Information inference method and device, computer device, and storage medium

Abstract

Embodiments of the present application provide an information reasoning method and device, computer equipment and a storage medium, belonging to the field of artificial intelligence and large model technology. The method comprises: obtaining reference configuration parameters and a reference key-value block of reference semantic fingerprint information; wherein the reference configuration parameters comprise: a reference application scenario category, a reference user level and a reference input frequency; performing priority division on the reference semantic fingerprint information according to the reference application scenario category, the reference user level and the reference input frequency to obtain a reference priority level; storing the reference key-value block to a preset target cache queue according to the reference priority level; receiving current semantic fingerprint information; extracting a target key-value block from the target cache queue according to the current semantic fingerprint information and the reference semantic fingerprint information; and performing information reasoning according to the target key-value block. Embodiments of the present application can reduce the loading time of information reasoning and improve recommendation efficiency.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and large model technology, and in particular to an information reasoning method and apparatus, computer equipment and storage medium. Background Technology

[0002] With the large-scale application of large models in fields such as natural language processing and computer vision, the response speed and computational cost during the inference phase have gradually become the core bottlenecks restricting system performance. Specifically, during the inference process of large models, the self-attention mechanism needs to calculate the association relationship between each token in the input sequence and all other tokens, and generate a key matrix and a value matrix for each token, i.e., a KV cache. The repeated computation of this KV cache and its high memory consumption are the main reasons for the increased inference latency and the escalating computational cost.

[0003] In related technologies, to address the issues of redundant computation and high GPU memory consumption, prompt word caching technology reuses already computed key-value (KV) caches, avoiding repeated calculations on similar or identical input prefixes. However, current methods using a unified cache queue for prompt words do not consider usage scenarios, user levels, or input frequency, which can easily affect the caching efficiency of popular KV caches, thereby impacting the efficiency of information inference in critical scenarios. Therefore, reducing the loading time for information inference and improving recommendation efficiency has become an urgent technical problem to be solved. Summary of the Invention

[0004] The main objective of this application is to propose an information reasoning method and apparatus, computer equipment and storage medium, which aims to reduce the loading time of information reasoning and improve recommendation efficiency.

[0005] To achieve the above objectives, a first aspect of this application proposes an information reasoning method, the method comprising: Obtain reference configuration parameters and reference key-value blocks for reference semantic fingerprint information; wherein, the reference configuration parameters include: reference application scenario category, reference user level, and reference input frequency; The reference semantic fingerprint information is prioritized according to the reference application scenario category, the reference user level, and the reference input frequency to obtain the reference priority level; The reference key-value block is stored in a preset target cache queue according to the reference priority level; Receive the current semantic fingerprint information; Extract the target key-value block from the target cache queue based on the current semantic fingerprint information and the reference semantic fingerprint information; Information reasoning is performed based on the target key value block.

[0006] In some embodiments, the reference semantic fingerprint information is multimodal semantic fingerprint information, and the reference configuration parameters for obtaining the reference semantic fingerprint information include: Acquire multimodal prompt data and the reference configuration parameters of the multimodal prompt data; wherein the multimodal prompt data includes: text prompt information, image prompt information, and voice prompt information; The text prompt information is semantically fingerprinted to obtain text fingerprint information; The image prompt information is semantically fingerprinted to obtain image fingerprint information; The voice prompt information is semantically fingerprinted to obtain voice fingerprint information; The text fingerprint information, the image fingerprint information, and the voice fingerprint information are concatenated according to the reference application scenario category to obtain the reference semantic fingerprint information.

[0007] In some embodiments, the step of concatenating the text fingerprint information, the image fingerprint information, and the voice fingerprint information according to the reference application scenario category to obtain the reference semantic fingerprint information includes: A selected fingerprint weight set is selected from a preset candidate fingerprint weight set according to the reference application scenario category; wherein, the selected fingerprint weight set includes: a first weight, a second weight, and a third weight; The text fingerprint information, the image fingerprint information, and the voice fingerprint information are concatenated according to the first weight, the second weight, and the third weight to obtain the reference semantic fingerprint information.

[0008] In some embodiments, prioritizing the reference semantic fingerprint information based on the reference application scenario category, the reference user level, and the reference input frequency to obtain a reference priority level includes: The first level is selected from the preset candidate levels based on the reference application scenario category; A second level is selected from the candidate levels based on the reference user level; A third level is selected from the candidate levels based on the reference input frequency; The first level, the second level, and the third level are concatenated according to a preset set of level weights to obtain the reference priority level.

[0009] In some embodiments, the reference semantic fingerprint information is stored in a vector retrieval library; The step of extracting the target key-value block from the target cache queue based on the current semantic fingerprint information and the reference semantic fingerprint information includes: Fingerprint similarity is obtained by calculating the similarity between the current semantic fingerprint information and the reference semantic fingerprint information in the vector retrieval library; Obtain the current configuration parameters of the current semantic fingerprint information; The current semantic fingerprint information is prioritized according to the current configuration parameters to obtain the current priority level; The current semantic fingerprint information is retrieved and evaluated based on the current priority level and the fingerprint similarity to obtain retrieval evaluation information; wherein, the retrieval evaluation information indicates whether the vector retrieval library contains or does not contain the reference semantic fingerprint information that matches the current semantic fingerprint information; If the retrieval evaluation information indicates that the vector retrieval library contains reference semantic fingerprint information that matches the current semantic fingerprint information, the target key-value block is extracted from the target cache queue according to the current priority level and the current semantic fingerprint information.

[0010] In some embodiments, the step of retrieving and evaluating the current semantic fingerprint information based on the current priority level and the fingerprint similarity to obtain retrieval evaluation information includes: The target similarity threshold is selected from the preset candidate similarity thresholds based on the current priority level; The fingerprint similarity is compared with the target similarity threshold to obtain the comparison result; The retrieval evaluation information is determined based on the comparison results.

[0011] In some embodiments, after retrieving and evaluating the current semantic fingerprint information based on the current priority level and the fingerprint similarity to obtain retrieval evaluation information, the method further includes: The hit ratio is obtained by collecting the retrieval evaluation information according to a preset time period, which represents the ratio of the reference semantic fingerprint information that matches the current semantic fingerprint information in the vector retrieval store. The target similarity threshold is adjusted based on the hit ratio to obtain an updated similarity threshold; The retrieval evaluation information is updated based on the updated similarity threshold.

[0012] To achieve the above objectives, a second aspect of this application provides an information reasoning apparatus, the apparatus comprising: The acquisition module is used to acquire reference configuration parameters and reference key-value blocks for reference semantic fingerprint information; wherein, the reference configuration parameters include: reference application scenario category, reference user level, and reference input frequency; The priority division module is used to divide the reference semantic fingerprint information into priorities based on the reference application scenario category, the reference user level, and the reference input frequency to obtain a reference priority level. The storage module is used to store the reference key-value block into a preset target cache queue according to the reference priority level; The receiving module is used to receive the current semantic fingerprint information; The key-value block extraction module is used to extract target key-value blocks from the target cache queue based on the current semantic fingerprint information and the reference semantic fingerprint information; The information reasoning module is used to perform information reasoning based on the target key value block.

[0013] To achieve the above objectives, a third aspect of the present application provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method described in the first aspect.

[0014] To achieve the above objectives, a fourth aspect of the present application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect.

[0015] The information reasoning method, apparatus, computer device, and storage medium proposed in this application acquire configuration parameters in three dimensions—reference application scenario category, reference user level, and reference input frequency—of the reference semantic fingerprint information. Reference key-value blocks are then prioritized and stored in a target cache queue according to their assigned priority levels. Compared to the "unified cache queue" approach used in existing technologies, this application prioritizes high-priority key-value blocks, preventing them from being flushed by low-priority blocks and increasing the probability of cache hits. Furthermore, after receiving the current semantic fingerprint information, the application directly extracts the target key-value blocks stored according to priority from the target cache queue based on the matching relationship between the current semantic fingerprint information and the reference semantic fingerprint information, and performs reasoning accordingly. Because high-priority, frequently reused key-value blocks are cached at the head of the queue or in the active area, and are pre-cached, the target key-value blocks can be quickly extracted. This not only avoids the overhead of repeatedly calculating key-value blocks but also shortens the target key-value block retrieval time, reduces information reasoning latency, and improves the efficiency of information reasoning. Attached Figure Description

[0016] Figure 1 This is a flowchart of the information reasoning method provided in the embodiments of this application; Figure 2 yes Figure 1 The flowchart of step S101 in the text; Figure 3 yes Figure 2 The flowchart of step S205 in the document; Figure 4 yes Figure 1 The flowchart of step S102 in the document; Figure 5 yes Figure 1 The flowchart of step S105 in the process; Figure 6 yes Figure 5 The flowchart of step S504 in the process; Figure 7 This is a flowchart of an information reasoning method provided in another embodiment of this application; Figure 8 This is an overall flowchart of the information reasoning method provided in the embodiments of this application; Figure 9 This is a schematic diagram of the structure of the information reasoning device provided in the embodiments of this application; Figure 10 This is a schematic diagram of the hardware structure of the computer device provided in the embodiments of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0018] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0020] First, let's analyze some of the terms used in this application: Prompt caching is a performance optimization technique in the field of Large Language Models (LLM). It caches and reuses the computational state of large, repetitive parts of user requests (typically system instructions, tool definitions, long texts, etc.).

[0021] Large Language Models (LLMs) are deep learning models trained on large amounts of text data that can generate natural language text or understand the meaning of language text. LLMs can handle various natural language tasks, such as text classification, question answering, and dialogue, and are an important pathway to artificial intelligence.

[0022] Key-Value Cache (KV Cache) is a key optimization technique used to accelerate the autoregressive inference process of large language models such as Transformers. Its core idea is to cache the computed key and value vectors during the autoregressive generation process for reuse in subsequent generation steps, thereby avoiding redundant computation.

[0023] With the large-scale application of language models, response speed and computational cost during the inference phase have gradually become the core bottlenecks restricting system performance. During large model inference, the self-attention mechanism needs to calculate the relationships between each token and other tokens in the input sequence, and generate a key matrix and a value matrix (KV Cache) for each token. The repetitive computation of this KV Cache and its high memory consumption are the main reasons for increased inference latency and escalating computational costs.

[0024] To address the aforementioned issues, related technologies have proposed prompt word caching techniques. The core idea is to reuse the already computed key-value (KV) cache, thereby avoiding redundant computations on similar or identical input prefixes. Currently publicly available prompt word caching schemes include: OpenAI's prefix caching technology, which reuses the KV cache by matching repeated prefixes in the input sequence, thus shortening the first token generation time; and the EPIC framework, which employs a position-independent caching mechanism to effectively solve the cache invalidation problem caused by changes in text position. However, most related technologies focus on single-dimensional cache optimization—either emphasizing precise prefix matching, position independence, semantic similarity, or computational parallelization. Specifically, these technologies commonly employ a "unified cache queue" management approach, treating high-priority and low-priority requests equally. When cache space is limited, the hit latency of high-priority requests increases, severely impacting the efficiency of information reasoning.

[0025] Based on this, embodiments of this application provide an information reasoning method and apparatus, computer device, and storage medium, aiming to obtain configuration parameters in three dimensions—reference application scenario category, reference user level, and reference input frequency—of reference semantic fingerprint information, and to prioritize reference key-value blocks, storing them in a target cache queue according to the assigned priority level. Compared to the "unified cache queue" method used in the prior art, this application allows high-priority key-value blocks to be stored in the cache first, avoiding being flushed by low-priority key-value blocks and increasing the probability of cache hits. Furthermore, after receiving the current semantic fingerprint information, based on the matching relationship between the current semantic fingerprint information and the reference semantic fingerprint information, the target key-value blocks stored according to priority are directly extracted from the target cache queue, and reasoning is performed accordingly. Because high-priority, frequently reused key-value blocks are cached at the head of the queue or in the active area, and are pre-cached, target key-value blocks can be quickly extracted, avoiding the overhead of repeatedly calculating key-value blocks, shortening the target key-value block extraction time, reducing information reasoning latency, and improving the efficiency of information reasoning.

[0026] The information reasoning method, apparatus, computer equipment, and storage medium provided in this application are specifically described through the following embodiments. First, the information reasoning method in this application is described.

[0027] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0028] Foundational technologies for artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0029] The information reasoning method provided in this application relates to the field of artificial intelligence technology. The information reasoning method provided in this application can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application implementing the information reasoning method, but is not limited to the above forms.

[0030] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer computer devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0031] Figure 1 This is an optional flowchart of the information reasoning method provided in the embodiments of this application. Figure 1 The method may include, but is not limited to, steps S101 to S106.

[0032] Step S101: Obtain the reference configuration parameters and reference key-value blocks of the reference semantic fingerprint information; wherein, the reference configuration parameters include: reference application scenario category, reference user level and reference input frequency; Step S102: Prioritize the reference semantic fingerprint information according to the reference application scenario category, reference user level, and reference input frequency to obtain the reference priority level; Step S103: Store the reference key-value block to the preset target cache queue according to the reference priority level; Step S104: Receive the current semantic fingerprint information; Step S105: Extract the target key-value block from the target cache queue based on the current semantic fingerprint information and the reference semantic fingerprint information; Step S106: Perform information reasoning based on the target key value block.

[0033] Steps S101 to S106 of this embodiment prioritize and store the semantic fingerprint information in the target cache queue based on the reference application scenario category, reference user level, and reference input frequency. First, the hierarchical caching mechanism based on multi-dimensional priorities effectively avoids the problem of hot and cold data crowding each other in a unified cache queue, significantly improving the overall cache hit rate. Second, upon receiving the current semantic fingerprint information, the target key-value blocks stored according to priority are directly extracted from the target cache queue based on the matching relationship between the current semantic fingerprint information and the reference semantic fingerprint information, and inference is performed accordingly. Since high-priority, high-frequency reused key-value blocks are likely located at the head of the cache queue or in the active area, the matching and extraction speed is faster, avoiding the overhead of repeatedly calculating the KV cache and shortening the time for extracting target key-value blocks, thus improving information inference efficiency.

[0034] In step S101 of some embodiments, the reference semantic fingerprint information is stored in a vector retrieval library. It can be extracted directly from the vector retrieval library, or it can be obtained first and then stored in the vector retrieval library, and is not limited thereto.

[0035] In one application scenario, using the intelligent medical diagnostic assistant scenario, Weili collects reference semantic fingerprint information related to medical diagnosis in advance. Specifically, the reference semantic fingerprint information is semantic fingerprint information generated from at least one of the text, voice or images uploaded by users on the medical diagnostic platform. Specifically, it is disease inquiry information, casual conversation information and disease response information represented by at least one of the images, text and voice.

[0036] As disclosed above, the reference semantic fingerprint information is a multimodal semantic fingerprint information. The semantic fingerprint information represents the input content of any modality, which is transformed into a fixed-length vector or hash code that can represent the semantic features of the input content through a deep learning model.

[0037] Please see Figure 2 In some embodiments, step S101 may include, but is not limited to, steps S201 to S205: Step S201: Obtain multimodal prompt data and reference configuration parameters for the multimodal prompt data; wherein, the multimodal prompt data includes: text prompt information, image prompt information, and voice prompt information; Step S202: Perform semantic fingerprint conversion on the text prompt information to obtain text fingerprint information; Step S203: Perform semantic fingerprint conversion on the image prompt information to obtain image fingerprint information; Step S204: Perform semantic fingerprint conversion on the voice prompt information to obtain voice fingerprint information; Step S205: The text fingerprint information, image fingerprint information and voice fingerprint information are concatenated according to the reference application scenario category to obtain the reference semantic fingerprint information.

[0038] In step S201 of some embodiments, the multimodal prompt data consists of prompt information of at least one modality, which may include at least one of text prompt information, image prompt information, and voice prompt information. Taking an intelligent medical diagnostic assistant as an example, the text prompt information is the consultation text entered by the user in the intelligent medical diagnostic assistant, the image prompt information is the image or video related to disease symptoms uploaded by the user in the intelligent medical diagnostic assistant, and the voice prompt information is the voice information about medical descriptions uploaded by the user in the intelligent medical diagnostic assistant.

[0039] In step S202 of some embodiments, this embodiment uses a large language model to extract text semantic vectors from text prompts, and the text semantic vectors have a dimension of 768. The text semantic vectors are then combined with preset keyword weights to generate text fingerprint information.

[0040] In step S203 of some embodiments, this embodiment uses a large language model to extract image visual features from image prompt information, and the dimension of the image visual features is 2048. Then, image fingerprint information is generated by combining the preset target detection box coordinates and image visual features.

[0041] In step S204 of some embodiments, this embodiment uses a large language model to extract speech semantic vectors from audio prompt information, and combines preset speech emotion tags and speech semantic vectors to generate speech fingerprint information.

[0042] In step S205 of some embodiments, the text fingerprint information, image fingerprint information, and voice fingerprint information are concatenated according to the application scenario category. Specifically, the text fingerprint information, image fingerprint information, and voice fingerprint information are weighted and fused according to the application scenario category to generate unified reference semantic fingerprint information. It should be noted that the generated reference semantic fingerprint information is uniformly stored in a vector retrieval library for subsequent semantic fingerprint information queries.

[0043] In steps S201 to S204 of this embodiment, text prompts are converted into text fingerprints, image prompts are converted into image fingerprints, and voice prompts are converted into voice fingerprints. Then, the text fingerprints, image fingerprints, and voice fingerprints are concatenated into reference semantic fingerprints according to the reference application scenario category to construct multimodal reference semantic fingerprints. This enables similarity retrieval of text, image, and voice prompts, avoids redundant reasoning, and improves application speed.

[0044] Please see Figure 3 In some embodiments, step S205 may include, but is not limited to, steps S301 to S302: Step S301: Select a set of fingerprint weights from a preset set of candidate fingerprint weights according to the reference application scenario category; wherein, the selected set of fingerprint weights includes: a first weight, a second weight, and a third weight; Step S302: The text fingerprint information, image fingerprint information and voice fingerprint information are concatenated according to the first weight, the second weight and the third weight to obtain the reference semantic fingerprint information.

[0045] In step S301 of some embodiments, the importance of images, voice, and text varies in different scenarios. Specifically, in medical scenarios, images are more important than voice and text; in financial scenarios, the importance of images, voice, and text is similar; and in government public platform consultation scenarios, the importance of text and voice is higher than images. Therefore, this embodiment extracts a selected fingerprint weight set from the candidate fingerprint weight set according to different reference application scenario categories, and the selected fingerprint weight set includes: a first weight, a second weight, and a third weight. It should be noted that the first weight represents the importance of text fingerprint information, the second weight represents the importance of image fingerprint information, and the third weight represents the importance of voice fingerprint information. In another embodiment, dynamic features are extracted from text fingerprint information, image fingerprint information, and voice fingerprint information according to the reference application scenario category, and the first weight, second weight, and third weight are obtained by weighting the dynamic features (such as medical, autonomous driving, entertainment, etc.) through an attention mechanism.

[0046] In step S302 of some embodiments, this embodiment uses an attention mechanism, a first weight, a second weight, and a third weight to weight and fuse text fingerprint information, image fingerprint information, and voice fingerprint information to generate unified reference semantic fingerprint information.

[0047] For example, if the reference application scenario information is a medical scenario, the first weight is determined to be 0.3, the second weight to be 0.6, and the third weight to be 0.1. The text fingerprint information, image fingerprint information and voice fingerprint information are fused according to 0.3, 0.6 and 0.1, and a dense vector of fixed length is output as the final semantic fingerprint of the multimodal input, which is used for subsequent tasks such as cache matching and similarity retrieval.

[0048] In steps S301 to S302 of this embodiment, multimodal text fingerprint information, image fingerprint information and voice fingerprint information are fused into unified and multimodal reference semantic fingerprint information. The reference semantic fingerprint information can be used as a subsequent fingerprint information retrieval and matching, which facilitates the rapid retrieval of semantically matching key value blocks from subsequent inputs of multiple modalities, improves the extraction speed of key value blocks, and thus improves the efficiency of information reasoning.

[0049] Please see Figure 4 In some embodiments, step S102 may include, but is not limited to, steps S401 to S404: Step S401: Select the first level from the preset candidate levels according to the reference application scenario category; Step S402: Select the second level from the candidate levels based on the reference user level; Step S403: Select the third level from the candidate levels based on the reference input frequency; Step S404: The first level, the second level, and the third level are concatenated according to the preset level weight set to obtain the reference priority level.

[0050] In step S401 of some embodiments, the candidate level represents the priority level of the candidate, and the priority level is different for different application scenarios. In this embodiment, a first mapping relationship table between each application scenario category and the candidate level is set in advance, and the first level is directly looked up from the first mapping relationship table according to the reference application scenario category. For example, in this embodiment, the first level for application scenario categories such as medical, financial, and industrial is set as Level 1, and the second level for application scenario categories such as casual chat or entertainment is set as Level 2.

[0051] In step S402 of some embodiments, different priority levels are configured for different user levels, thus determining the second level matching the user level. In this embodiment, user levels include paid users and ordinary users. If the user level is a paid user, the second level is determined to be level one; if the user level is an ordinary user, the second level is determined to be level two.

[0052] In step S403 of some embodiments, this embodiment sets three input frequency ranges, namely a first frequency range, a second frequency range, and a third frequency range, and the three input frequency ranges decrease in number sequentially. If the reference input frequency is located in the first frequency range, the third level is determined to be level one; if the reference input frequency is located in the second frequency range, the third level is determined to be level two; if the reference input frequency is located in the third frequency range, the third level is determined to be level three.

[0053] In step S404 of some embodiments, the level weight set includes a fourth weight, a fifth weight, and a sixth weight. The fourth weight represents the importance of the scene category, the fifth weight represents the importance of the user level, and the sixth weight represents the importance of the input frequency. Specifically, the first level is converted into a first score, the second level into a second score, and the third level into a third score. Then, the first score, the second score, and the third score are weighted and summed according to the fourth weight, the fifth weight, and the sixth weight to obtain the target score. The reference priority level is determined by comparing the target score with the score range corresponding to each level.

[0054] In steps S401 to S404 of this embodiment, the reference application scenario category, reference user level, and the level corresponding to the parameter input frequency are identified respectively. Then, the weighted sum of each level is used to determine the reference priority level, making the determination of the reference priority level more accurate and the representation of the priority of the reference key value block more accurate.

[0055] In step S103 of some embodiments, this embodiment sets up a preset candidate cache queue, selects a target cache queue from the candidate cache queues according to the reference priority level, and then stores the reference key-value block into the target cache queue. Specifically, this embodiment sets up three candidate cache queues, defined as the first cache queue, the second cache queue, and the third cache queue. It should be noted that the cache configuration parameters of each candidate cache queue are different, and the cache configuration parameters include cache duration and cache frequency. The cache duration represents the duration for which the reference key-value block can be cached. When the cache duration is exceeded, the reference key-value block is removed from the target cache queue. The cache frequency represents the frequency at which the reference key-value block is cached. The higher the frequency, the shorter the cache time. In this embodiment, the cache duration of the first cache queue is greater than the cache duration of the second cache queue, and the cache duration of the second cache queue is greater than the cache duration of the third cache queue; the cache frequency of the first cache queue is higher than the cache frequency of the second cache queue, and the cache frequency of the second cache queue is higher than the cache frequency of the third cache queue. If the reference priority level is level one, the target cache queue is the first cache queue; if the reference priority level is level two, the target cache queue is the second cache queue; if the reference priority level is level three, the target cache queue is the third cache queue.

[0056] In some embodiments, the caching duration of the candidate cache queues is different, and the eviction strategy for each candidate cache queue is different. Furthermore, the eviction strategy can also evict reference key-value blocks based on the similarity between the scene features of the reference key-value blocks. For example, in this embodiment, the caching duration of the first cache queue is set to 24 hours, the second cache queue to 6 hours, and the third cache queue to 1 hour. Therefore, the higher the reference priority, the longer the caching duration, which facilitates the direct reuse of reference key-value blocks in subsequent information inference without the need to regenerate them, thus improving the efficiency of information inference. It should be noted that this embodiment can also set a similarity threshold for each candidate cache queue. The current similarity is calculated for the scene features of each reference key-value block. If the current similarity is greater than the similarity threshold, one of the reference key-value blocks is removed. For example, in the first cache queue, in a medical cold scenario, if the similarity of cold symptoms in spring and summer is determined to be greater than the similarity threshold, and the current season is summer, then the reference key-value blocks for spring in the first cache queue can be cleared, saving cache space in the first cache queue.

[0057] In step S104 of some embodiments, after the newly generated reference key value block is cached in advance, new current semantic fingerprint information is received. The current semantic fingerprint information can be generated from at least one of the following: current text information, current image information, and current voice information. It should be noted that the process of generating the current semantic fingerprint information is the same as that of generating the reference semantic fingerprint information, and will not be repeated here.

[0058] In step S105 of some embodiments, as disclosed above, the reference semantic fingerprint information is stored in a vector retrieval library. The vector retrieval library stores and indexes the reference semantic fingerprint information, supporting similarity retrieval. In other words, the vector retrieval library serves as a database for querying key-value blocks of the current semantic fingerprint information, determining whether a key-value block matching the current semantic fingerprint information exists.

[0059] Please see Figure 5 In some embodiments, step S105 may include, but is not limited to, steps S501 to S505: Step S501: Calculate the similarity between the current semantic fingerprint information and the reference semantic fingerprint information in the vector retrieval library to obtain the fingerprint similarity. Step S502: Obtain the current configuration parameters of the current semantic fingerprint information; Step S503: Prioritize the current semantic fingerprint information according to the current configuration parameters to obtain the current priority level; Step S504: The current semantic fingerprint information is retrieved and evaluated based on the current priority level and fingerprint similarity to obtain retrieval evaluation information; wherein, the retrieval evaluation information represents whether or not reference semantic fingerprint information matching the current semantic fingerprint information exists in the vector retrieval library. Step S505: If the retrieval evaluation information representation vector retrieval library contains reference semantic fingerprint information that matches the current semantic fingerprint information, extract the target key-value block from the target cache queue according to the current priority level and the current semantic fingerprint information.

[0060] In step S501 of some embodiments, the current semantic fingerprint information is converted into a current semantic fingerprint vector, and each reference semantic fingerprint information in the vector retrieval library is also represented by a vector, also called a reference semantic fingerprint vector. In this embodiment, the fingerprint similarity is obtained by performing similarity calculation on the current semantic fingerprint vector and the reference semantic fingerprint vector, and the similarity calculation method can be cosine similarity calculation, Euclidean distance, dot product, etc. This embodiment does not limit the similarity calculation method.

[0061] In steps S502 and S503 of some embodiments, the current configuration parameters include the current application scenario category, the current user level, and the current input frequency. The current semantic fingerprint information is prioritized according to the current application scenario category, the current user level, and the current input frequency to obtain the current priority level. The current priority level is determined in the same way as the reference priority level, which will not be described again here.

[0062] In step S504 of some embodiments, the matching criteria of the current semantic fingerprint information in the vector retrieval library are jointly evaluated using the current priority level and fingerprint similarity to determine that the retrieval evaluation information is more accurate. Simultaneously, finding the target key value block of the current semantic fingerprint information based on the retrieval evaluation information is more accurate, facilitating information inference based on the target key value block to conform to the current semantic fingerprint information.

[0063] In step S505 of some embodiments, if the retrieval evaluation information characterization vector retrieval library contains reference semantic fingerprint information that matches the current semantic fingerprint information, it is equivalent to hitting the reference key value block that matches the current semantic fingerprint information in the vector retrieval library. The vector retrieval library also stores the location information of each key value block that matches the reference semantic fingerprint information. The selected cache queue is selected from the target cache queue according to the key value block location information, and the target key value block is extracted from the selected cache queue according to the current priority level and the current semantic fingerprint information.

[0064] In steps S501 to S505 of this embodiment, firstly, similarity calculation ensures the semantic accuracy of cache matching, avoiding the omission of semantically equivalent fingerprint information due to differences in input expression; secondly, only after the retrieval evaluation confirms that a matching reference fingerprint exists in the vector retrieval library, the target key-value block is extracted from the target cache queue according to the current priority, effectively avoiding the invalid preemption of high-priority cache resources by low-priority requests, and suppressing cache misuse caused by fingerprint similarity but insufficient priority, thus improving the overall hit accuracy, service differentiation guarantee capability, and resource utilization efficiency of the cache system.

[0065] Please see Figure 6 In some embodiments, step S504 includes, but is not limited to, steps S601 to S603: Step S601: Select the target similarity threshold from the preset candidate similarity thresholds according to the current priority level; Step S602: Compare the fingerprint similarity with the target similarity threshold to obtain the comparison result; Step S603: Determine the retrieval evaluation information based on the comparison results.

[0066] In step S601 of some embodiments, in order to improve the accuracy of information reasoning, the extracted target key value block can be adapted to the current semantic fingerprint information, and different target similarity thresholds are set for different current priority levels. For example, if the current priority level is level one, the target similarity threshold is determined to be 0.9; if the current priority level is level two, the target similarity threshold is determined to be 0.8; if the current priority level is level three, the target similarity threshold is determined to be 0.7.

[0067] In steps S602 and S603 of some embodiments, if the fingerprint similarity is greater than or equal to the target similarity threshold, it is determined that the retrieval evaluation information represents reference semantic fingerprint information that matches the current semantic fingerprint information in the vector retrieval library; if the fingerprint similarity is less than the target similarity threshold, it is determined that the retrieval evaluation information represents reference semantic fingerprint information that does not match the current semantic fingerprint information in the vector retrieval library.

[0068] In some embodiments, if the retrieval evaluation information representation vector retrieval library does not contain reference semantic fingerprint information that matches the current semantic fingerprint information, the large language model is invoked to infer the current semantic fingerprint information and generate the current key value block, and the information inference is completed based on the current key value block.

[0069] In steps S601 to S603 of this embodiment, different target similarity thresholds are set according to different priority levels. This allows for the push of more accurate target key value blocks for the current semantic fingerprint information of the importance level, thereby realizing hierarchical information reasoning.

[0070] Please see Figure 7 In some embodiments, after step S504, the information reasoning method may also include, but is not limited to, steps S701 to S703: Step S701: Collect retrieval evaluation information according to a preset time period, representing the ratio of reference semantic fingerprint information that matches the current semantic fingerprint information in the vector retrieval database, and obtain the hit ratio. Step S702: Adjust the target similarity threshold according to the hit ratio to obtain the updated similarity threshold; Step S703: Update the retrieval evaluation information according to the updated similarity threshold.

[0071] In steps S701 and S702 of some embodiments, the hit ratio also represents the reuse status of the currently cached reference key-value blocks. A higher hit ratio indicates a higher key-value block reuse rate, and a lower hit ratio indicates a lower key-value block reuse rate. Therefore, in order to improve the key-value block reuse rate, improve the efficiency of information reasoning, and reduce the time spent on information reasoning, this embodiment dynamically adjusts the target similarity threshold based on the hit ratio. Specifically, if the hit ratio is lower than a preset ratio, the target similarity threshold is lowered to increase the hit rate of the reference semantic fingerprint information. If the hit ratio is higher than the preset ratio, the target similarity threshold is not adjusted.

[0072] For example, with a preset ratio of 85%, the real-time hit rate of each target cache queue (such as the first cache queue) is calculated every hour. When the hit rate of the first cache queue is lower than 85%, the target similarity threshold is automatically reduced from 0.9 to 0.85, thereby relaxing the matching conditions and enabling more semantically similar but slightly different multimodal fingerprints to successfully hit the cached key-value blocks.

[0073] In steps S701 to S702 of this embodiment, closed-loop feedback and adaptive parameter adjustment can respond in real time to the hit rate fluctuations of different target cache queues, prioritizing the cache hit rate of high-priority requests (such as medical diagnosis and high-level users) to remain stable above the target level, avoiding frequent cache misses in critical businesses due to overly strict fixed thresholds; at the same time, the threshold reduction operation is only dynamically triggered for queues with insufficient hit rates and does not affect the matching accuracy of other cache queues.

[0074] In step S106 of some embodiments, the target key-value block can be used as primary information for information reasoning or as auxiliary information. Specifically, if the similarity between the current semantic fingerprint information and the reference semantic fingerprint information is 100%, information reasoning is completed directly based on the target key-value block; if the similarity between the current semantic fingerprint information and the reference semantic fingerprint information is less than 100%, information reasoning is completed by combining the target key-value block and the current semantic fingerprint information.

[0075] Reference Figure 8This application embodiment takes an intelligent medical diagnostic assistant as an example. It pre-collects multimodal prompt data input from the doctor's end, specifically including text prompts, image prompts, and voice prompts. The text prompt is: the patient complains of "fever and cough for three days"; the image prompt is: a chest X-ray; and the voice prompt is: the patient reports "chest tightness and night sweats." The system extracts feature vectors from the text, image, and voice prompts respectively, and outputs a unified 512-dimensional reference semantic fingerprint through a weighted fusion using an attention mechanism. Further, based on the reference configuration parameters of the multimodal prompt data, if the reference configuration parameters are: reference application scenario is medical diagnosis, reference user level is paid user level, and reference input frequency is high frequency, the reference priority level is determined to be level one, equivalent to the highest level. Then, the reference key-value blocks generated before the multimodal prompt data are stored in the first cache queue. It should be noted that the reference key-value blocks record the data of the multimodal prompt information completing the information reasoning process and the disease response content. When a patient asks a question and uploads an image on the intelligent medical diagnostic assistant, with the question being "fever and cough for three days" and the image being a "chest X-ray," the reference semantic fingerprint information of the current question content is matched in the vector retrieval library, and the current priority level is determined to be level one. Therefore, the matching reference key-value block is directly extracted from the first cache queue and information reasoning is completed. This information reasoning can supplement the disease response content with the difference in the "chest X-ray" image. Furthermore, if the patient asks a casual question on the intelligent medical diagnostic assistant, although the vector retrieval library is matched, the priority level of the matching reference key-value block is level three. It is necessary to wait for the key-value blocks in the first and second cache queues to be extracted before extracting the target key-value block from the third cache queue to complete information reasoning. Therefore, this embodiment constructs a key-value block caching technology to achieve key-value block reuse, pre-compile multimodal fingerprint information caching, and set cache priorities for different scenario categories, user levels, and input frequencies, and complete the key-value block extraction order according to the cache priority. This embodiment not only reduces redundant KV cache calculations, lowering GPU computing power costs by approximately 50%, but also offers faster response times and improved diagnostic efficiency. Furthermore, the multimodal reference semantic fingerprint information setting allows for rapid retrieval of matching key-value blocks even when only one modality's hint information is subsequently uploaded, enabling cross-modal reuse. Moreover, high-priority requests can be quickly and stably fed back to the KV cache, while low-priority requests do not consume the cache resources of high-priority requests, adapting to more advanced information reasoning scenarios.

[0076] Please see Figure 9 This application also provides an information reasoning apparatus that can implement the above-described information reasoning method. The apparatus includes: The acquisition module 901 is used to acquire reference configuration parameters and reference key value blocks of reference semantic fingerprint information; wherein, the reference configuration parameters include: reference application scenario category, reference user level and reference input frequency; The priority division module 902 is used to prioritize the reference semantic fingerprint information according to the reference application scenario category, reference user level and reference input frequency to obtain the reference priority level. Storage module 903 is used to store reference key-value blocks to a preset target cache queue according to reference priority level; The receiving module 904 is used to receive the current semantic fingerprint information; The key-value block extraction module 905 is used to extract the target key-value block from the target cache queue based on the current semantic fingerprint information and the reference semantic fingerprint information; Information reasoning module 906 is used to perform information reasoning based on the target key value block.

[0077] The specific implementation of this information reasoning device is basically the same as the specific implementation of the information reasoning method described above, and will not be repeated here.

[0078] This application also provides a computer device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described information reasoning method. This computer device can include any smart terminal such as a tablet computer or an in-vehicle computer.

[0079] Please see Figure 10 , Figure 10 The hardware structure of a computer device according to another embodiment is illustrated. The computer device includes: The processor 1001 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 1002 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 1002 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1002 and is called and executed by the processor 1001 using the information reasoning method of the embodiments of this application. Input / output interface 1003 is used to implement information input and output; The communication interface 1004 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, Wi-Fi, Bluetooth, etc.). Bus 1005 transmits information between various components of the device (e.g., processor 1001, memory 1002, input / output interface 1003, and communication interface 1004); The processor 1001, memory 1002, input / output interface 1003 and communication interface 1004 are connected to each other within the device via bus 1005.

[0080] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described information reasoning method.

[0081] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0082] The information reasoning method, apparatus, computer device, and storage medium provided in this application obtain configuration parameters in three dimensions—reference application scenario category, reference user level, and reference input frequency—of the reference semantic fingerprint information, and prioritize the reference key-value blocks, storing them in a target cache queue according to the assigned priority level. Compared to the "unified cache queue" method used in the prior art, this application allows high-priority key-value blocks to be stored in the cache first, avoiding being flushed by low-priority key-value blocks and increasing the probability of cache hits. Furthermore, after receiving the current semantic fingerprint information, based on the matching relationship between the current semantic fingerprint information and the reference semantic fingerprint information, the target key-value blocks stored according to priority are directly extracted from the target cache queue, and reasoning is performed accordingly. Because high-priority, frequently reused key-value blocks are cached at the head of the queue or in the active area, and are pre-cached, the target key-value blocks can be quickly extracted, avoiding the overhead of repeatedly calculating key-value blocks, shortening the target key-value block extraction time, reducing information reasoning latency, and improving the efficiency of information reasoning.

[0083] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0084] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0085] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0086] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0087] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0088] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0089] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0090] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0091] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0092] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0093] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. An information reasoning method, characterized in that, The method includes: Obtain reference configuration parameters and reference key-value blocks for reference semantic fingerprint information; wherein, the reference configuration parameters include: reference application scenario category, reference user level, and reference input frequency; The reference semantic fingerprint information is prioritized according to the reference application scenario category, the reference user level, and the reference input frequency to obtain the reference priority level; The reference key-value block is stored in a preset target cache queue according to the reference priority level; Receive the current semantic fingerprint information; Extract the target key-value block from the target cache queue based on the current semantic fingerprint information and the reference semantic fingerprint information; Information reasoning is performed based on the target key value block.

2. The method according to claim 1, characterized in that, The reference semantic fingerprint information is multimodal semantic fingerprint information, and the reference configuration parameters for obtaining the reference semantic fingerprint information include: Acquire multimodal prompt data and the reference configuration parameters of the multimodal prompt data; wherein the multimodal prompt data includes: text prompt information, image prompt information, and voice prompt information; The text prompt information is semantically fingerprinted to obtain text fingerprint information; The image prompt information is semantically fingerprinted to obtain image fingerprint information; The voice prompt information is semantically fingerprinted to obtain voice fingerprint information; The text fingerprint information, the image fingerprint information, and the voice fingerprint information are concatenated according to the reference application scenario category to obtain the reference semantic fingerprint information.

3. The method according to claim 2, characterized in that, The step of concatenating the text fingerprint information, the image fingerprint information, and the voice fingerprint information according to the reference application scenario category to obtain the reference semantic fingerprint information includes: A selected fingerprint weight set is selected from a preset candidate fingerprint weight set according to the reference application scenario category; wherein, the selected fingerprint weight set includes: a first weight, a second weight, and a third weight; The text fingerprint information, the image fingerprint information, and the voice fingerprint information are concatenated according to the first weight, the second weight, and the third weight to obtain the reference semantic fingerprint information.

4. The method according to claim 1, characterized in that, The step of prioritizing the reference semantic fingerprint information based on the reference application scenario category, the reference user level, and the reference input frequency to obtain a reference priority level includes: The first level is selected from the preset candidate levels based on the reference application scenario category; A second level is selected from the candidate levels based on the reference user level; A third level is selected from the candidate levels based on the reference input frequency; The first level, the second level, and the third level are concatenated according to a preset set of level weights to obtain the reference priority level.

5. The method according to any one of claims 1 to 4, characterized in that, The reference semantic fingerprint information is stored in a vector retrieval library; The step of extracting the target key-value block from the target cache queue based on the current semantic fingerprint information and the reference semantic fingerprint information includes: Fingerprint similarity is obtained by calculating the similarity between the current semantic fingerprint information and the reference semantic fingerprint information in the vector retrieval library; Obtain the current configuration parameters of the current semantic fingerprint information; The current semantic fingerprint information is prioritized according to the current configuration parameters to obtain the current priority level; The current semantic fingerprint information is retrieved and evaluated based on the current priority level and the fingerprint similarity to obtain retrieval evaluation information; wherein, the retrieval evaluation information indicates whether the vector retrieval library contains or does not contain the reference semantic fingerprint information that matches the current semantic fingerprint information; If the retrieval evaluation information indicates that the vector retrieval library contains reference semantic fingerprint information that matches the current semantic fingerprint information, the target key-value block is extracted from the target cache queue according to the current priority level and the current semantic fingerprint information.

6. The method according to claim 5, characterized in that, The step of retrieving and evaluating the current semantic fingerprint information based on the current priority level and the fingerprint similarity to obtain retrieval evaluation information includes: The target similarity threshold is selected from the preset candidate similarity thresholds based on the current priority level; The fingerprint similarity is compared with the target similarity threshold to obtain the comparison result; The retrieval evaluation information is determined based on the comparison results.

7. The method according to claim 6, characterized in that, After performing a retrieval evaluation on the current semantic fingerprint information based on the current priority level and the fingerprint similarity to obtain retrieval evaluation information, the method further includes: The hit ratio is obtained by collecting the retrieval evaluation information according to a preset time period, which represents the ratio of the reference semantic fingerprint information that matches the current semantic fingerprint information in the vector retrieval store. The target similarity threshold is adjusted based on the hit ratio to obtain an updated similarity threshold; The retrieval evaluation information is updated based on the updated similarity threshold.

8. An information reasoning device, characterized in that, The device includes: The acquisition module is used to acquire reference configuration parameters and reference key-value blocks for reference semantic fingerprint information; wherein, the reference configuration parameters include: reference application scenario category, reference user level, and reference input frequency; The priority division module is used to divide the reference semantic fingerprint information into priorities based on the reference application scenario category, the reference user level, and the reference input frequency to obtain a reference priority level. The storage module is used to store the reference key-value block into a preset target cache queue according to the reference priority level; The receiving module is used to receive the current semantic fingerprint information; The key-value block extraction module is used to extract target key-value blocks from the target cache queue based on the current semantic fingerprint information and the reference semantic fingerprint information; The information reasoning module is used to perform information reasoning based on the target key value block.

9. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the information reasoning method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the information reasoning method according to any one of claims 1 to 7.

Images