Request processing method and device based on fast and slow double-inference system, and electronic equipment

Abstract

The invention provides a request processing method and device based on a fast and slow double inference system and electronic equipment, and relates to the technical field of artificial intelligence such as large models, multi-agent collaboration, inference systems and content indexing. The method comprises the steps that a preset fast reasoning system is used for processing an instant request initiated by a first user to obtain an instant response, and the fast reasoning system generates the instant response based on the content index technology; in response to the fact that the matching degree between associated knowledge used for generating the instant response and the instant request is lower than the preset matching degree, the instant request is processed through a preset slow reasoning system, a request reply is obtained, and the associated knowledge is determined and obtained based on information related to the instant request extracted from a shared information base through the content index technology; the slow reasoning system generates a request reply based on task disassembly and multi-intelligence cooperation; and supplementing the new knowledge contained in the request reply to the shared information base. According to the method, the balance between the request processing timeliness and the reply quality is realized through the collaborative architecture of the fast and slow double-inference system.

Description

Technical Field

[0001] This disclosure relates to the field of data processing technology, specifically to artificial intelligence technologies such as large models, multi-agent collaboration, inference systems, and content indexing, and particularly to a request processing method, apparatus, electronic device, computer-readable storage medium, and computer program product based on a fast and slow dual inference system. Background Technology

[0002] In scenarios where users interact with Large Language Models (LLMs), users have high demands for the timeliness of responses and the accuracy and depth of replies, but existing technologies struggle to meet both requirements simultaneously.

[0003] Furthermore, existing solutions lack effective knowledge iteration mechanisms, resulting in an inability to continuously evolve and hindering long-term processing capabilities. Summary of the Invention

[0004] This disclosure presents a request processing method, apparatus, electronic device, computer-readable storage medium, and computer program product based on a fast and slow dual inference system.

[0005] In a first aspect, embodiments of this disclosure propose a request processing method based on a fast and slow dual-inference system, comprising: processing an immediate request initiated by a first user using a preset fast inference system to obtain an immediate response; wherein the fast inference system generates an immediate response for the immediate request based on content indexing technology; in response to the fact that the matching degree between the associated knowledge used to generate the immediate response and the immediate request is lower than a preset matching degree, processing the immediate request using a preset slow inference system to obtain a request reply; wherein the associated knowledge is determined based on extracting information related to the immediate request from a preset shared information database through content indexing technology, and the slow inference system generates a request reply for the immediate request based on task decomposition and multi-intelligence collaboration; and supplementing the shared information database with the new knowledge contained in the request reply.

[0006] Secondly, embodiments of this disclosure propose a request processing apparatus based on a fast and slow dual-inference system, comprising: a fast inference system processing unit configured to process an immediate request initiated by a first user using a preset fast inference system to obtain an immediate response; wherein the fast inference system generates an immediate response for the immediate request based on content indexing technology; a slow inference system supplementary processing unit configured to process the immediate request using a preset slow inference system in response to a lower-than-preset matching degree between the associated knowledge used to generate the immediate response and the immediate request; wherein the associated knowledge is determined based on information related to the immediate request extracted from a preset shared information database through content indexing technology, and the slow inference system generates a request response for the immediate request based on task decomposition and multi-intelligence collaboration; and a first knowledge supplementary unit configured to supplement the shared information database with new knowledge contained in the request response.

[0007] Thirdly, embodiments of this disclosure provide an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to implement the request processing method based on a fast and slow dual inference system as described in the first aspect.

[0008] Fourthly, embodiments of this disclosure provide a non-transitory computer-readable storage medium storing computer instructions that enable a computer to implement the request processing method based on a fast and slow dual inference system as described in the first aspect.

[0009] Fifthly, embodiments of this disclosure provide a computer program product including a computer program that, when executed by a processor, can implement the steps of the request processing method based on a fast and slow dual inference system as described in the first aspect.

[0010] The request processing scheme based on a fast and slow dual-inference system disclosed herein achieves a balance between request processing timeliness and response quality through the collaborative architecture of the two systems. The fast inference system responds to user requests instantly based on content indexing technology, ensuring immediacy. Simultaneously, by judging the matching degree between related knowledge and requests, requests with low matching degrees are further transferred to the slow inference system, rationally allocating computing resources while avoiding unnecessary deep inference overhead. The slow inference system, employing task decomposition and multi-intelligence collaboration, can perform in-depth processing of complex requests, effectively improving the accuracy and depth of responses compared to the fast inference system. Furthermore, by supplementing the shared information base with the new knowledge generated by slow inference, the fast inference system can gradually improve the quality and accuracy of its rapid responses in subsequent response generation processes due to the continuous enrichment of related knowledge. This enables the continuous evolution of the request processing system, and in the long run, it improves the response quality of fast inference, reduces the frequency of slow inference calls, and optimizes the overall efficiency and effectiveness of request processing.

[0011] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0012] Other features, objects, and advantages of this disclosure will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is an exemplary system architecture to which this disclosure can be applied; Figure 2 A flowchart of a request processing method based on a fast and slow dual inference system provided in this disclosure embodiment; Figure 3 A flowchart illustrating a method for processing immediate requests using a fast inference system, provided as an embodiment of this disclosure; Figure 4 A flowchart illustrating a method for processing immediate requests using a slow inference system, provided as an embodiment of this disclosure; Figure 5 A flowchart of a method for determining the complete task corresponding to an immediate request using an agent constituting a slow inference system, provided in an embodiment of this disclosure; Figure 6 A flowchart illustrating a method for processing problem investigation requests and performing knowledge updates using a slow inference system, as provided in this embodiment of the disclosure; Figure 7 A flowchart illustrating a method for generating and processing problem investigation results using multiple intelligent agents, as provided in this embodiment of the disclosure; Figure 8A flowchart illustrating a method for deleting knowledge from a shared information base and triggering updates based on the amount of knowledge stored, provided in this embodiment of the disclosure; Figure 9 A schematic diagram of a human-machine co-creation system architecture based on end-to-end multi-agent technology for sustainable evolution is provided in this embodiment of the disclosure. Figure 10 A structural block diagram of a request processing device based on a fast and slow dual inference system provided in this disclosure embodiment; Figure 11 This is a schematic diagram of the structure of an electronic device suitable for executing a request processing method based on a fast and slow dual inference system, provided as an embodiment of the present disclosure. Detailed Implementation

[0013] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding; these should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description. It should be noted that, unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0014] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0015] Figure 1 An exemplary system architecture 100 is shown, to which embodiments of the request processing method, apparatus, electronic device, and computer-readable storage medium based on the fast and slow dual inference system disclosed herein can be applied.

[0016] like Figure 1 As shown, system architecture 100 may include terminal devices 101, 102, and 103, a network 104, and a server 105. Network 104 serves as the medium for providing communication links between terminal devices 101, 102, and 103 and server 105. Network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0017] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Various applications for enabling information communication between the terminal devices 101, 102, and 103 and server 105 can be installed. These applications include dialogue and question-and-answer applications, large model update applications, and instant messaging applications.

[0018] Terminal devices 101, 102, and 103 and server 105 can be either hardware or software. When terminal devices 101, 102, and 103 are hardware, they can be various electronic devices with displays, including but not limited to smartphones, tablets, laptops, and desktop computers. When terminal devices 101, 102, and 103 are software, they can be installed in the aforementioned electronic devices, and can be implemented as multiple software programs or software modules, or as a single software program or software module; no specific limitation is made here. When server 105 is hardware, it can be implemented as a distributed server cluster composed of multiple servers, or as a single server. When server 105 is software, it can be implemented as multiple software programs or software modules, or as a single software program or software module; no specific limitation is made here.

[0019] Server 105 can provide various services through its built-in applications. Taking a dialogue-based question-and-answer application that provides dialogue services to users as an example, server 105 can achieve the following effects when running this application: First, it receives instant requests from users using terminal devices 101, 102, and 103 via network 104, and processes the instant requests using a preset fast inference system to obtain an instant response. The fast inference system generates the instant response for the instant request based on content indexing technology. Next, when the matching degree between the relevant knowledge used to generate the instant response and the instant request is lower than the preset matching degree, the server processes the instant request using a preset slow inference system to obtain a request reply. The relevant knowledge is determined based on information related to the instant request extracted from a preset shared information database using content indexing technology. The slow inference system generates the request reply for the instant request based on task decomposition and multi-intelligence collaboration. Finally, the new knowledge contained in the request reply is added to the shared information database.

[0020] It should be noted that, in addition to being obtained from terminal devices 101, 102, and 103 via network 104, immediate requests can also be pre-stored locally on server 105 through various means. Therefore, when server 105 detects that this data is already stored locally (e.g., when it begins processing previously retained pending requests), it can choose to retrieve this data directly from its local storage. In this case, the exemplary system architecture 100 may also exclude terminal devices 101, 102, and 103 and network 104.

[0021] Since processing user requests requires significant computing resources and power, the request processing methods based on the fast-slow dual-inference system provided in the subsequent embodiments of this disclosure are generally executed by a server 105 with strong computing power and abundant computing resources. Correspondingly, the request processing device based on the fast-slow dual-inference system is also generally located in the server 105. However, it should also be noted that when terminal devices 101, 102, and 103 also possess sufficient computing power and resources, they can also complete the aforementioned calculations performed by the server 105 through their installed dialog-based question-and-answer applications, thereby outputting the same results as the server 105. Especially when multiple terminal devices with different computing capabilities exist simultaneously, but the dialog-based question-and-answer application determines that its terminal device has strong computing power and abundant remaining computing resources, it can allow the terminal device to perform the aforementioned calculations, thereby appropriately reducing the computing pressure on the server 105. Accordingly, the request processing device based on the fast-slow dual-inference system can also be located in the terminal devices 101, 102, and 103. In this case, the exemplary system architecture 100 may also exclude the server 105 and the network 104.

[0022] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0023] Please refer to Figure 2 , Figure 2 A flowchart of a request processing method based on a fast and slow dual inference system provided in this disclosure embodiment, wherein process 200 includes the following steps: Step 201: Use the preset fast inference system to process the immediate request initiated by the first user and obtain an immediate response; This step is intended for the execution body of the request processing method based on the fast and slow dual inference system (e.g., Figure 1 The server 105 shown uses a pre-set fast inference system to process the immediate request initiated by the first user and obtain an immediate response. The purpose of this step is to quickly meet the query needs of ordinary users (i.e., the first user) who have high requirements for immediacy (such as "What is the weather like in Beijing today?" "Where is my package?" etc.) through a lightweight and high-response-speed inference mechanism. The fast inference system, as the core module for processing immediate requests, has the core advantage of speed, that is, it needs to return results within 1-2 seconds. The key technology that supports this speed is the content indexing technology used.

[0024] From a technical perspective, this "content indexing technology" involves the aforementioned entity pre-converting the massive amounts of information required for the dialogue Q&A application (such as weather data, express delivery tracking information, and frequently asked questions) into a structured data structure that can be quickly retrieved, similar to a library's "catalog system." On one hand, the entity can construct an indexing mechanism similar to an "inverted index," associating keywords in the information (such as "Beijing," "weather," and "express delivery tracking number") with corresponding content. For example, the keyword "Beijing" can be associated with all documents containing "Beijing weather." On the other hand, to handle more complex semantic queries (such as "What clothes should I wear today?"), the entity can also construct a "vector index." For instance, it can convert text content into high-dimensional vectors (numerical representations) using a large language model and utilize the Approximate Nearest Neighbor (ANN) algorithm to quickly find content semantically most similar to the user's request. Furthermore, by combining these two types of indexes, information relevant to the immediate request can be quickly extracted from massive amounts of data.

[0025] In practical terms, when the first user sends an immediate request (such as "When will my express tracking number 12345 arrive?") through the dialogue Q&A application, the aforementioned execution entity can first parse the request into two parts: one is keywords ("express tracking number 12345" "arrival time"), and the other is a semantic vector (representing "querying the arrival time of a specific express delivery"). Then, the fast inference system can call content indexing technology: first, it quickly locates all content containing "express tracking number 12345" (such as express tracking records) through an inverted index, and then calculates the similarity between these contents and the request semantic vector through vector indexing (such as the vector of "the latest status of express 12345 is 'arrived at the Beijing transit station, expected to be delivered tomorrow'" has the highest similarity to the request vector), thereby extracting the most relevant "related knowledge" (i.e., information directly related to the request). Afterwards, the fast inference system can organize this related knowledge into a natural language response (such as "Your express tracking number 12345 has arrived at the Beijing transit station and is expected to be delivered before 10:00 AM tomorrow"), and return it to the first user through the dialogue Q&A application. Throughout the process, content indexing technology, through its "pre-built" feature (the server updates the index periodically, such as obtaining the latest tracking data from the courier service provider's interface every hour to update the inverted index and vector index), ensures that the fast inference system does not need to process all data in real time, but directly extracts the latest and most relevant information from the index, thereby achieving instant response.

[0026] Of course, besides directly using the indexed relevant information as the immediate response, natural language processing techniques can also be used to organize and generate the immediate response based on the indexed related knowledge, such as using a large language model. Furthermore, the indexed relevant information can undergo some sorting or filtering before being used as related knowledge by the large language model to improve the accuracy of the related knowledge.

[0027] Furthermore, the content indexing techniques mentioned above can be further optimized to ensure the fast inference system maintains high speed. For example, the execution entity can cache the results of frequently queried queries (such as "today's weather"). When multiple users repeatedly query the same query, the cached results are returned directly without needing to call the index again. Simultaneously, the execution entity can periodically clean up expired data in the index (such as yesterday's weather information) to ensure the index remains lightweight and avoids impacting retrieval speed due to excessive data volume. These optimization measures work together to enable the fast inference system to maintain high responsiveness while providing accurate and up-to-date information when processing immediate requests.

[0028] Step 202: In response to the fact that the matching degree between the associated knowledge used to generate the immediate response and the immediate request is lower than the preset matching degree, the immediate request is processed using the preset slow inference system to obtain the request response; Building upon step 201, this step aims to have the aforementioned executing entity process the immediate request using a preset slow inference system when it discovers that the matching degree between the associated knowledge used to generate the immediate response and the immediate request is lower than a preset matching degree, thereby obtaining a request response. Specifically, when the matching degree between the associated knowledge used to generate the immediate response and the immediate request initiated by the first user is lower than a preset matching degree, the aforementioned executing entity will activate the preset slow inference system to process the request again, in order to generate a more accurate and comprehensive request response. The core logic of this step is that the fast inference system relies on associated knowledge in a shared information database to achieve an immediate response, but when the associated knowledge is insufficient to match the user's needs, a deeper inference mechanism is needed to fill the information gap.

[0029] Among them, "related knowledge" refers to the information related to the immediate request extracted by the aforementioned executing entity from a pre-set shared information database. Its source is content indexing technology, that is, the shared information database serves as a "knowledge base" storing various structured and unstructured knowledge (such as historical dialogue records, industry data, frequently asked questions, etc.), and content indexing technology (such as inverted index, vector index) is like a "library catalog," which can quickly locate knowledge related to the semantics or keywords of the immediate request (for example, if a user asks "the distribution package has decreased today," the related knowledge may be "yesterday's distribution package data" or "common reasons for distribution package anomalies" in the shared information database); "matching degree" is a similarity measure between the related knowledge and the immediate request (such as semantic similarity, keyword coverage). The pre-set matching degree can be a pre-set threshold (such as 0.8 or 80%). If it is lower than this value, it means that the related knowledge cannot meet the user's needs for "accuracy" or "depth" (for example, "yesterday's distribution package data" cannot explain the specific reason for "the decrease in today's distribution package").

[0030] When the matching degree is low, the aforementioned executing agents will attempt to switch to a "slow inference system" to perform a more comprehensive and complex analysis and reasoning on the immediate request, which is also relatively time-consuming, in order to obtain a more accurate and comprehensive answer. The core mechanism of this slow inference system is task decomposition and multi-agent collaboration. The task decomposition technique breaks down complex immediate requests into multiple processable sub-tasks (reducing problem complexity). The latter is completed by multiple intelligent agents (leveraging the expertise of each intelligent agent). For example, when a user requests "the reasons for the decline in Beijing distribution packages today and solutions", the aforementioned executing agents can first determine that the matching degree between the related knowledge (such as "yesterday's distribution package data") and the request is less than 80%, and then activate the slow inference system to decompose the request into three sub-tasks: "obtain today's Beijing distribution package data", "analyze the reasons for the decline", and "generate targeted solutions". This is precisely because complex problems often involve knowledge from multiple domains (such as data acquisition, logical analysis, and strategy generation). Therefore, breaking them down into multiple sub-tasks reduces the processing difficulty of each step, thereby improving the accuracy of the final answer. The "multi-intelligence collaboration" technology involves the slow inference system assigning the decomposed sub-tasks to matching agents (such as data acquisition agents, cause analysis agents, and solution generation agents). Each agent is responsible for completing one sub-task. For example, the data acquisition agent obtains specific data on today's Beijing distribution packages from a shared information database or external interface (such as the distribution package system database) (e.g., "Today's distribution package volume is 1000 orders, a decrease of 20% compared to yesterday"). The cause analysis agent analyzes the reasons for the decrease by combining historical data and system change records (e.g., "yesterday's server upgrade") ("The server upgrade caused system delays, affecting distribution package efficiency"). The solution generation agent proposes suggestions such as "roll back the server upgrade" and "optimize the upgrade process" based on the reasons. At the same time, the agents can exchange information through internal communication mechanisms (such as message queues and API interfaces) (e.g., the data agent transmits the acquired data to the analysis agent) to collaboratively complete all sub-tasks. This process fully follows the technical principle that division of labor and cooperation can leverage the expertise of each agent (e.g., data agents are good at quickly acquiring structured data, while analytical agents are good at logical reasoning) and avoid the inefficiency or errors of a single agent when dealing with complex problems.

[0031] Finally, the slow inference system integrates the results of the sub-tasks completed by each agent to generate a complete response to the request (e.g., "Today, the number of packages distributed in Beijing has decreased by 20% due to system delays caused by yesterday's server upgrade. It is recommended to immediately roll back the upgrade and conduct stress testing before any subsequent upgrades"). This response not only addresses the user's core question ("Why is it decreasing?") but also provides an actionable solution ("What should be done?"), far exceeding the depth that the fast inference system can achieve by relying on associative knowledge.

[0032] From a practical perspective, the value of slow inference systems lies in balancing "immediacy" and "accuracy": fast inference systems handle simple, highly relevant requests (ensuring immediacy), while slow inference systems handle complex, less relevant requests (ensuring accuracy). The "task decomposition and multi-agent collaboration" mechanism ensures that slow inference systems can efficiently handle complex problems. By decomposing tasks and dividing labor, it improves both processing efficiency (avoiding excessive time spent by a single agent) and response quality (leveraging the strengths of each agent). For example, in an e-commerce customer service scenario, when a user asks, "Why hasn't my package arrived yet? Is it lost?", if the fast reasoning system's extraction of relevant knowledge from the shared information database (such as "the package has been sent") has a low matching degree (unable to explain the reason for "not arriving"), the slow reasoning system will break it down into three sub-tasks: "obtaining the latest trajectory of the package," "analyzing the reasons for the delay," and "generating a reassurance plan." These are handled by the trajectory query agent, the reason analysis agent, and the customer service response agent, respectively. Ultimately, it generates a response such as, "Your package has arrived at the Beijing transit station. Due to the blizzard, it is delayed and is expected to arrive tomorrow. We will provide you with a 5 yuan no-threshold coupon as compensation," which not only resolves the user's question but also improves the user experience.

[0033] Step 203: Add the new knowledge contained in the request response to the shared information repository.

[0034] Building upon step 202, this step aims to have the aforementioned executing entity supplement the shared information repository with the new knowledge contained in the request response. This step is the core mechanism for achieving "continuous evolution," that is, using the deep processing results of slow reasoning to feed back into the knowledge source of fast reasoning, so that subsequent similar requests can obtain more accurate, comprehensive, and timely responses.

[0035] Since the shared information repository serves as the "knowledge base" for the fast inference system to process immediate requests, storing related knowledge such as historical dialogue records, frequently asked questions, and industry data, fast inference quickly matches user requests using library-like cataloging techniques (such as inverted indexes and vector indexes). When slow inference processes requests with low matching rates, the response usually includes "new knowledge" not found in the shared repository. For example, if a user asks "Why is the distribution package rate dropping today?", slow inference might reply "It's because yesterday's server upgrade caused system latency." The causal relationship of "server upgrade → distribution package rate drop" is the new knowledge. Adding this knowledge to the shared repository expands the fast inference's "knowledge base." When other users ask similar questions later (such as "The distribution package rate in Shanghai has dropped today"), fast inference can directly extract more comprehensive related knowledge from the updated shared repository (such as "Server upgrades may cause a drop in distribution packages"), generating a more accurate immediate response (such as "The drop in distribution packages may be related to recent server upgrades; we recommend checking system latency").

[0036] The essence of this process is the "data flywheel" effect: slow inference generates new knowledge → shared library updates → fast inference response quality improves → user experience improves → more users use it → more slow inference requests → more new knowledge generation, forming a positive cycle. In practical applications, supplementing new knowledge requires four steps: "extraction—verification—structuring—storage." First, structured knowledge is extracted from the response (e.g., "Problem: Distribution package decline; Reason: Server upgrade; Solution: Rollback upgrade"). Then, the accuracy of the knowledge is ensured through timestamp verification, external interface verification, etc. Next, it is converted into a machine-processable format (e.g., key-value pairs) and metadata (e.g., generation time, keywords) is added. Finally, it is stored in the shared library using content indexing technology. To optimize efficiency, the above-mentioned execution entities can filter meaningless content (e.g., "thank you"), cache high-frequency knowledge (e.g., "reason for distribution package decline"), and update the index regularly (e.g., hourly) to ensure that new knowledge can be obtained by fast inference in a timely manner.

[0037] The request processing method based on a fast and slow dual-inference system provided in this disclosure achieves a balance between request processing timeliness and response quality through the collaborative architecture of the fast and slow dual-inference systems. The fast inference system responds to user requests instantly based on content indexing technology, ensuring immediacy. Simultaneously, by judging the matching degree between related knowledge and requests, requests with low matching degrees are further transferred to the slow inference system, rationally allocating computing resources while avoiding unnecessary deep inference overhead. The slow inference system, employing task decomposition and multi-intelligence collaboration, can perform in-depth processing of complex requests, effectively improving the accuracy and depth of responses compared to the fast inference system. Furthermore, by supplementing the shared information base with the new knowledge generated by slow inference, the fast inference system can gradually improve the quality and accuracy of its rapid responses in subsequent response generation processes due to the continuous enrichment of related knowledge. This enables the continuous evolution of the request processing system, and in the long run, it improves the response quality of fast inference, reduces the frequency of slow inference calls, and optimizes the overall efficiency and effectiveness of request processing.

[0038] To gain a deeper understanding of how fast inference systems specifically handle immediate requests, please refer to [link / reference needed]. Figure 3 , Figure 3 A flowchart of a method for processing immediate requests using a fast inference system, provided as an embodiment of this disclosure, is included in process 300, which comprises the following steps: Step 301: Utilize the recall mechanism that constitutes the fast reasoning system to recall the candidate knowledge set related to the immediate request from the shared information database through the approximate nearest neighbor search algorithm; The recall mechanism serves as the "information entry point" of the fast reasoning system. Its function is to quickly locate candidate knowledge relevant to the immediate request from a shared information repository (which stores massive amounts of knowledge such as historical dialogues, frequently asked questions, and real-time data). To achieve "speed" (typically requiring completion within 100 milliseconds), the server uses the ANN algorithm to quickly find candidate knowledge (rather than absolutely precise results) that is semantically similar to the request through vector transformation and indexing techniques. For example, the server might convert a user's request "What's the weather like in Beijing today?" into a high-dimensional vector (through the embedding layer of a large language model), and then use the ANN algorithm to recall 100 candidate knowledge entries semantically similar to that vector from the vector index in the shared information repository (such as "Yesterday's weather in Beijing was sunny," "Beijing weather forecast for the next three days," "Clothing suggestions for Beijing today," etc.). The technical principle behind this step is: using vectors to represent the semantic information of the text, and using ANN to quickly calculate vector similarity to achieve rapid retrieval of massive amounts of data.

[0039] Step 302: Use the sorting mechanism that constitutes the fast reasoning system to sort each candidate knowledge in the candidate knowledge set according to the preset sorting algorithm to obtain the sorted candidate knowledge set; Considering that not all of the retrieved candidate knowledge is highly relevant to the user's request (for example, "Yesterday's weather in Beijing was sunny" might not answer the question about "today"), a ranking mechanism is needed to filter out the most relevant candidates. The ranking algorithm used by this mechanism can score and rank candidate knowledge based on multiple features (such as keyword coverage, timeliness, and user historical interaction data). For example, traditional information retrieval algorithms can be used to calculate the frequency and distribution of keywords such as "today," "Beijing," and "weather" in candidate knowledge, and then score 100 candidate knowledge items, ranking "Beijing weather forecast for the next three days" (containing keywords such as "today," "Beijing," and "weather") first, and "Yesterday's weather in Beijing was sunny" (not containing "today") 20th. For more precise ranking, the aforementioned execution entity may also use machine learning ranking algorithms (such as XGBoost), combining features such as timeliness (e.g., "forecast for the next three days" is more relevant than "yesterday's weather") and user click history (e.g., users view "forecast for the next three days" more frequently) to improve ranking accuracy. The purpose of this step is to sort the candidate knowledge and place the Top N (e.g., Top 10) most relevant to the user's request at the forefront, laying the foundation for subsequent steps.

[0040] Step 303: Use the intervention and adjustment mechanism that constitutes the fast reasoning system to filter relevant knowledge from the sorted candidate knowledge set; Even after ranking, the candidate knowledge set (e.g., Top 10) may still contain biases (e.g., the algorithm may not have considered the priority of "urgent information"). Therefore, an intervention and adjustment mechanism is needed for manual or rule-based correction. The intervention and adjustment mechanism works by filtering the final related knowledge from the ranked candidate knowledge set according to preset rules or manual instructions. For example, the server might pre-set a rule: "When candidate knowledge contains the 'urgent' tag, place it at the top." If the 5th ranked candidate knowledge is "Urgent: Yellow Rainstorm Warning for Beijing Today" (containing the "urgent" tag), the intervention and adjustment mechanism will move it to the top position, making it the final related knowledge. Furthermore, operators can also intervene in real-time through the backend system (e.g., manually adding "Sudden Temperature Drop in Beijing Today" to the related knowledge) to ensure that the related knowledge meets the user's actual needs. The technical principle behind this step is that the algorithm's ranking may overlook certain scenario-based needs (e.g., emergency situations). The intervention and adjustment mechanism can compensate for the algorithm's limitations and improve the accuracy of the related knowledge.

[0041] Step 304: Utilize the large language model that constitutes the fast reasoning system to generate immediate responses for immediate requests based on relational knowledge.

[0042] Relevant knowledge (such as "Beijing weather forecast for the next three days: Today, sunny, temperature 15-22℃; Tomorrow, cloudy, temperature 13-20℃") is structured or semi-structured information, therefore it needs to be converted into a natural, conversational immediate response (such as "Today, Beijing's weather is sunny and the temperature is suitable for outdoor activities; tomorrow, cloudy, remember to bring an umbrella"). The core of this step is that the large language model in the fast reasoning system generates a response that meets the user's needs based on the input relevant knowledge and prompt (such as prompt: "Convert the following weather forecast into a conversational response suitable for ordinary users: Beijing weather forecast for the next three days: Today, sunny, temperature 15-22℃; Tomorrow, cloudy, temperature 13-20℃"), and returns a natural, coherent immediate response to the first user.

[0043] Furthermore, the number of candidate knowledge items recalled can be dynamically adjusted based on the size of the shared information repository (e.g., recalling 100 items when the shared repository has 1 million data items; recalling 200 items when there are 100 million data items) to ensure a balance between comprehensiveness and speed in recall; features of the ranking algorithm can be updated regularly (e.g., adding the "user's geographical location" feature, so that when the user is in Shanghai, "Shanghai weather" has a higher ranking weight than "Beijing weather") to improve the personalization of the ranking; intervention rules can be adjusted based on user feedback (e.g., if a user complains "no emergency warning received", the intervention priority of the "emergency" tag is strengthened) to improve the relevance of related knowledge; and the resource consumption of the lightweight LLM used by the fast inference model can be further reduced and the inference speed improved through model compression (e.g., quantization, pruning).

[0044] This embodiment forms a closed loop of "rapidly extracting effective information from massive amounts of knowledge and converting it into a natural language response" through the solutions provided in steps 301-304: that is, the recall mechanism uses ANN to quickly locate candidate knowledge, the ranking mechanism selects the most relevant candidates, the intervention and adjustment mechanism corrects deviations, and the lightweight LLM generates conversational responses. Each step revolves around the balance between speed and accuracy, ensuring that the first user can receive an immediate response that meets their needs within 1-2 seconds. While improving the user experience, it also indirectly saves computing resources for the slow inference system and realizes the rational allocation of system resources.

[0045] To deepen your understanding of how to specifically handle low-match immediate requests using a slow inference system, please also refer to [link to relevant documentation]. Figure 4 , Figure 4 A flowchart of a method for processing immediate requests using a slow inference system, provided as an embodiment of this disclosure, is included in process 400, comprising the following steps: Step 401: The main agent constituting the slow inference system determines the complete task based on the immediate request and the shared information base, and then breaks down the complete task into multiple sub-tasks and distributes them to the matching sub-agents. This step aims to enable the aforementioned executing entity to utilize the main intelligent agent constituting the slow inference system to confirm the complete task, break down subtasks, and match and distribute them. That is, the main intelligent agent is the "brain" of the slow inference system, and its responsibility is to combine the immediate request with existing knowledge in the shared information base to identify the user's true needs (the complete task) and break it down into executable subtasks. For example, when a user asks, "The number of distribution packages in Beijing has decreased today," the shared information base only stores basic data such as "1000 distribution packages yesterday" and "800 distribution packages today," without including related knowledge such as "reasons for the decrease" or "solutions." The main intelligent agent can then determine that the user's true need is not just "to know that the number has decreased," but "to understand the reasons and solve the problem." Therefore, the complete task is defined as "to analyze the reasons for the decrease in distribution packages in Beijing today and propose targeted solutions."

[0046] The essence of subtask decomposition is to modularize complex tasks. The main agent, based on the "complete task" and "existing knowledge gaps," breaks it down into multiple subtasks with clearly defined dependencies and responsibilities. Taking the example above, the main agent decomposes the complete task into three subtasks: 1) Obtain detailed data on today's Beijing distribution packets (such as the number of packets distributed in each time period, system latency logs, and third-party interface status); 2) Analyze the specific reasons for the decline (such as whether it's due to server upgrades causing system latency, third-party interface failures, or business process anomalies); 3) Generate targeted solutions (such as rolling back server upgrades, fixing interface vulnerabilities, or adjusting business strategies). Each subtask corresponds to a matching sub-agent (such as a data acquisition agent, a cause analysis agent, and a solution generation agent). The principle is that different subtasks require expertise in different domains (data acquisition requires interfacing with external systems, cause analysis requires logical reasoning, and solution generation requires business experience). Decomposition leverages the strengths of each agent, improving processing accuracy.

[0047] Step 402: Using the main agent constituting the slow inference system, control each sub-agent to execute the corresponding sub-task in an orderly manner according to the execution order of each sub-task, and obtain the request response.

[0048] After breaking down tasks into subtasks, the main agent transforms into a "coordinator," controlling each sub-agent to complete its task in an orderly manner based on the dependencies and execution logic of the subtasks. For example, "acquiring detailed data" is a prerequisite for "analyzing the cause" (without data, it's impossible to deduce the cause), and "analyzing the cause" is the foundation for "generating a solution" (without the cause, it's impossible to propose an effective solution). Therefore, the main agent must distribute tasks in the order of "data acquisition → cause analysis → solution generation." The core technical principle of this step is that there are strong dependencies between subtasks, and orderly execution avoids logical confusion (such as analyzing the cause without acquiring data), while ensuring "contextual consistency" between sub-agents (e.g., the information "the system was delayed by 10 seconds at 14:00 today," acquired by the data agent, needs to be accurately transmitted to the cause analysis agent).

[0049] In practice, the main agent can use the Model Context Protocol (MCP) to achieve information transmission, state synchronization, and invocation of external tools. For example, after the data acquisition agent completes the subtask of "obtaining detailed data of today's Beijing distribution packages," it returns the data to the main agent in a protocol format (such as structured information including "task ID, data type, timestamp, and specific value"). The main agent then uses this data as "context" and sends it to the cause analysis agent, specifying the task requirements ("analyze the cause of the decline based on today's data and historical logs"). The cause analysis agent combines historical cases in the shared information database (such as "the server upgrade in October 2025 caused a decline in distribution packages") with the current data to deduce that "today's decline is caused by the system delay due to the server upgrade at 20:00 yesterday," and then passes the result to the solution generation agent. The solution generation agent then generates a solution based on the cause: "roll back the server upgrade to the previous version and conduct stress testing before subsequent upgrades." Throughout the process, the main agent constantly monitors the execution status of each sub-agent (such as "data acquisition in progress" or "cause analysis completed"). If a sub-task fails (such as data acquisition timeout), a retry mechanism will be triggered or the task order will be adjusted (such as acquiring data from other channels) to ensure that the process is not interrupted.

[0050] Furthermore, the "orderly control" of the master agent is not merely about "delivering tasks sequentially," but can also include dynamic adjustment and adaptive processes. For example, if the cause analysis agent discovers that "the decline is not only due to server upgrades but also to third-party interface failures," the master agent can temporarily add a subtask of "verifying the status of the third-party interface" and send it to the data acquisition agent to supplement the data. If the "rollback upgrade" solution proposed by the solution generation agent conflicts with cases of "historical rollback leading to data loss" in the shared information database, the master agent will require the solution generation agent to adjust the solution (e.g., "first restore the interface, then gradually roll back the upgrade"). In other words, slow inference systems need to handle complex "unstructured" problems, and the master agent, as the central hub, needs to adjust the task flow based on real-time feedback to ensure the accuracy and feasibility of the results.

[0051] This embodiment describes how the main intelligent agent, through steps 401-402, transforms complex requests into a collaboratively achievable task chain using a "decomposition-control" mechanism: First, the main intelligent agent identifies the user's true needs (the complete task), reduces the problem's complexity by decomposing it; then, through orderly control and context synchronization, it allows each sub-agent to leverage its expertise, ultimately generating a deep response containing "cause-solution." The core value lies in the fact that the slow inference system does not rely on "existing knowledge," but rather creates "new knowledge" through "active decomposition and collaboration." This new knowledge (such as the causal relationship of "server upgrade + interface failure → reduced distribution packets") is added to the shared information base, feeding back into the fast inference system and enabling subsequent similar requests to receive better, more immediate responses.

[0052] Building upon the above embodiments, when a first user's immediate request needs to trigger slow inference due to low matching degree of related knowledge, the aforementioned execution entity may not submit the immediate request to the slow inference system. Instead, it can first store the request in a preset message queue. This message queue serves as the entry point for the slow inference system, buffering high-concurrency requests, optimizing resource allocation, and preventing the slow inference system from crashing due to overload. In other words, before using the slow inference system to process immediate requests, immediate requests with a matching degree lower than a preset matching degree of related knowledge can be stored in a preset message queue. This allows the slow inference system to process only the requests retrieved from the message queue and obtain a response.

[0053] Furthermore, the request processing priority of immediate requests in the message queue can be determined. This request processing priority can be determined based on at least one of the following: the problem domain of the immediate request, the proportion of relevant knowledge in the shared information base, and the importance evaluation. Then, when the request processing priority is lower than the preset priority, the corresponding immediate request is deleted from the message queue; and when the request processing priority is not lower than the preset priority, the corresponding immediate request is retained in the message queue.

[0054] In other words, requests in the message queue are not processed in a first-come, first-served manner, but are sorted by priority. An empty server can calculate the request processing priority of each request based on three dimensions: the problem domain, the proportion of relevant knowledge in the shared information base, and the importance evaluation, so as to ensure that limited slow inference resources (such as GPU and memory) are used on the most valuable requests. Regarding the question domain dimension, this takes into account the different values ​​of requests in different domains (e.g., the medical question "Why isn't my medicine working?" is more important than the entertainment question "What's fun to do today?"). The aforementioned execution entities can preset weights for each domain (e.g., medical weight 10, e-commerce weight 8, entertainment weight 5). The higher the domain weight, the higher the priority. As for the relevant knowledge percentage dimension, if the knowledge percentage of a certain domain in the shared information library is low (e.g., the knowledge of "reasons for the decline in distribution packages" only accounts for 1% of the shared library), it indicates that the domain has a "knowledge gap." Processing such requests can supplement the knowledge in the shared library (e.g., generating new knowledge of "reasons for the decline in distribution packages"), so the priority is high. For example, if a user asks "The distribution packages in Shanghai have decreased today," and the knowledge percentage of "Shanghai distribution packages" in the shared library is only 0.5%, the priority will be increased. For the importance evaluation dimension, the user's "subjective importance" is also key (e.g., the user adds keywords such as "urgent!" or "important" to the request, or marks it as "urgent"). These signals can be identified through natural language processing (NLP) to improve the priority.

[0055] After calculating the request priority, the executing entity can compare the actual request priority with the preset priority. If the actual request priority is lower than the preset priority, it means that the "value" of the request is insufficient to consume slow inference resources (e.g., the entertainment question "What's fun today?", with a domain weight of 5, a high proportion of relevant knowledge (there is a lot of entertainment information in the shared library), no importance marker, priority 5, below the threshold of 6). The executing entity can delete it from the message queue without slow inference processing. However, if the actual request priority is not lower than the preset priority, it means that the request is worth processing (e.g., the medical question "Why isn't my medicine working?", with a domain weight of 10, a low proportion of relevant knowledge (the proportion of "drug effect" knowledge in the shared library is 2%), a "urgent!" marker, priority 10, above the threshold of 6). The executing entity can keep it in the message queue, waiting for the slow inference system to process it.

[0056] Furthermore, resource allocation can be optimized by dynamically adjusting priorities. For example, during specific periods (such as promotional festivals), the priority weight of requests in specific areas can be temporarily increased (from 8 to 10). The priority of requests from high-level or high-privilege users can also be increased to ensure that the needs of high-value users are met first. The length of the message queue can also be monitored in real time (e.g., if there are 1000 requests in the queue). If the queue is too long, the preset priority can be temporarily increased (e.g., from 6 to 7) to filter out more low-value requests and avoid queue overflow.

[0057] The message queue and priority filtering scheme provided in this embodiment is essentially a rational scheduling of the slow inference system. It buffers high-concurrency requests through message queues to prevent slow inference from becoming overloaded, while prioritizing requests to identify high-value requests, ensuring resources are used where they are most needed. In other words, it improves system efficiency by filtering out low-value requests. The core value lies in allowing the slow inference system to focus on processing high-value, high-demand requests, avoiding resource waste, and ensuring system stability and throughput through message queue buffering.

[0058] To further understand how the main intelligent agent specifically determines the complete task corresponding to an immediate request, please refer to [link to relevant documentation]. Figure 5 , Figure 5 A flowchart of a method for determining the complete task corresponding to an immediate request using an agent constituting a slow inference system, provided as an embodiment of this disclosure, is included in the following steps: Step 501: Use the main agent to determine the completeness of the actual intent expressed in the immediate request; The intent completeness described in this step is an indicator that measures whether a user request contains enough key information. For example, "The distribution package price in Beijing has dropped today" contains three key elements: "time (today), region (Beijing), and event (distribution package price drop)," thus the intent is complete. However, "The distribution package price has dropped today" lacks "region," making the intent incomplete. The main agent's responsibility is to extract the key elements from the user request using natural language processing technology and calculate the actual intent completeness.

[0059] At the technical level, the main intelligent agent typically employs an intent recognition model (such as a BERT-based sequence labeling model) to extract key elements from user requests, such as "entities (region, time, product), actions (decline, query, solution), and needs (reason, solution)." The number of extracted elements is then compared to the number of elements required for the expected complete intent (e.g., "time + region + event + need"), calculating a completeness score (e.g., if two elements are extracted but four are expected, the completeness score is 0.5). For example, if a user requests "The distribution package has decreased today," the model extracts two elements: "time (today)" and "event (distribution package decreased)." The expected complete intent requires four elements: "time + region + event + need," therefore the actual intent completeness score is 0.5 (2 / 4).

[0060] In practical terms, the aforementioned execution entities can pre-set key element templates for different types of requests (e.g., the template for "distribution package issue" is "time + region + distribution package quantity + requirement"). The main intelligent agent can also quickly determine the completeness by matching the elements in the template. For example, the "distribution package issue" template requires "time (today), region (Beijing), distribution package quantity (decreased), requirement (reason)". The user request "distribution packages decreased today" is missing "region" and "requirement", and the completeness is 0.5 (2 / 4).

[0061] Step 502: In response to the fact that the actual intent completeness is lower than the preset intent completeness threshold, the main agent performs intent completion operation on the instant request based on the shared information database to obtain the complete intent; When the actual intent completeness is lower than a preset intent completeness threshold (e.g., 80%), the main agent can extract contextual information from the shared information base to attempt to fill in the missing key elements, thereby obtaining a "complete intent (an intent with a completeness not lower than the preset intent completeness threshold)". The context in the shared information base includes user history dialogues, frequently asked questions (FAQs), industry common sense data, etc. The principle is that user requests are often related to historical behavior or common scenarios, and the data in the shared information base can infer the missing elements.

[0062] For example, a user's actual intent to request "Distribution packages have decreased today" has a completeness of 50% (missing "region" and "need"), below the threshold of 80%. The agent can first query the user's historical dialogues in the shared information database and find that the user asked "What is the distribution package volume in Beijing?" yesterday, thus inferring that "region" is Beijing; then, it queries frequently asked questions and finds that a common need for "distribution package decrease" is "analyze the reason", so it completes the "need" to "analyze the reason". Finally, the complete intent is "Distribution packages in Beijing have decreased today, I want to know the reason".

[0063] In practical terms, the completion logic of the main intelligent agent can be specifically "probability-first". For example, if the region of "user's historical dialogue" in the shared information database is Beijing (probability 70%) and the region of "frequently asked questions" is Shanghai (probability 30%), then the priority is to complete it as Beijing.

[0064] Step 503: Use the main agent to determine the complete task corresponding to the complete intent.

[0065] The complete intent is the "user's real need," while the complete task is the "specific work that needs to be done to satisfy that need." Therefore, the main agent's responsibility is to map the complete intent to a preset task template and, in conjunction with the existing knowledge in the shared information base, determine the complete task that needs to be executed.

[0066] At the technical principle level, the main intelligent agent can use a task matching model (such as a rule-based or machine learning-based classification model) to match the complete intent with a preset task template (such as "analyze the cause", "generate a solution", "provide data"). For example, the complete intent "The distribution package in Beijing dropped today, and I want to know the reason" corresponds to the task template "analyze the cause"; if the complete intent is "The distribution package in Beijing dropped today, and I want to solve the problem", then it corresponds to the task template "analyze the cause + generate a solution".

[0067] At the practical level, existing knowledge in the shared information base influences the determination of the complete task. For example, if the shared information base already contains "common reasons for the decline in distribution packages" (such as server upgrades or third-party interface failures), the main agent will determine the complete task as "analyze the reasons for the decline in distribution packages in Beijing today (based on existing knowledge)." If the shared information base does not contain relevant knowledge, the main agent may expand the complete task to "obtain detailed data on distribution packages in Beijing today → analyze the reasons → generate a solution" (but according to user requirements, task decomposition is not involved here; only the complete task is determined). For example, if the complete intent is "Distribution packages in Beijing have declined today, and I want to know the reason," the main agent, combining the "common reasons for the decline in distribution packages" in the shared information base, determines the complete task as "analyze the reasons for the decline in distribution packages in Beijing today (including server status, third-party interfaces, business processes, etc.)."

[0068] This embodiment, through the process of the main intelligent agent processing intent provided in steps 501-503, essentially transforms the user's "fuzzy needs" into "explicit tasks." Specifically, it identifies missing elements in the user's request by judging intent completeness, completes the request using a shared information base, transforming the incomplete intent into a complete one, and maps the complete intent to an executable task through task matching. This process ensures that subsequent steps of the slow inference system (splitting subtasks and executing subtasks) have a clear direction, improving the accuracy and efficiency of slow inference.

[0069] exist Figure 5 Based on the illustrated embodiment, a more specific implementation of step 502 can be as follows: First, the main agent generates a supplementary request to complete the intent expressed by the immediate request. Then, the main agent searches for supplementary knowledge corresponding to the supplementary request in the shared information database. Finally, the main agent completes the current intent expressed by the immediate request based on the supplementary knowledge to obtain the complete intent.

[0070] The "supplementation request" is a query generated by the main agent based on the missing elements of the current intent, with the purpose of clarifying the information that needs to be supplemented. After generating the supplementation request, the main agent first searches for supplementary knowledge corresponding to the supplementation request in the shared information base, because the shared information base stores data such as user history dialogues, frequently asked questions (FAQs), and industry common sense, that is, it attempts to "use past knowledge to solve the current problem".

[0071] Furthermore, if it is still impossible to complete the current intent into a complete intent by supplementing knowledge, the main agent can initiate a supplementary query to the first user, obtain a query response, and then use the main agent to complete the current intent based on the query response until a complete intent is obtained.

[0072] If the shared information base lacks supplementary knowledge corresponding to the supplementary request (e.g., the user is a new user, there is no history of conversations, and the default region is not listed in the FAQs), the main agent will initiate a supplementary query to the first user. This is "proactively requesting information from the user" to ensure the completeness of the intent. For example, if the supplementary request is "the region where the distribution package has decreased," and there is no relevant data in the shared information base, the main agent will send a message to the user through the dialogue question-and-answer function: "Which region are you referring to when you say the distribution package has decreased?" After the user replies "Shanghai," the main agent will use "Shanghai" as supplementary knowledge to complete the "region" element of the current intent, resulting in the complete intent: "The distribution package in Shanghai has decreased today."

[0073] Furthermore, the main AI can record the process of intent completion (such as "completing the region as Beijing through user history dialogue" or "completing the region as Shanghai through user inquiry"), so that the source of the intent can be traced during subsequent task decomposition, thereby improving the accuracy of task execution.

[0074] The overall approach of the solution provided in this embodiment is as follows: first, search for existing knowledge, then ask the user. That is, the main intelligent agent generates a supplementary request, searches for supplementary knowledge in the shared information base, and tries to complete the intent with existing knowledge. If no existing knowledge is available, it then initiates a supplementary query to the user to complete the intent with the information provided by the user. This process ensures that the slow inference system can handle vague and incomplete user requests and transform them into clear, executable, and complete intents.

[0075] Building upon the above embodiments which describe how slow inference systems can handle immediate requests with low matching degrees, please also refer to... Figure 6 , Figure 6 The flowchart of a method for processing problem investigation requests and performing knowledge updates using a slow inference system provided in this embodiment of the disclosure, that is, this embodiment also provides a method for using a slow inference system to process problem investigation requests initiated by a second user (to distinguish them from the first user) (to distinguish them from immediate requests with low matching degree), and its process 600 includes the following steps: Step 601: Use the slow inference system to process the troubleshooting request initiated by the second user in response to the abnormal situation, and obtain the troubleshooting results; Unlike the primary user, who is a regular user, the secondary user (usually a system operator or administrator), who is a higher-level user, often initiates "problem investigation requests" that point to complex and wide-ranging anomalies (such as "a sudden 50% drop in package distribution today" or "users unable to submit orders"). The causes of these problems typically involve multiple aspects (such as servers, interfaces, and business processes), and a single investigation method can easily miss crucial information. Therefore, slow inference systems require at least two investigation methods to cross-validate and pinpoint the true cause. The technical principle behind this design is that the causes of anomalies in complex systems have a "multiple causes, one effect" characteristic; multiple investigation methods can cover different dimensions of possibilities, avoiding the one-sidedness of a single method.

[0076] In a practical scenario, for example, a second user discovers that "the number of packets distributed suddenly decreased after 2 PM today" and initiates a troubleshooting request. The Slow Inference system breaks down the troubleshooting into two core tasks: data trend analysis (Method 1) and system log auditing (Method 2). Data trend analysis, by accessing the real-time database of the packet distribution system, finds that after 2 PM, the number of "failed packets" surged from 10 to 200 per hour, with the "failure reasons" concentrated on "server timeout." System log auditing, by parsing the server's Nginx logs, finds a large number of "504 Gateway Timeout" errors at 2 PM (the server could not respond in time). Both methods point to "server timeout" as the cause, thus the Slow Inference system determines that "server timeout caused a decrease in packet distribution" as the real reason. If only data trend analysis is used, it might be mistakenly judged as "reduced business demand"; if only log auditing is used, the business impact of "the surge in failed packets" might be overlooked. Verification through multiple methods ensures the accuracy of the cause.

[0077] Step 602: Add the new knowledge contained in the problem investigation results to the shared information database.

[0078] New knowledge refers to causal relationships or solutions that explain the cause of anomalies but are not stored in the shared information database during problem investigation (such as "server timeout → reduced packet distribution" or "504 error → user unable to submit order"). Once this knowledge is added to the shared information database, it will become the key related knowledge for the fast inference system to handle similar requests from ordinary users.

[0079] Furthermore, multi-faceted investigation not only ensures accurate cause identification but may also uncover hidden correlations. For example, when investigating a "distribution package decline," simultaneously employing "user feedback analysis" (method 3, distinguishing it from methods 1 and 2) might reveal that "users are churned due to their inability to submit orders," thus supplementing the knowledge of "server timeout → user churn" and further enriching the shared information database. The value of knowledge looping lies in reducing system maintenance costs: as the knowledge of anomaly causes accumulates in the shared information database, the fast inference system can handle more requests from ordinary users, reducing the frequency of slow inference calls and allowing slow inference resources to be concentrated on more complex problems (such as "distribution package decline accompanied by user data leakage").

[0080] In this embodiment, step 601 ensures the accuracy of the cause of the anomaly through multiple investigations, while step 602 forms the system's "self-evolution" through knowledge supplementation. The combination of the two not only solves the current anomaly problem, but also provides a better solution for similar requests in the future, achieving the effect of "handling one problem and optimizing a class of problems".

[0081] Based on the above embodiments, please also refer to Figure 7 , Figure 7 A flowchart of a method for generating and processing problem investigation results using multiple intelligent agents is provided in this disclosure embodiment. This method aims to fully utilize the problem investigation results generated by a slow inference system. The process 700 includes the following steps: Step 701: Use the analysis report to generate an intelligent agent to generate a problem analysis report based on the problem investigation results; In this step, the problem analysis report serves as a structured review of the abnormal situation by the slow inference system. Its core function is to present the "investigation process," "causal conclusions," and "solutions" to a second user (such as team or enterprise operators) in a professional and clear manner. The responsibility of the analysis report generating agent is to integrate key information from the investigation results (such as data trends, log errors, and verification methods) and generate a report according to a preset template.

[0082] For example, when the investigation result is "server timeout caused a decrease in the number of distributed packets", the agent will extract the following elements: Problem description: After 14:00 today, the number of packets sent dropped sharply by 50%, while the number of failed packets surged; Troubleshooting methods: Data trend analysis (accessing the real-time database revealed that the failures were mainly due to "server timeout") and system log auditing (analyzing Nginx logs revealed a large number of 504 errors). Cause and conclusion: The server was unable to respond to the request in a timely manner, resulting in the failure to distribute the package; Solution: Restart the server and optimize interface response time (e.g., by adding caching).

[0083] At the technical level, the report generation agent can employ template-based natural language generation technology. It pre-sets a fixed structure of "problem-investigation method-cause-solution," and fills the template with structured data extracted during the investigation process (such as "200 failed packets after 14:00" and "80% of 504 errors") to generate a professional report that conforms to the reading habits of operators. This approach ensures the consistency (uniform report structure for different problems) and accuracy (avoiding the omission of key information), facilitating operators to quickly understand the root cause of the problem and take appropriate measures.

[0084] Step 702: Utilize the content generation agent to generate new content corresponding to the abnormal state based on the problem analysis report; Considering that the analysis report is intended for operators, while ordinary first-time users (such as merchants using the distribution package service) need concise, conversational information (such as "What caused the distribution package delay?" "Has it been fixed?"), the content generation agent's responsibility is to reconstruct the key information in the professional report and transform it into "new content" (such as user notifications, help documents, and short videos) that users can quickly understand. For example, for the analysis report above, the agent will extract three core pieces of information: "Problem (distribution package delay), Cause (server timeout), Solution (fixed)," and generate the following user notification: "Dear merchants, the delay in some distribution packages after 14:00 today was due to a temporary server timeout. We have urgently fixed it, and the distribution packages are now back to normal. We apologize for any inconvenience caused!" At the technical level, content-generating agents can employ content reconstruction models (such as BART-based text summarization models) and style adaptation modules (such as conversational conversion): First, extract "information that users care about" (such as "has it been fixed?" and "what impact will it have on me") from the analysis report. Then, reorganize the content using simple, direct language (avoiding technical terms such as "504 error" and "server timeout") to ensure users can quickly grasp the key points. In practice, the agent will also adjust the content style according to user type; for example, notifications to individual users will be more friendly ("Dear User"), while notifications to enterprise users will be more formal ("Dear Enterprise Customer").

[0085] Step 703: Use the content publishing agent to push new content to the first matched user.

[0086] The first user in this matching process refers to users affected by the abnormal situation or who may be concerned about the issue (such as merchants with distribution records in the past 7 days, or users in the Beijing area). Therefore, the responsibility of the content publishing agent is to push new content to these users based on user profile matching, avoiding interference to irrelevant users.

[0087] For example, regarding the issue of "server timeout causing packet delivery delays," the agent will target the following recipients: Region matching: Users in the Beijing area (because the server timeout affected the Beijing node); Behavioral matching: Merchants with distribution records in the last 7 days (because they are most likely to be affected); Interest matching: Users who have previously searched for "distribution delay" (because they are concerned about this type of issue).

[0088] At the technical level, the content delivery agent relies on a user profiling system (such as a collaborative filtering-based recommendation model) to generate user profiles based on users' historical behavior (such as query records and usage scenarios) and attributes (such as region and industry). Then, it matches the "tags" of new content (such as "distribution package," "Beijing," and "server issue") with these user profiles to filter out target users. In practice, the agent sends new content to target users through the "message center" or "push notification" functions of the dialogue response application. For example, it might send a text message to a merchant in Beijing stating, "The distribution package delay issue you are concerned about has been resolved. The reason was a temporary server timeout, but it has now returned to normal." Or it might pop up a notification within the application to ensure users receive the information promptly.

[0089] In the solution provided in this embodiment, the report generation agent ensures that operators can quickly understand the root cause of the problem, the content generation agent ensures that users can understand the problem resolution, and the content publishing agent ensures that information is accurately delivered to those who need it. The collaboration of these three agents achieves a closed loop from "technical problem resolution" to "user experience restoration." For example, without the content generation agent, users might receive reports full of technical jargon ("Server 504 error caused distribution package failure"), which they cannot understand; without the content publishing agent, users might not see the repair notification and continue to worry about the distribution package issue. This collaboration not only improves user satisfaction but also reduces repeated user queries (such as "Has the distribution package been restored?") and reduces the pressure on the fast inference system. In other words, the core logic of the solution provided in this embodiment is: to summarize the problem with professional reports, to convey information with easy-to-understand content, and to reach users with precise push notifications, ensuring that the impact of abnormal situations is minimized, while providing experience for handling similar problems in the future (such as adding knowledge about "server timeout → distribution package delay" to the shared information database).

[0090] Based on any of the above embodiments, please also refer to Figure 8 , Figure 8 The flowchart of a method for deleting knowledge from a shared information base and triggering updates based on the amount of knowledge stored, provided for embodiments of this disclosure, is intended to provide relevant optimization solutions for the shared information base. The process 800 includes the following steps: Step 801: Assign a corresponding popularity value to the corresponding knowledge based on the number of times it is used in the shared information database; The popularity value is directly proportional to the number of times it is used within a unit time period; The "popularity score" is a quantitative indicator of the "frequency of use" of each piece of knowledge in the shared information database. Its core logic is that the more times a piece of knowledge is used, the greater its value to users. The calculation method can be the number of times it is used within a unit time period (such as the last 7 days), and the popularity score is directly proportional to the number of times it is used (for example, if a piece of knowledge is called 50 times by the fast reasoning system, the popularity score is 50; if it is called 5 times, the popularity score is 5).

[0091] At the technical principle level, the calculation of popularity value can rely on log recording (the server records the time and scenario when each piece of knowledge is called by the fast inference / slow inference system), and the timeliness of the data is ensured by using sliding window statistics (such as the cumulative number of times in the last 7 days). In practice, for example, the knowledge "server timeout caused the distribution package to drop" in the shared information database was used by the fast inference system to answer the request "the distribution package dropped today" 50 times in the last 7 days, while "2023 Q1 distribution package data" was only called once. The former's popularity value (50) is much higher than the latter's (1), indicating that the former is more in line with the current user's needs.

[0092] Step 802: Delete knowledge with a popularity value lower than the preset popularity lower limit from the shared information database; The "preset minimum popularity threshold" is the minimum usage threshold set by the server for knowledge (e.g., ≥3 times per week). Its purpose is to filter out "low-value knowledge" that has not been used for a long time and to prevent the shared information database from reducing retrieval efficiency due to storing too much redundant information.

[0093] At the technical level, the presence of low-profile knowledge increases the retrieval burden on fast reasoning systems. (For example, when searching for "reasons for the decline in distribution packages," it is necessary to traverse 1000 pieces of knowledge, of which 900 are old, unused data.) Deleting this knowledge can improve search speed (e.g., reducing it to 100 pieces). In practice, for example, if the preset minimum popularity is "weekly usage ≥ 3 times," then "2023 Q1 distribution package data" (weekly usage 1 time) will be marked as low-popularity knowledge, and the server will periodically perform a deletion operation every Sunday evening, removing it from the shared information database. This method ensures that the knowledge in the shared information database is always "what the current user needs most" (such as "server timeout → distribution package decline," "504 error → user cannot submit order," etc., which are recent high-frequency knowledge).

[0094] Step 803: In response to the knowledge storage in the shared information base exceeding the preset storage threshold, use the knowledge in the shared information base to trigger an update training of the base model that constitutes the fast reasoning system and the slow reasoning system.

[0095] The base model is the fundamental large language model for both fast reasoning systems (such as OR LLM, large online reasoning models) and slow reasoning systems (such as Reasoner LLM, large deep reasoning models), and its performance directly determines the accuracy of reasoning. Therefore, when the storage capacity of the shared information repository exceeds a preset threshold (such as 1 million records), the server can use the knowledge in the repository to fine-tune the base model, realizing the transformation of knowledge into model capabilities.

[0096] At the technical level, the "knowledge update" of the base model can employ incremental fine-tuning. This involves using the latest knowledge added to the shared information repository (such as causal relationship knowledge added in the last month, such as "server timeout → reduced distribution packages" or "third-party interface failure → order submission failure") as training data to adjust the model's parameters, allowing the model to better understand these new business logics. In practical terms, for example, when the shared information repository contains more than 1 million entries, the server selects the 100,000 most recently added knowledge entries (approximately 10% of the total) and uses LoRA (Low-Rank Adaptation) technology to fine-tune the model. After fine-tuning, when a user asks "the distribution packages have decreased today," the model can more accurately associate it with the cause "server timeout," rather than using old, potentially outdated knowledge (such as "rainy weather caused delivery delays").

[0097] Furthermore, to optimize the knowledge management process, in addition to the number of uses, the aforementioned implementing entities can also adjust the popularity value based on user feedback (such as the number of times users click on the knowledge links returned by quick reasoning). For example, if a piece of knowledge is called 50 times and a user clicks it 30 times, the popularity value can be weighted to 60, more accurately reflecting the "practical value" of the knowledge. They can also adjust the preset lower limit of popularity according to business scenarios (such as during e-commerce promotions, the popularity value of the knowledge "reason for order delay" will rise, and the server can temporarily lower the lower limit to avoid accidentally deleting useful knowledge). Knowledge distillation technology can also be used to condense the knowledge in the shared information base into small models, and then use the small models to fine-tune the base model to further reduce the consumption of computing resources.

[0098] This embodiment, through the solutions provided in steps 801 to 803, forms a closed loop of "knowledge generation → knowledge use → knowledge optimization → model evolution." Specifically, new knowledge generated by the slow inference system is added to the shared information database. The server uses popularity value calculation to filter out frequently used knowledge and delete low-popularity knowledge, thereby maintaining the efficiency of the shared information database. At the same time, when the storage volume of the shared information database exceeds the threshold, the base model is fine-tuned with the latest knowledge to improve the model's reasoning ability. The optimized model can then better handle user requests and generate more accurate immediate responses (fast inference) or deep replies (slow inference), further improving the user experience.

[0099] With the rapid development of large-scale model technology in fields such as AI search and general intelligent agents, its capabilities are becoming increasingly rich. However, its application in traditional engineering fields (such as enterprise-level search systems) remains relatively rare. Meanwhile, the training and inference costs of large-scale models remain high, and how to transform them into monetizable productivity has become a core problem plaguing enterprises. Traditional search systems cannot meet users' needs for "accurate cause analysis and in-depth solutions," and lack "reflection-memory" capabilities, meaning results do not iterate and optimize with usage. Using large models to handle all requests would lead to exorbitant costs, making large-scale application difficult.

[0100] To address the aforementioned issues, this embodiment proposes a sustainable evolutionary human-machine co-creation system architecture (such as one built around intelligent agent ecosystem business and based on end-to-end multi-agent technology) that revolves around intelligent agent ecosystem business. Figure 9 As shown, the traditional search system is reconstructed through "fast and slow reasoning collaboration + multi-agent scheduling" to achieve a balance between "intelligent evolution" and "cost convergence".

[0101] The system is divided into an online fast inference system and a near-line slow inference system, which work together to cover user needs in different scenarios: The online fast reasoning system, composed of a Search Agent, handles real-time, high-frequency user requests (such as "What's the weather like today?" or "Where is my order?"). It employs traditional retrieval techniques (inverted index, ANN) to quickly return basic results, ensuring fast response times (latency ≤ 1 second). For example, if a user inputs "Today's weather in Beijing," the fast reasoning system quickly matches the structured data for "Today's weather in Beijing" using an inverted index, directly returning "Sunny, 15-22℃," achieving low cost and high efficiency.

[0102] The near-line slow inference system, composed of Reasoner Agents, handles complex, low-frequency but high-value requests (such as "What caused the sudden drop in packet distribution volume today?" and "How to resolve server timeout issues?"). These requests require in-depth analysis (such as cross-system log auditing and multi-dimensional data correlation), which traditional technologies cannot meet, necessitating the use of large models for reflection and reasoning. For example, if a user asks "The packet distribution volume has decreased," the fast inference system recognizes the need for in-depth cause analysis and forwards the request to the slow inference system. The slow inference system, by retrieving data from multiple sources such as server logs and the packet distribution database, infers the conclusion that "Server timeouts have led to a surge in failed packet distributions" and generates a solution of "Restarting the server + optimizing the interface."

[0103] Specifically, the slow inference system adopts a multi-agent architecture, achieving end-to-end processing "from request to high-quality content" through the division of labor among three core agents: QA-Trigger Agent (Request-Agent Matching Agent): Responsible for accurately matching user requests (Q) with agents in the Wenxin Agent Platform (an ecosystem platform for developers to create agents). For example, if a user asks "How to choose a laptop", the agent, through semantic understanding, matches the user with the "Computer Selection Agent" on the platform (a specialized agent created by developers for the "Computer Selection" scenario), ensuring that the request and the agent's capabilities are "aligned".

[0104] Producer Agent (Content Production Agent): Based on the matched agent and user request, it generates high-quality structured content. For example, for a "computer buying" request, the Producer Agent calls upon the knowledge graph of the "computer buying agent" (including dimensions such as budget, configuration, and brand) to generate a "2024 Laptop Buying Guide." The content covers key points that users care about, such as "Recommended Models Under 5000 Yuan" and "Gaming Laptops vs. Ultrabooks," which is more valuable than the "fragmented results" of traditional search.

[0105] Chat Agent (Conversational Interaction Intelligent Agent): Responsible for engaging in natural language conversations with users, adjusting content based on user feedback, and achieving application-level conversion. For example, after reading a "Computer Buying Guide," a user might ask, "With a budget of 5000 yuan, which gaming laptop should I choose?" The Chat Agent will combine the user's budget (5000 yuan) with the core requirements of a gaming laptop (graphics card, processor), recommending specific models such as "XXXXX Model 1" and "XXXXX Model 2," and providing purchase links to guide the user to complete the conversion.

[0106] The following is a specific usage flow to demonstrate the entire process:

[0107] 1. For search users, it can provide a smarter "search experience".

[0108] The user experience is the same as traditional search (enter search terms → get results), but the underlying system has been implemented... "Traditional retrieval + large-scale model reflection + memory" Upgrade: 1) Reflection phase: For example, if a user asks "The distribution package has decreased today", the fast inference system first recalls "yesterday's distribution package data", and the slow inference system will "reflect" on: "Did we miss any server logs?" "Is there a third-party interface failure?" By cross-system data correlation, a more comprehensive reason can be found.

[0109] 2) Memory process: The system will remember the user's needs and preferences (such as the user repeatedly asking "solution" instead of "cause analysis"). The next time a similar request is encountered, "solution" will be returned first (such as "server has been repaired and distribution package has returned to normal"). The results become more and more accurate with the number of uses.

[0110] 3) Cost convergence: The division of labor between fast and slow inference avoids the high cost of "using a large model for all". Fast inference handles 80% of simple requests (using traditional technology, low cost) and slow inference handles 20% of complex requests (using a large model, high value). The overall cost is reduced by more than 60% compared to using a large model for all requests.

[0111] 2. For platform users, it provides a "data flywheel" for human-machine collaboration.

[0112] For platform developers (creating intelligent agents) and operators (managing content), the system achieves intelligent evolution through human-machine collaboration: 1) Intelligent prompts: The system analyzes the matching data of user requests and intelligent agents. If it finds that the number of requests for "computer selection" has increased (week-on-week +50%), but the number of "computer selection intelligent agents" on the platform is insufficient (only 3), it will automatically prompt the developer to "add a computer selection intelligent agent" to fill the demand gap.

[0113] 2) Data closed loop: The system tracks content distribution and conversion data (such as the number of clicks and forwarding rate of "Computer Buying Guide"). If the forwarding rate of a certain piece of content is as high as 20% (far higher than the average of 5%), its distribution will be increased (such as being placed at the top of the search results), forming an end-to-end data flywheel of "increased user demand → intelligent agent optimization / addition → improved content quality → improved user experience → further increase in demand", driving the prosperity of the ecosystem.

[0114] In other words, the solution provided in this embodiment helps the product by transforming the "cost" of a large model into "productivity," in order to achieve two key objectives: 1. Improved Distribution and Conversion Efficiency: Through precise matching of QA-Trigger Agents, high-quality content from Producer Agents, and conversion by Chat Agents, the conversion efficiency of Q2C (content found) and Q2V (value found) is significantly improved. For example, click-through rates for "computer shopping" requests increased by 40%, and the purchase rate increased by 30%, directly driving sales. 2. Model Optimization: The system uses large models in high-value scenarios (such as application transformation and deep inference) to avoid the waste of "intelligence for the sake of intelligence". For example, the "distribution package decline analysis" request processed by the slow inference system accounts for only 20%, but it can help enterprises quickly solve problems (reduce losses by 1 million yuan). Its value is far higher than the cost of processing simple requests, realizing a positive cycle of "model cost → application benefit".

[0115] In summary, the system provided in this embodiment has fundamentally reconstructed the underlying logic of traditional search systems through "fast and slow reasoning collaboration" and "multi-agent architecture," achieving a balance between "improved user experience," "cost convergence," and "application transformation," and providing a reference solution for the implementation of large models in traditional engineering fields.

[0116] Further reference Figure 10 As an implementation of the methods shown in the above figures, this disclosure provides an embodiment of a request processing apparatus based on a fast and slow dual inference system. This apparatus embodiment is similar to... Figure 2 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0117] like Figure 10As shown, the request processing device 1000 based on a fast and slow dual-inference system in this embodiment may include: a fast inference system processing unit 1001, a slow inference system supplementary processing unit 1002, and a first knowledge supplementary unit 1003. The fast inference system processing unit 1001 is configured to process an immediate request initiated by a first user using a preset fast inference system to obtain an immediate response; wherein the fast inference system generates an immediate response for the immediate request based on content indexing technology. The slow inference system supplementary processing unit 1002 is configured to process the immediate request using a preset slow inference system to obtain a request reply in response to a lower-than-preset matching degree between the associated knowledge used to generate the immediate response and the immediate request; wherein the associated knowledge is determined based on information related to the immediate request extracted from a preset shared information database through content indexing technology, and the slow inference system generates a request reply for the immediate request based on task decomposition and multi-intelligence collaboration. The first knowledge supplementary unit 1003 is configured to supplement the shared information database with new knowledge contained in the request reply.

[0118] In this embodiment, the specific processing and technical effects of the fast inference system processing unit 1001, the slow inference system supplementary processing unit 1002, and the first knowledge supplementary unit 1003 in the request processing device 1000 based on the fast and slow dual inference system can be referred to respectively. Figure 2 The relevant descriptions of steps 201-203 in the corresponding embodiments will not be repeated here.

[0119] In some other optional implementations of this embodiment, the fast inference system processing unit 1001 is further configured to: By utilizing the recall mechanism that constitutes the fast reasoning system, a candidate knowledge set related to the immediate request is retrieved from the shared information database through an approximate nearest neighbor search algorithm; The candidate knowledge set is sorted according to a preset sorting algorithm using the sorting mechanism that constitutes the fast reasoning system, resulting in a sorted candidate knowledge set. The intervention and adjustment mechanism that constitutes the fast reasoning system is used to filter relevant knowledge from the sorted candidate knowledge set; By leveraging the large language model that constitutes the fast reasoning system, we can generate instant responses for instant requests based on associative knowledge.

[0120] In some other optional implementations of this embodiment, the request processing apparatus 1000 based on the fast and slow dual inference system may further include: The message queue placement unit is configured to store immediate requests with a matching degree lower than a preset matching degree into a preset message queue before processing the immediate request using the slow inference system. Correspondingly, the fast inference system processing unit 1001 is further configured as follows: The slow inference system is used to process requests retrieved from the message queue and obtain request responses.

[0121] In some other optional implementations of this embodiment, the request processing apparatus 1000 based on the fast and slow dual inference system may further include: The priority determination unit is configured to determine the request processing priority of immediate requests in the message queue; wherein the request processing priority is determined based on at least one of the following: the problem domain of the immediate request, the proportion of relevant knowledge in the shared information base, and importance evaluation. The low-priority deletion unit is configured to delete the corresponding immediate request from the message queue in response to a request processing priority that is lower than the preset priority. The high-priority retention unit is configured to retain the corresponding immediate request in the message queue in response to a request processing priority that is not lower than the preset priority.

[0122] In some other optional implementations of this embodiment, the slow inference system supplementary processing unit 1002 includes: The complete task determination and allocation subunit is configured to use the main intelligent agent constituting the slow inference system to determine the complete task based on immediate requests and a shared information base, and then distribute the complete task into multiple subtasks to the matching sub-intelligent agents. The control execution subunit is configured to use the main agent constituting the slow inference system to control each sub-agent to execute the corresponding sub-task in an orderly manner according to the execution order of each sub-task, and to obtain a request response; wherein, the sub-agent executes the corresponding sub-task based on a shared information base and a model context protocol.

[0123] In some other optional implementations of this embodiment, the complete task determination and allocation subunit includes a complete task determination subunit configured to determine the complete task based on immediate requests and a shared information base using the main intelligent agent constituting the slow inference system. The complete task determination subunit includes: The intent completeness determination module is configured to use the main agent to determine the actual intent completeness expressed by the immediate request; The intent completion module is configured to respond to situations where the actual intent completeness is lower than a preset intent completeness threshold by using the main agent to perform intent completion operations on the immediate request based on a shared information database to obtain a complete intent; wherein, a complete intent is an intent with a completeness not lower than the preset intent completeness threshold. The complete task determination module is configured to use the main agent to determine the complete task corresponding to the complete intent.

[0124] In some other optional implementations of this embodiment, the intent completion module includes: The supplementary request generation submodule is configured to generate supplementary requests that utilize the main agent to complete the intent expressed by the immediate request. The supplementary knowledge search submodule is configured to use the main intelligent agent to search for and supplement the knowledge corresponding to the request in the shared information database; The intent completion submodule is configured to use the main agent to complete the current intent expressed by the immediate request based on supplementary knowledge, so as to obtain the complete intent.

[0125] In some other optional implementations of this embodiment, the intent completion module may further include: The supplementary inquiry submodule is configured to respond to situations where the current intent cannot be completed into a full intent by supplementing knowledge, and to initiate a supplementary inquiry to the first user using the main agent to obtain an inquiry response. The additional completion submodule is configured to use the main agent to complete the current intent based on the query response until a complete intent is obtained.

[0126] In some other optional implementations of this embodiment, the request processing apparatus 1000 based on the fast and slow dual inference system may further include: The slow inference system processing unit is configured to use the slow inference system to process the problem investigation request initiated by the second user in response to the abnormal situation, and obtain the problem investigation result; wherein, the problem requested for investigation needs to be determined by at least two investigation methods to determine the true cause of the abnormal situation. The second knowledge supplement unit is configured to add new knowledge contained in the problem investigation results to the shared information database.

[0127] In some other optional implementations of this embodiment, the request processing apparatus 1000 based on the fast and slow dual inference system may further include: The analysis report generation unit is configured to use the analysis report generation agent to generate a problem analysis report based on the problem investigation results; The new content generation unit is configured to use the content generation agent to generate new content corresponding to the abnormal state based on the problem analysis report; The push unit is configured to use a content publishing agent to push new content to the first matched user.

[0128] In some other optional implementations of this embodiment, the request processing apparatus 1000 based on the fast and slow dual inference system may further include: The popularity value determination unit is configured to assign a corresponding popularity value to the corresponding knowledge based on the number of times different knowledge is used in the shared information database; wherein, the size of the popularity value is directly proportional to the number of times it is used within a unit time period; The low-popularity knowledge deletion unit is configured to delete knowledge with a popularity value lower than a preset popularity threshold from the shared information database.

[0129] In some other optional implementations of this embodiment, the request processing apparatus 1000 based on the fast and slow dual inference system may further include: The trigger update unit is configured to respond to the amount of knowledge stored in the shared information base exceeding a preset storage threshold by using the knowledge in the shared information base to trigger an update training of the base model that constitutes the fast reasoning system and the slow reasoning system.

[0130] This embodiment is a device embodiment corresponding to the above method embodiment. The request processing device based on a fast and slow dual-inference system provided in this embodiment achieves a balance between request processing timeliness and response quality through the collaborative architecture of the fast and slow dual-inference systems. The fast inference system responds to user requests in real time based on content indexing technology, ensuring immediacy. At the same time, by judging the matching degree between related knowledge and requests, requests with low matching degree are further transferred to the slow inference system, which not only reasonably allocates computing resources but also avoids unnecessary deep inference overhead. The slow inference system adopts task decomposition and multi-intelligent collaboration, which can perform deep processing on complex requests, effectively improving the accuracy and depth of responses compared to the fast inference system. Furthermore, by supplementing the shared information base with the new knowledge generated by slow inference, the fast inference system can gradually improve the quality and accuracy of rapid responses in subsequent response generation processes due to the continuous enrichment of related knowledge, thereby achieving continuous evolution of the request processing system. In the long run, it not only improves the response quality of fast inference but also reduces the call frequency of slow inference, and optimizes the efficiency and effect of request processing at the overall level.

[0131] According to embodiments of this disclosure, this disclosure also provides an electronic device, the electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to implement the request processing method based on the fast and slow dual inference system described in any of the above embodiments.

[0132] According to embodiments of this disclosure, this disclosure also provides a readable storage medium storing computer instructions that enable a computer to implement the request processing method based on a fast and slow dual inference system as described in any of the above embodiments.

[0133] According to embodiments of this disclosure, this disclosure also provides a computer program product that, when executed by a processor, can implement the request processing method based on a fast and slow dual inference system described in any of the above embodiments.

[0134] Figure 11 A schematic block diagram of an example electronic device 1100 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0135] like Figure 11 As shown, device 1100 includes a computing unit 1101, which can perform various appropriate actions and processes according to a computer program stored in read-only memory (ROM) 1102 or a computer program loaded into random access memory (RAM) 1103 from storage unit 1108. The RAM 1103 may also store various programs and data required for the operation of device 1100. The computing unit 1101, ROM 1102, and RAM 1103 are interconnected via bus 1104. Input / output (I / O) interface 1105 is also connected to bus 1104.

[0136] Multiple components in device 1100 are connected to I / O interface 1105, including: input unit 1106, such as keyboard, mouse, etc.; output unit 1107, such as various types of monitors, speakers, etc.; storage unit 1108, such as disk, optical disk, etc.; and communication unit 1109, such as network card, modem, wireless transceiver, etc. Communication unit 1109 allows device 1100 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0137] The computing unit 1101 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1101 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 1101 performs the various methods and processes described above, such as a request processing method based on a fast and slow dual-inference system. For example, in some embodiments, the request processing method based on a fast and slow dual-inference system can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 1108. In some embodiments, part or all of the computer program can be loaded and / or installed on device 1100 via ROM 1102 and / or communication unit 1109. When the computer program is loaded into RAM 1103 and executed by the computing unit 1101, one or more steps of the request processing method based on a fast and slow dual-inference system described above can be performed. Alternatively, in other embodiments, computing unit 1101 may be configured by any other suitable means (e.g., by means of firmware) to perform a request processing method based on a fast and slow dual inference system.

[0138] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0139] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0140] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0141] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0142] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0143] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and Virtual Private Server (VPS) services, such as high management difficulty and weak business scalability.

[0144] According to the technical solution of this disclosure, a balance between request processing timeliness and response quality is achieved through a collaborative architecture of fast and slow dual inference systems. The fast inference system responds to user requests instantly based on content indexing technology, ensuring immediacy. Simultaneously, by judging the matching degree between related knowledge and requests, requests with low matching degrees are further transferred to the slow inference system, rationally allocating computing resources while avoiding unnecessary deep inference overhead. The slow inference system, employing task decomposition and multi-intelligence collaboration, can perform in-depth processing of complex requests, effectively improving response accuracy and depth compared to the fast inference system. Furthermore, by supplementing the shared information base with the new knowledge generated by slow inference, the fast inference system can gradually improve the quality and accuracy of its rapid responses in subsequent response generation processes due to the continuous enrichment of related knowledge. This enables the continuous evolution of the request processing system, and in the long run, it improves the response quality of fast inference, reduces the frequency of slow inference calls, and optimizes the overall efficiency and effectiveness of request processing.

[0145] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0146] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A request processing method based on a fast and slow double reasoning system, comprising: processing an instant request initiated by a first user using a preset fast reasoning system to obtain an instant response; wherein the fast reasoning system generates the instant response for the instant request based on a content indexing technology; in response to the matching degree of associated knowledge used to generate the instant response and the instant request being lower than a preset matching degree, processing the instant request using a preset slow reasoning system to obtain a request reply; wherein the associated knowledge is determined based on information related to the instant request extracted from a preset shared information base through the content indexing technology, and the slow reasoning system generates the request reply for the instant request based on task decomposition and multi-intelligent collaboration; supplementing new knowledge contained in the request reply to the shared information base.

2. The method of claim 1, wherein, The processing of the instant request initiated by the first user using the preset fast reasoning system to obtain the instant response comprises: using a recall mechanism constituting the fast reasoning system to recall a candidate knowledge set related to the instant request from the shared information base through an approximate nearest neighbor search algorithm; using a ranking mechanism constituting the fast reasoning system to rank each candidate knowledge in the candidate knowledge set according to a preset ranking algorithm to obtain a ranked candidate knowledge set; using an intervention adjustment mechanism constituting the fast reasoning system to filter the associated knowledge from the ranked candidate knowledge set; using a large language model constituting the fast reasoning system to generate the instant response for the instant request based on the associated knowledge.

3. The method of claim 1, further comprising: storing instant requests with a matching degree lower than the preset matching degree with the associated knowledge in a preset message queue before processing the instant request using the slow reasoning system; correspondingly, the processing of the instant request using the preset slow reasoning system to obtain the request reply comprises: processing the request taken out of the message queue using the slow reasoning system to obtain the request reply.

4. The method of claim 3, further comprising: determining the request processing priority of the instant request in the message queue; wherein the request processing priority is determined based on at least one of the problem domain of the instant request, the proportion of related knowledge in the shared information base, and the importance evaluation; in response to the request processing priority being lower than a preset priority, deleting the corresponding instant request from the message queue; in response to the request processing priority not being lower than the preset priority, retaining the corresponding instant request in the message queue.

5. The method of claim 1, wherein, The processing of the instant request using the preset slow reasoning system to obtain the request reply comprises: using a master agent constituting the slow reasoning system to determine a complete task based on the instant request and the shared information base, and to split the complete task into a plurality of sub-tasks and then issue them to matched sub-agents. The main agent of the slow reasoning system is used to control the sub-agents to execute the sub-tasks in sequence, and the request reply is obtained; wherein, the sub-agents execute the sub-tasks based on the shared information base and the model context protocol.

6. The method of claim 5, wherein, The main agent of the slow reasoning system is used to determine the complete task based on the instant request and the shared information base, including: The main agent is used to determine the actual intent completeness of the intent expressed by the instant request; In response to the actual intent completeness being lower than a preset intent completeness threshold, the main agent is used to perform intent completion operation on the instant request based on the shared information base, and a complete intent is obtained; wherein, the complete intent has an intent that is not lower than the preset intent completeness threshold; The main agent is used to determine the complete task corresponding to the complete intent.

7. The method of claim 6, wherein, The main agent is used to perform intent completion operation on the instant request based on the shared information base, and a complete intent is obtained, including: The main agent is used to generate a supplementary request for completing the intent expressed by the instant request; The main agent is used to search for supplementary knowledge corresponding to the supplementary request in the shared information base; The main agent is used to complete the current intent expressed by the instant request based on the supplementary knowledge, and the complete intent is obtained.

8. The method of claim 7, further comprising: In response to being unable to complete the current intent to the complete intent through the supplementary knowledge, the main agent is used to initiate a supplementary inquiry to the first user, and an inquiry reply is obtained; The main agent is used to complete the current intent according to the inquiry reply until the complete intent is obtained.

9. The method of claim 1, further comprising: The slow reasoning system is used to process a problem investigation request initiated by a second user for an abnormal condition, and a problem investigation result is obtained; wherein, the problem to be investigated by the problem investigation request needs to be determined through at least two investigation methods to determine the real cause of the abnormal condition; New knowledge contained in the problem investigation result is supplemented to the shared information base.

10. The method of claim 9, further comprising: An analysis report generation agent is used to generate a problem analysis report for the problem investigation result; A content generation agent is used to generate new content corresponding to the abnormal state according to the problem analysis report; A content publishing agent is used to push the new content to the matched first user.

11. The method of any one of claims 1-10, further comprising: According to the number of times different knowledge stored in the shared information base is used, a corresponding heat value is assigned to the corresponding knowledge; wherein, the size of the heat value is in a positive relationship with the number of times used in a unit time period; Knowledge with a heat value lower than a preset heat lower limit value is deleted from the shared information base.

12. The method of claim 11, further comprising: In response to the amount of knowledge in the shared information base exceeding a preset storage amount threshold, triggering an update training of base models constituting the fast reasoning system and the slow reasoning system using the knowledge in the shared information base.

13. A request processing apparatus based on a fast-slow dual reasoning system, comprising: a fast reasoning system processing unit configured to process a first user-initiated instant request using a preset fast reasoning system to obtain an instant response; wherein the fast reasoning system generates the instant response for the instant request based on a content indexing technology; a slow reasoning system supplementary processing unit configured to, in response to a matching degree of associated knowledge used to generate the instant response with the instant request being lower than a preset matching degree, process the instant request using a preset slow reasoning system to obtain a request reply; wherein the associated knowledge is determined based on information related to the instant request extracted from a preset shared information base through the content indexing technology, and the slow reasoning system generates the request reply for the instant request based on task decomposition and multi-intelligence collaboration; a first knowledge supplementing unit configured to supplement new knowledge contained in the request reply to the shared information base.

14. An electronic device, comprising: at least one processor; and a memory connected in communication with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the request processing method based on a fast-slow dual reasoning system according to any one of claims 1-12.

15. A non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform the request processing method based on a fast-slow dual reasoning system according to any one of claims 1-12.

16. A computer program product comprising a computer program, the computer program being executed by a processor to implement the steps of the request processing method based on a fast-slow dual reasoning system according to any one of claims 1-12.

Images