devices can be peripheral devices of users such as smart speakers and car systems; the second stage can be within 5 seconds (unit: s), and the second-level reminder can be triggered within 5 seconds, the smart speaker will give a voice broadcast reminder, and the screen of the car system will flash a screen reminder.

[0074] Step 1083: In the third stage, at least one of automatic dialing reminder and email follow-up reminder is performed through at least one of the remote service layer devices.

[0075] Among them, the remote service layer device can be a telephone outbound call system or an email system; the third stage can be 1 minute, and the third level reminder can be triggered within 1 minute. The telephone outbound call system performs automatic dialing reminder, and the email system performs email follow-up reminder.

[0076] In the embodiment of the present application, efficient access to reminder tasks can be achieved through phased, graded, and multi-modal reminder methods, and reliable execution of reminder tasks can be achieved through multi-device collaboration.

[0077] In another exemplary embodiment of the present application, Figure 4 As shown, the intelligent reminder method further includes steps 1011 to 1016:

[0078] Step 1011: Select the second conversation content in the domain data conversation record, time series event table, and log;

[0079] Among them, the domain data conversation record can be a children's watch question and answer record (for example, it can be "wake me up at 3:20 pm"); the time series event table can be an event log that records timestamps (such as historical reminder records); the log can be a system operation log, which may be used to verify conditional execution rules; the second conversation content can be the user's historical question and answer record.

[0080] Step 1012: Generate a second conditional execution rule based on the spoken speech time in the second conversation content;

[0081] The spoken time may refer to a non-standardized time description expressed by a user through natural language, such as "a quarter of an hour later" or "tomorrow morning".

[0082] Step 1013: Execute a rule based on the second conversation content and the second condition to generate question-answer pair data;

[0083] The question-answer pair data may be a data pair consisting of a question and an answer generated by binding the user's second conversation content (i.e., the question) with the generated second conditional execution rule (i.e., the answer); illustratively, the question-answer pair data may include:

[0084] Question: "Remind me to drink water in 35 minutes"

[0085] Answer: IF time==15:30 THEN alert("drink water")

[0086] It should be noted that the intelligent reminder method of the embodiment of the present application supports cross-domain adaptation, pairs the time expression (question) in the user's original conversation with the automatically generated IF-THEN rule (answer) to generate question-answer pair data, which can be migrated from the field of children's watches to industrial operation and maintenance, smart home, medical care and other fields.

[0087] Step 1014: fine-tune the question-answer pair data into a large model base to obtain a dedicated model;

[0088] Among them, the large model base can be fine-tuned through question-answering data to obtain a dedicated model;

[0089] Step 1015: Setting an agent in a prompt manner based on the exclusive model to obtain the exclusive model agent;

[0090] Prompts are user-provided input instructions or questions that guide the model to generate specific responses. Their core purpose is to define the task scope, clarify the output format, or inject domain knowledge into the model.

[0091] Step 1016: Associate the dedicated model agent with the multimodal perception tool.

[0092] Among them, multimodal data includes voice, text, image, video and other data. The multimodal perception tool can be a tool that can process data of one or more modalities. The multimodal perception tool includes a voice recognition tool for processing voice data, a vision API (Application Programming Interface) that can process image data, a pop-up notification tool for processing text data and image data, and a weather API that can process voice data, text data, and image data.

[0093] Multimodal tools require the integration of speech recognition and speech synthesis, enabling standardized interaction via REST APIs. After data returned by multimodal tools is processed by dedicated model agents, multimodal responses can be generated according to pre-set templates, ensuring end-to-end controllability and a consistent user experience.

[0094] In this application, highly adaptable and feasible domain intelligent agent development is achieved through domain data fusion, rule-guided data generation, model fine-tuning, intelligent agent construction, and multimodal expansion.

[0095] The embodiments of this application solve the limitations of related systems in complex intent recognition, task management, and cross-device collaboration through deep semantic analysis, dynamic task engines, and multimodal collaboration technologies. The core of the embodiments of this application is to improve semantic understanding capabilities. The accuracy of parsing complex expressions (such as "two hours after taking the medicine") is significantly improved, dynamic logic is quickly generated, conditional statements are converted with one click, and the setup time can be reduced by 80%. It realizes dynamic generation and adjustment of tasks, and builds a low-latency cross-device reminder matrix. Multiple devices collaborate efficiently and respond synchronously at 300ms level, significantly improving the accuracy, efficiency, and real-time performance of the reminder system. It is suitable for various scenarios such as medical monitoring and business management.

[0096] In another exemplary embodiment of the present application, Figure 5 As shown, step 1012 can be replaced by the following steps 10121 to 10127:

[0097] Step 10121: Convert the spoken time in the second dialogue content into absolute time;

[0098] Among them, absolute time is usually time expressed in specific numbers or standard time units. For example, the spoken time is "three quarters of an hour later" and the current time is 14:45, then the absolute time corresponding to "three quarters of an hour later" is 15:30.

[0099] Step 10122: Identify conditional keywords in the second conversation content;

[0100] The “conditional keywords” include “if”, “when”, etc.; the conditional keywords in the input text can be parsed using a natural language processing (NLP) model.

[0101] Step 10123: Based on the absolute time and the conditional keywords, a bidirectional attention mechanism is used to identify entity associations and temporal logic in the second conversation content to obtain semantic roles in the second conversation content.

[0102] Among them, the bidirectional attention mechanism is a key attention mechanism design in natural language processing. By simultaneously considering forward (left to right) and backward (right to left) context information, it can more comprehensively capture the dependencies between elements in the sequence. Identifying entity associations in the second dialogue content refers to identifying hidden relationships between different entities (people, objects, events) in the text and establishing logical connections; identifying temporal logic refers to analyzing the time sequence or dependency between events to ensure that actions are triggered at the correct time point; the bidirectional attention mechanism can be used to analyze the absolute time and the conditional keywords to identify entity associations and temporal logic in the second dialogue content; semantic roles include conditions and actions. For example, the second dialogue content is "turn off all lights after 10 pm", and the identified condition in the second dialogue content is "10 pm", and the action is "turn off all lights."

[0103] Step 10124: Building a structured logical framework of conditions and actions based on the semantic roles;

[0104] Among them, a condition-action logical framework is then constructed based on semantic role labeling (condition "10 p.m.", action "turn off all lights"), and fuzzy expressions (such as "the value is too high") can also be mapped to precise threshold comparison expressions (for example, "IF the temperature > 30℃ THEN...").

[0105] Step 10125: Generate a third conditional execution rule based on the structured logic framework and the rule template engine;

[0106] Among them, the structured logic can be converted into executable code (i.e., the third condition execution rule) through the rule template engine.

[0107] Step 10126: Optimize the rule weights and conflict resolution strategies in the third conditional execution rule using a genetic algorithm to obtain a fourth conditional execution rule.

[0108] For example, the third conditional execution rule may include Rule A and Rule B. Rule A: "Turn on lights when arriving home" (weight 0.6), Rule B: "Turn off all lights after 10 pm" (weight 0.8). The conflict is: if arriving home after 10 pm, should the lights be turned on or off? A genetic algorithm can use historical data to test the effectiveness of each set of weights, retaining the best ones and eliminating the worst ones to generate a new generation of weights, ultimately achieving the optimal weights. For example, the weight of Rule B can be increased to 0.9, giving it priority execution.

[0109] Step 10127: Use a test case to perform boundary verification on the fourth conditional execution rule. After the verification passes, determine the fourth conditional execution rule as the second conditional execution rule.

[0110] Among them, the fourth conditional execution rule can be tested with extreme and abnormal examples to ensure that the generated fourth conditional execution rule meets the system execution requirements while retaining semantic accuracy. The test cases may include time boundary test cases, numerical boundary test cases, and logical boundary test cases. The time boundary test case may be to set a "reminder at 23:59 every day" and verify whether an error occurs when crossing midnight; the numerical boundary test case may be "alarm when temperature > 30°C" and verify what happens if the temperature is exactly 30.0°C; the logical boundary test case may be to determine what to do when the user says "If I am not at home and it is raining...", but "I am not at home" and "it is raining" are not true at the same time. If the verification fails, the fourth conditional execution rule can be modified or exception handling can be added based on the test results. The method of the embodiment of the present application can handle nested compound conditions (such as "A and (B or C)").

[0111] In an embodiment of the present application, by performing time sequence standardization, semantic parsing, rule generation and optimization verification on the second conversation content, a second conditional execution rule is obtained, thereby making the constructed second conditional execution rule more accurate.

[0112] In another exemplary embodiment of the present application, Figure 6 As shown, the question-answer pair data includes a question text set and an answer text set, and step 1014 can be replaced by the following steps 10141 to 10148:

[0113] Step 10141: Structural processing is performed on the question-answer pair data to obtain a first data set including an input sequence and an output sequence, wherein the input sequence is obtained by embedding the question text set into a preset input template, and the output sequence is obtained by embedding the answer text set into a preset output template;

[0114] The question-and-answer data can be structured, using the generative model type with a "question: {text}" input template and an "answer: {text}" output template format. The contextual question and answer are combined with the relevant text as input, and the output is the answer text or the answer's location index. Alternatively, multiple rounds of conversation data can be spliced together using historical conversations as input. For example, the input sequence might be "question: {remind me in 35 minutes}," and the output sequence might be "answer: {IF current_time == 15:30 THEN alert('drink water')"}."

[0115] Step 10142: performing data preprocessing including normalization, truncation, and alignment and segmentation on the first data set to obtain a second data set;

[0116] Among them, standardization processing includes unifying upper and lower cases, and tokens longer than 512 can be truncated.

[0117] Step 10143: performing data enhancement processing including inverse problem processing and synonym replacement processing on the second data set to obtain a third data set;

[0118] For example, from the answer "IF time == 15:30 THEN remind me to drink water," we can infer the questions "What time do you remind me to drink water?", "What do I need to do at 15:30?", or "Help me set a reminder to drink water." We can also replace the question "Remind me to take medicine every 4 hours" with "Notify me to take medicine every 4 hours." The augmented second dataset can also be converted into a third dataset file in JSON format.

[0119] Step 10144: Use the LlamaFactory framework as a training tool and the Llama3.2-8B model as the large model base;

[0120] Among them, the Llama3.2-8B model is an open source large model with 8 billion parameters that can balance performance and computing power requirements. The LlamaFactory framework is an efficient training tool optimized for the Llama series that can simplify the training process.

[0121] Step 10145: Based on the low-rank adaptation technology LoRA, the Attention layer parameters of the large model base are updated through the low-rank matrix, and a lightweight module is inserted into the adapter to freeze the original parameters of the Llama3.2-8B model;

[0122] Among them, parameter-efficient technology (LoRA, Low-Rank Adaptation) can be used according to the size of the dataset, the attention layer parameters can be updated through the low-rank matrix, and adapters can be inserted into lightweight modules to freeze the original model.

[0123] Step 10146: Using a hint fine-tuning technique to guide semantic alignment of the output dataset of the completed large model base with the third dataset using a learnable vector;

[0124] Step 10147: Perform mixed precision training on the large model base using the AdamW optimizer, cosine decay learning rate, and gradient clipping constraints to obtain a trained model.

[0125] Among them, the norm of the gradient can be ≤1.0.

[0126] Step 10148: Verify the trained model. If the verification is successful, determine the trained model as an exclusive model.

[0127] The trained model can be verified after weight deployment.

[0128] In the embodiments of the present application, data preprocessing can be used to improve data quality and computing efficiency; data enhancement processing can be used to improve model generalization ability, cover edge scenarios, and increase data diversity; LoRA and adapters can be used to reduce the amount of training parameters; prompt fine-tuning technology can be used to more efficiently adapt parameters and more accurately achieve semantic alignment; the AdamW optimizer can prevent overfitting; and learning rate adjustment and gradient clipping constraints can be used to improve training stability and save computing resources.

[0129] In another exemplary embodiment of the present application, Figure 7 As shown, the step 10148 of "verifying the trained model" includes the following steps 201 to 205:

[0130] Step 201: Based on BLEU and / or ROUGE, perform generation task verification on the trained model;

[0131] Among them, BLEU can measure the vocabulary overlap between the generated answers output by the trained model and the reference answers, and ROUGE can measure the longest common subsequence between the generated answers output by the trained model and the reference answers.

[0132] Step 202: Based on the exact match value and F1 (taking into account both precision and recall) value, the trained model is verified for the extraction task;

[0133] The Exact Match (EM) value requires that the generated answer is identical to the reference answer in order to be returned. A successful match returns 1, while an unsuccessful match returns 0. The F1 score measures the classification performance of a trained model. It is the harmonic mean of precision and recall, balancing the two and providing a more comprehensive evaluation result.

[0134] Step 203: Based on the classification accuracy and / or ROC-AUC value, the trained model is verified for the classification task;

[0135] Among them, the classification accuracy is used to measure the proportion of samples correctly predicted by the trained model to the total number of samples; the ROC-AUC value is used to measure the ability of the trained model to distinguish between positive and negative samples.

[0136] Step 204: Obtaining a comprehensive verification result based on at least one of the generation task verification, the extraction task verification, the classification task verification, and the verification result of the manual verification;

[0137] Among them, manual sampling evaluation can be carried out, random sampling can be used to generate results, and logical consistency can be manually judged (such as whether the answer is irrelevant to the question).

[0138] It should be noted that staged validation and adversarial validation can also be performed. Staged validation can evaluate every 100-500 steps to monitor loss and indicator fluctuations. Adversarial validation can check the distribution differences between the validation set and the training set to avoid data leakage.

[0139] Step 205: If the comprehensive verification result meets the preset conditions, determine that the verification is passed.

[0140] For example, a comprehensive verification result can be obtained based on the verification results of the generated task verification and the manual verification; when BLEU>25, ROUGE>40, the manual evaluation pass rate ≥80%, and the answer is logically correct and has no factual errors, the test is passed and the exclusive model training is completed.

[0141] In the embodiment of the present application, multi-task joint verification of generation tasks, extraction tasks and classification tasks can be performed, and combined with the verification results of manual verification to obtain comprehensive verification results, so that the model can be verified from multiple angles. Only after the verification is passed, the trained model is determined as an exclusive model, thereby improving the accuracy and reliability of the model output answers and improving the generalization performance of the model.

[0142] In one embodiment, the prompt includes task type, calling tool, step structured instruction template, example pair and output format constraint, such as Figure 8As shown, step 1015 includes steps 10151 to 10155:

[0143] Step 10151: Identify the task type of the reminder task through structured markup in the Prompt prompt;

[0144] Prompt can be a pre-designed structured instruction template that guides the dedicated model to understand the task, invoke tools, and generate standardized outputs. Its core function is like a "detailed operating manual for the model," ensuring that the intelligent agent accurately executes tasks in a multimodal environment. For example, if the second dialogue content is "If the temperature is below 10 degrees tomorrow morning, remind me to wear a down jacket using voice and pop-up notifications," the Prompt can clearly inform the model that the current task type is a "temperature condition reminder task."

[0145] Step 10152: Embed a tool registry in the prompt, wherein the tool registry defines the name, input format, and permissions of the calling tool;

[0146] The Prompt may call a weather API to check the temperature tomorrow morning, and prepare a speech synthesis tool and a pop-up notification tool to use voice reminders and pop-up reminders for subsequent reminders.

[0147] Step 10153: guiding the dedicated model to operate step by step through the process instructions of the step-structured instruction template in the Prompt;

[0148] Step 10154: inserting an example pair into the prompt to show the mapping relationship between input and output;

[0149] Step 10155: Declare output format constraints in the Prompt.

[0150] The format constraint may be a JSON format constraint, which forces the model to generate results in a specified format (such as JSON format) to facilitate subsequent parsing.

[0151] When building an agent based on a proprietary model, the prompt for the reminder agent is designed as follows: a structured instruction template for analyzing task types, invoking tools, and integrating results, along with example pairs and JSON format constraints. The agent architecture utilizes a hierarchical decision-making process: after the model interprets user intent, it calls the engine-matched tool through a tool invocation mechanism. Failures trigger retries or downgraded responses, and error logging optimizes the process. Multimodal perception tool association requires the integration of speech recognition and speech synthesis (for voice replies), with standardized interactions implemented through a REST API. The data returned by the multimodal perception tool is then processed by the model and a multimodal response is generated according to a preset template, ensuring end-to-end controllability and a consistent user experience.

[0152] In the embodiment of the present application, by adding Prompt on the basis of the exclusive model, task orientation can be performed more accurately according to the task type at the identification location; by integrating external multimodal perception tools, the model capabilities can be expanded; through step-structured instruction templates, complex tasks can be decomposed and multi-step reasoning automation can be achieved; through example pairs, analogy prompts can be given to the model to activate the model's pattern recognition ability, accelerate the model's learning, improve the accuracy of the answer, and quickly adapt to new tasks; by constraining the output format, the output quality can be controlled.

[0153] Based on the same inventive concept, the present application also provides a large-model-based intelligent reminder device for implementing the large-model-based intelligent reminder method mentioned above. The implementation solution provided by this device is similar to the implementation solution described in the above method. Therefore, the specific limitations of one or more large-model-based intelligent reminder device embodiments provided below can be found in the above limitations of the large-model-based intelligent reminder method, and will not be repeated here.

[0154] In an exemplary embodiment, Figure 9 As shown, a smart reminder device 300 based on a large model is provided, including:

[0155] Identification module 301, configured to identify the first conversation content input by the user through a dedicated model agent, and obtain a first conditional execution rule corresponding to the reminder task, wherein the dedicated model agent is obtained by fine-tuning parameters of a pre-trained large model;

[0156] A conversion module 302 is configured to convert the first conditional execution rule corresponding to the reminder task into a corresponding target CRON expression;

[0157] An establishing module 303 is configured to establish connection channels with multiple user terminals of the user;

[0158] The execution module 304 is used to actively trigger the reminder task based on the pre-configured execution mode of the reminder task and the target CRON expression when the triggering timing of the target CRON expression is met. The reminder task is triggered in sequence according to the multi-level reminder matrix to actively send reminder information to the user through the connection channel of the user terminal corresponding to each level of reminder.

[0159] As an optional implementation, the conversion module 302 includes: a parsing submodule, which is used to use a parameter parser to parse the structure in the first conditional execution rule corresponding to the reminder task to obtain a component structure; an assembling submodule, which is used to assemble the component structure into a first CRON expression through a hierarchical replacement strategy based on a pre-established rule mapping table, and the rule mapping table is used to associate the first conditional execution rule and the CRON expression; a first processing submodule, which is used to standardize the heterogeneous event source data in the first CRON expression to obtain the target CRON expression.

[0160] As an optional implementation, the user terminal includes at least one smart terminal layer device, a near-field device layer device and a remote service layer device, and the execution module 304 includes: a first reminder sub-module, which is used to execute the following three steps in sequence based on the pre-configured execution mode of the reminder task: in the first stage, at least one of pop-up reminders, vibration reminders and light-emitting diode (LED) reminders is executed through at least one of the smart terminal layer devices; a second reminder sub-module is used to execute at least one of voice broadcast reminders and screen flashing reminders through at least one of the near-field device layer devices in the second stage; a third reminder sub-module is used to perform at least one of automatic dialing reminders and email follow-up reminders through at least one of the remote service layer devices in the third stage.

[0161] As an optional implementation, the device also includes: a selection module for selecting a second dialogue content in domain data dialogue records, time series event tables, and logs; a first generation module for generating a second conditional execution rule based on the spoken speech time in the second dialogue content; a second generation module for generating question-answer pair data based on the second dialogue content and the second conditional execution rule; a training module for training the question-answer pair data into a large model base in a fine-tuning manner to obtain an exclusive model; a setting module for setting an intelligent agent in a prompt manner on the basis of the exclusive model to obtain the exclusive model intelligent agent; and an association module for associating the exclusive model intelligent agent with a multimodal perception tool.

[0162] As an optional implementation, the first generation module includes: a conversion submodule for converting the colloquial time in the second dialogue content into absolute time; a first recognition submodule for identifying conditional keywords in the second dialogue content; a second recognition submodule for identifying entity associations and temporal logic in the second dialogue content based on the absolute time and the conditional keywords using a bidirectional attention mechanism to obtain semantic roles in the second dialogue content; a construction submodule for constructing a structured logical framework of conditions and actions based on the semantic roles; a first generation submodule for generating a third conditional execution rule based on the structured logical framework and a rule template engine; a second generation submodule for optimizing the rule weights and conflict resolution strategies in the third conditional execution rule using a genetic algorithm to obtain a fourth conditional execution rule; a first verification submodule for performing boundary verification on the fourth conditional execution rule using a test case, and after passing the verification, determining the fourth conditional execution rule as the second conditional execution rule.

[0163] As an optional implementation, the question-answer pair data includes a question text set and an answer text set, and the training module includes: a second processing submodule for performing structured processing on the question-answer pair data to obtain a first data set including an input sequence and an output sequence, wherein the input sequence is obtained by embedding the question text set into a preset input template, and the output sequence is obtained by embedding the answer text set into a preset output template; a third processing submodule for performing data preprocessing including standardization processing, truncation processing, and alignment and segmentation processing on the first data set to obtain a second data set; a fourth processing submodule for performing data enhancement processing including inverse question processing and synonym replacement processing on the second data set to obtain a third data set; an updating submodule for adopting the LlamaFactory framework As a training tool, the Llama3.2-8B model is used as the large model base; based on the low-rank adaptive technology LoRA, the Attention layer parameters of the large model base are updated through the low-rank matrix, and a lightweight module is inserted into the adapter to freeze the original parameters of the Llama3.2-8B model; the alignment submodule is used to adopt the prompt fine-tuning technology to guide the semantic alignment of the output data set of the completed large model base with the learnable vector; the training submodule is used to combine the AdamW optimizer, cosine decay learning rate and gradient clipping constraint to perform mixed precision training on the large model base to obtain the trained model; the second verification submodule is used to verify the trained model. If the verification passes, the trained model is determined as an exclusive model.

[0164] As an optional embodiment, the second verification submodule includes: a first verification unit for performing generation task verification on the trained model based on BLEU and / or ROUGE; a second verification unit for performing extraction task verification on the trained model based on exact match value and F1 score; a third verification unit for performing classification task verification on the trained model based on classification accuracy and / or ROC-AUC value; a fourth verification unit for obtaining a comprehensive verification result based on at least one of the generation task verification, the extraction task verification, and the classification task verification, as well as the manual verification results; and a fifth verification unit for determining that verification has passed if the comprehensive verification result meets preset conditions.

[0165] As an optional implementation, the Prompt prompt includes a task type, a calling tool, a step structured instruction template, an example pair, and an output format constraint, and the setting module includes: a third identification submodule, used to identify the task type of the reminder task through structured tags in the Prompt prompt; an embedding submodule, used to embed a tool registry in the Prompt, wherein the tool registry defines the name, input format, and permissions of the calling tool; a guiding submodule, used to guide the exclusive model to operate step by step through the process-based instructions of the step structured instruction template in the Prompt; an inserting submodule, used to insert an example pair in the Prompt to show the mapping relationship between input and output; and a declaration submodule, used to declare output format constraints in the Prompt.

[0166] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Figure 10 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O) and a communication interface. The processor, memory and input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device is used to store data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a large model-based intelligent reminder method is implemented.

[0167] Those skilled in the art will understand that Figure 10 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0168] In an exemplary embodiment, a computer device is further provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps in the above method embodiments when executing the computer program.

[0169] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0170] In an exemplary embodiment, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0171] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0172] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. In particular, any reference to memory, database, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0173] The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processors involved in the various embodiments provided herein may include, but are not limited to, general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic units, data processing logic units based on quantum computing, and the like.

[0174] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0175] This document uses specific examples to illustrate the principles and implementation methods of this application. The description of the above examples is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. An intelligent reminder method based on a large model, characterized in that: The intelligent reminder method based on the large model includes: Recognizing the first conversation content input by the user through a dedicated model agent to obtain a first conditional execution rule corresponding to the reminder task, wherein the dedicated model agent is obtained by fine-tuning parameters of a pre-trained large model; Convert the first conditional execution rule corresponding to the reminder task into a corresponding target CRON expression; establishing connection channels with multiple user terminals of the user; When the triggering timing of the target CRON expression is met, the reminder task is actively triggered based on the pre-configured execution mode of the reminder task and the target CRON expression. The reminder task is triggered in sequence according to the multi-level reminder matrix to actively send reminder information to the user through the connection channel of the user terminal corresponding to each level of reminder; The method for constructing the exclusive model agent includes: Select the second conversation content in the domain data conversation record, time series event table, and log; converting the spoken time in the second dialogue content into absolute time; identifying conditional keywords in the second conversation content; Based on the absolute time and the conditional keywords, using a bidirectional attention mechanism to identify entity associations and temporal logic in the second dialogue content to obtain semantic roles in the second dialogue content; Based on the semantic roles, a structured logical framework of conditions and actions is constructed; Based on the structured logic framework and the rule template engine, generating a third conditional execution rule; Using a genetic algorithm to optimize the rule weights and conflict resolution strategies in the third conditional execution rule to obtain a fourth conditional execution rule; Using a test case to perform boundary verification on the fourth conditional execution rule, and after passing the verification, determining the fourth conditional execution rule as the second conditional execution rule; Executing a rule based on the second conversation content and the second condition to generate question-answer pair data; The question-answer pair data is trained into a large model base in a fine-tuning manner to obtain a dedicated model; On the basis of the exclusive model, an agent is set in a prompt manner to obtain the exclusive model agent.

2. The intelligent reminder method based on a large model according to claim 1 is characterized in that: The step of converting the first conditional execution rule corresponding to the reminder task into a corresponding target CRON expression includes: Using a parameter parser, the structure in the first conditional execution rule corresponding to the reminder task is parsed to obtain a component structure; Based on a pre-established rule mapping table, assembling the component structure into a first CRON expression through a hierarchical replacement strategy, wherein the rule mapping table is used to associate the first conditional execution rule with the CRON expression; The heterogeneous event source data in the first CRON expression is standardized to obtain the target CRON expression.

3. The intelligent reminder method based on a large model according to claim 1 is characterized in that: The user terminal includes at least one intelligent terminal layer device, a near field device layer device, and a remote service layer device. When the triggering timing of the target CRON expression is met, the reminder task is actively triggered based on the pre-configured execution mode of the reminder task and the target CRON expression, including: Based on the pre-configured execution mode of the reminder task, the following three phases are performed in sequence: In the first stage, at least one of a pop-up reminder, a vibration reminder, and a light-emitting diode (LED) reminder is executed by at least one of the smart terminal layer devices; In the second stage, at least one of a voice broadcast reminder and a screen flashing reminder is performed by at least one of the near-field device layer devices; In the third stage, at least one of automatic dialing reminder and email follow-up reminder is performed through at least one of the remote service layer devices.

4. The intelligent reminder method based on a large model according to claim 1 is characterized in that: The intelligent reminder method based on the large model also includes: Associating the dedicated model agent with a multimodal perception tool.

5. The intelligent reminder method based on a large model according to claim 4 is characterized in that: The question-answer pair data includes a question text set and an answer text set. The question-answer pair data is fine-tuned into a large model base to obtain a dedicated model, including: Performing structured processing on the question-answer pair data to obtain a first data set comprising an input sequence and an output sequence, wherein the input sequence is obtained by embedding the question text set into a preset input template, and the output sequence is obtained by embedding the answer text set into a preset output template; Performing data preprocessing including standardization, truncation, and alignment and segmentation on the first data set to obtain a second data set; performing data enhancement processing including inverse problem processing and synonym replacement processing on the second data set to obtain a third data set; The LlamaFactory framework is used as a training tool, and the Llama3.2-8B model is used as the large model base; Based on the low-rank adaptation technology LoRA, the Attention layer parameters of the large model base are updated through the low-rank matrix, and a lightweight module is inserted into the adapter to freeze the original parameters of the Llama3.2-8B model; Using a hint fine-tuning technique to guide the completion of semantic alignment of the output dataset of the large model base and the third dataset using a learnable vector; Perform mixed precision training on the large model base using the AdamW optimizer, cosine decay learning rate, and gradient clipping constraints to obtain a trained model. The trained model is verified, and if the verification passes, the trained model is determined as an exclusive model.

6. The intelligent reminder method based on a large model according to claim 5 is characterized in that: The verifying of the trained model includes: Performing generation task verification on the trained model based on BLEU and / or ROUGE; Based on the exact match value and F1 value, the trained model is verified for the extraction task; Performing classification task verification on the trained model based on the classification accuracy and / or ROC-AUC value; Obtaining a comprehensive verification result based on at least one of the generation task verification, the extraction task verification, the classification task verification, and the verification results of the manual verification; If the comprehensive verification result meets the preset conditions, it is determined that the verification is passed.

7. The intelligent reminder method based on a large model according to claim 1 is characterized in that: The prompt includes a task type, a calling tool, a step structured instruction template, an example pair, and an output format constraint. The setting of an agent in the form of a prompt on the basis of the exclusive model to obtain the exclusive model agent includes: Identify the task type of the reminder task through structured markup in the Prompt prompt; Embed a tool registry in the prompt, wherein the tool registry defines the name, input format and permissions of the calling tool; In the Prompt, the dedicated model is guided to operate step by step through the process instructions of the step-structured instruction template; Insert example pairs into the prompt to show the mapping relationship between input and output; Declare output format constraints in the prompt.

8. An intelligent reminder device based on a large model, characterized in that: The intelligent reminder device based on the large model includes: a recognition module configured to recognize the first conversation content input by the user through a dedicated model agent, and obtain a first conditional execution rule corresponding to the reminder task, wherein the dedicated model agent is obtained by fine-tuning parameters of a pre-trained large model; A conversion module, configured to convert the first conditional execution rule corresponding to the reminder task into a corresponding target CRON expression; An establishing module, configured to establish connection channels with multiple user terminals of the user; an execution module, configured to proactively trigger a reminder task based on a pre-configured execution mode of the reminder task and the target CRON expression when a triggering timing of the target CRON expression is met, wherein the reminder task is triggered sequentially according to a multi-level reminder matrix to proactively send a reminder message to the user through a connection channel of a user terminal corresponding to each level of reminder; A construction module is provided for selecting a second conversation content from a domain data conversation record, a time series event table, or a log; converting the spoken time in the second conversation content into an absolute time; identifying conditional keywords in the second conversation content; using a bidirectional attention mechanism to identify entity associations and temporal logic in the second conversation content based on the absolute time and the conditional keywords, so as to obtain semantic roles in the second conversation content; constructing a structured logic framework of conditions and actions based on the semantic roles; generating a third conditional execution rule based on the structured logic framework and a rule template engine; optimizing the rule weights and conflict resolution strategies in the third conditional execution rule using a genetic algorithm to obtain a fourth conditional execution rule; performing boundary verification on the fourth conditional execution rule using a test case, and determining the fourth conditional execution rule as the second conditional execution rule after passing the verification; generating question-answer pair data based on the second conversation content and the second conditional execution rule; training the question-answer pair data into a large model base in a fine-tuning manner to obtain a dedicated model; and setting an intelligent agent in a prompt manner based on the dedicated model to obtain the dedicated model intelligent agent.

9. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the large model-based intelligent reminder method according to any one of claims 1 to 7.

Images