7 results about "Online search" patented technology

This invention discloses an online search method for particle spelling variants and properties usable by intelligent agents. The steps are as follows: inputting the baseline spelling of a particle and the constructed particle-specific prompt words into a large language model to generate a core variant corresponding to the particle; integrating the core variant of each particle with the authoritative benchmark information corresponding to the particle to form a full variant set for each particle; each record in the full variant set contains a unique identifier for a particle, a multi-scene benchmark name, a core variant, and extended attributes; dynamically mapping a remotely callable interface set on the server to a local proxy object on the client; after receiving a query request, the local proxy object parses the particle spelling parameters in the query request and performs multi-dimensional spelling matching in the full variant set to realize online search for particle spelling variants and properties. This invention effectively improves the ability of AI intelligent agents to handle domain-specific tasks.

This invention discloses an automated tuning method and system for large language model service configurations, belonging to the fields of artificial intelligence and cloud computing technology. The invention includes an offline analysis module for performance profiling of the large language model (LLM) service system under a finite configuration set and for collecting training data; an online search module for exploring and evaluating the performance of different configurations in the configuration space based on a Monte Carlo tree search algorithm; and a simulator module for receiving the configurations generated by the online search module and predicting the performance metrics of the configurations under a target workload based on a logical step simulation model. This invention effectively solves the problems of low search efficiency and poor performance of existing automatic tuning methods in existing large language model service frameworks (such as vLLM, SGLang, etc.) due to the huge number of configuration parameters (reaching a petabyte-scale search space) and the complex dependencies and mutual exclusion relationships between parameters.

The application provides a training method of a search relevance model, a relevance evaluation method and related devices. The method comprises: inputting a sample into a plurality of teacher models to obtain a basic relevance score output by each teacher model; inputting the sample, a search term type and a teacher model scoring confidence into a gating model to dynamically obtain the weight of different teacher models; weighting and fusing the basic relevance scores output by the plurality of teacher models based on the weight output by the gating model to obtain a target relevance score; inputting the sample into a student model to train the student model with the target relevance score as a target; and adding an entity extraction auxiliary task to the teacher model and extracting key attributes by using an external knowledge graph, and improving the attention weight of the key attributes through a post-interaction link when training the student model. The student model serves as an online search relevance model. The application can improve the relevance discrimination ability of search terms and commodities in a search system to a certain extent.

Embodiments of the present application provide an end-side model inference dynamic optimization method, device, equipment and medium. The method surrounds the end-side single-flow CPU inference path, establishes a performance, energy consumption and thermal state measurement system covering the prompt processing stage and the self-recurrence generation stage, and adopts a fixed word window unified index statistics and configuration decision. The offline measurement is performed on the CPU frequency, thread number and affinity configuration space, the feasible boundary and preferred priori are refined, and a discrete neighborhood search dynamic optimization method under the offline priori constraint is proposed. Thus, taking a unified comprehensive cost function as an optimization target, the online search is limited in the feasible neighborhood obtained through offline screening under the first word delay, thermal safety and frequency accessibility constraint, and combined with the minimum residence, anti-shake, limited memory and controlled jump-out mechanism, the low-overhead joint adjustment of the CPU frequency and thread number is realized, while the adaptability to state changes is improved.

This invention discloses a snow ablation-optimized weak-sensitivity Kalman filter method and system for UAV state estimation, belonging to the field of UAV autonomous navigation and state estimation technology. The method includes: obtaining the posterior state estimate, posterior covariance matrix, and posterior sensitivity matrix of the previous time step; obtaining the prior statistics of the current time step based on ensemble Kalman filtering with ensemble point sampling and prediction propagation; searching online for the optimal scaling factor using the snow ablation optimization algorithm and scaling the preset sensitivity weight matrix; calculating the Kalman gain based on the weighted prior sensitivity matrix; updating the posterior state estimate using measurement data and outputting the flight state estimation result. This invention, while retaining the advantages of fast analytical calculation of the weak-sensitivity ensemble Kalman filter gain, achieves online adaptive adjustment of the sensitivity weights, improving the accuracy and robustness of state estimation in dynamic flight environments, and has low computational overhead, making it suitable for UAV airborne navigation systems.

The application relates to an artificial intelligence-based safety production danger prediction method and management system, and belongs to the safety monitoring and early warning field. The method comprises the following steps: acquiring multi-source heterogeneous data through an edge sampling module; inputting the data into a self-supervised prediction engine, generating predicted observation values of a future period by using a generative adversarial network model, identifying abnormal hidden dangers by point-by-point residual error calculation and dynamic threshold determination; when the abnormal hidden dangers are identified, triggering a reinforcement learning control module to search for an optimal risk avoidance control instruction sequence online under preset safety constraints; and based on a hard real-time kernel and zero-copy technology, the edge execution interface delivers the instruction to a field executor at a low time delay to execute a risk avoidance operation. The system realizes millisecond-level response and zero-sample abnormality identification, builds a closed-loop risk management system integrating perception and decision-making, has the functions of offline autonomous operation and equipment health degree evaluation, and improves the real-time performance, accuracy and intelligent level of safety production monitoring in complex industrial scenes.