221 results about "Nearest neighbor search" patented technology

The invention discloses a document analysis method based on an RAG model and a dynamic knowledge graph, and relates to the technical field of artificial intelligence. The method is combined with an RAG model and a dynamic knowledge graph technology, and is realized by the following steps of: performing entity relationship joint extraction on an input document, generating a structural triple, and constructing a dynamically updatable knowledge graph; based on the knowledge graph, mapping entities and relationships into low-dimensional vectors by adopting a graph embedding model, and constructing a local vector knowledge base with a topological structure; receiving user questions in real time, encoding the user questions into query vectors, executing approximate nearest neighbor search based on the vector knowledge base, and matching related map fragments; and combining the retrieved graph fragments with the large language model, and generating a structured answer through path constraint of the injection knowledge graph. The method is used for solving the problem that in the prior art, a model cannot capture document deep semantics and dynamic relations insufficiently.

The invention discloses a few-sample visual defect detection method based on approximate proximity search, and belongs to the technical field of industrial intellectualization. The system mainly comprises an image acquisition module, a feature extraction module, a memory bank construction module, a feature compression module, an approximate nearest neighbor search module and an anomaly detection and result output module. The feature extraction module is used for performing feature extraction on the collected image; the feature compression module is used for reducing the feature data volume and reducing the internal memory and calculation burden; the memory bank construction module is mainly operated during system initialization or model training and is used for constructing a feature memory bank of normal samples; the approximate nearest neighbor search module is responsible for quickly finding normal sample features closest to the input features in the memory bank; and the anomaly detection and result output module is used for converting distance information obtained by neighbor search into anomaly. According to the few-sample visual defect detection method based on the approximate proximity search, the effect of few-sample anomaly detection is achieved.

The invention provides a retrieval enhancement generation method and system for multi-modal semantic alignment in the technical field of computer vision and natural language processing. The method comprises the steps that S1, a large amount of text data, image data and 3D model data are acquired to construct a multi-modal vector knowledge base; s2, acquiring an input query image and a natural language question, and performing joint semantic coding on the query image and the natural language question to obtain a multi-modal query feature vector; s3, evaluating complexity, dynamically adjusting retrieval parameters based on the complexity, and retrieving Top-K candidate knowledge from the multi-modal vector knowledge base through an approximate nearest neighbor search method based on the retrieval parameters and the multi-modal query feature vectors; and S4, in combination with multi-dimensional similarity and token-level interaction alignment, screening each candidate knowledge, and outputting a retrieval result. The method has the advantages that the reasoning precision, the answer credibility and the response efficiency in a complex question and answer scene are greatly improved.

To accelerate search speeds for approximate nearest neighbor searches of vector databases, compute-in-memory techniques using NAND memory structures are introduced. For each element of the database, a kernel of its M nearest neighbors is determined. For each vector of the database, both the vector and its kernel are programmed in the arrays of a NAND memory based accelerator card, so that the vectors will be written into the memory arrays both as themselves and also in kernels of vectors for which they are a nearest neighbor. Metadata, associating the locations of the kernel members with the correspond vector is also stored in the memory system. After determining the input's nearest neighbor at one level of search, the metadata is then used to locate that nearest neighbor's nearest neighbors and their distances to the input vector are then computed in parallel in a compute-in-memory vector-vector dot product multiplication.

The invention discloses an obstacle avoidance motion planning method and system for redundant drive of an intelligent process of a welding robot and computer equipment. According to the method, firstly, a motion path of a robot is planned based on welding process redundancy, then a sampling strategy that the position and posture of a welding gun are separated is adopted, then the sampled posture is mapped to a planar two-dimensional space, and the relevance between an obstacle collision boundary and the obstacle avoidance posture of the welding gun is constructed; an incremental extreme gradient lifting model is adopted to carry out probability modeling on the collision boundary and obstacle avoidance attitude relevance, nearest neighbor search and new planning node expansion are carried out based on target cost to obtain new planning nodes, collision detection is carried out on the new planning nodes, and the new planning nodes passing the collision detection are used for incremental learning of an obstacle avoidance model; and finally, an optimization strategy of spherical linear interpolation and cosine slow motion time mapping is adopted to ensure continuity and smoothness of a motion planning path trace. The problem that in the prior art, it is difficult to quickly calculate and generate a collision-free welding track is solved.

The invention discloses an electric power multi-modal corpus construction query method and system based on a sliding window, which is applied to the field of electric power data query, and comprises the following steps: obtaining a structured document according to electric power multi-modal data, segmenting the structured document to obtain a plurality of segmented text blocks, and storing the segmented text blocks into a database; inputting each segmented text block into a large language model to generate a to-be-stored text vector and construct an electric power multi-mode corpus, when a user query request is received, generating a plurality of query variants according to query data, performing nearest neighbor search on each query variant in the electric power multi-mode corpus to obtain a corresponding nearest neighbor search result, and storing the nearest neighbor search result in the electric power multi-mode corpus. And fusing each nearest neighbor search result to generate an electric power related document set comprising a multi-modal association mark. According to the method, the semantic units can be accurately captured, semantic breakage is avoided, the retrieval continuity and coverage rate are improved, the comprehensiveness and context adaptability of retrieval results are improved, and the data retrieval requirement under the complex scene of the power industry is met.

The invention relates to the cross technical field of artificial intelligence and data retrieval, and particularly provides an RAG question and answer optimization method and device.Firstly, original knowledge documents of various formats and natural language query of a user are received on an input layer, then a knowledge base construction module and a knowledge retrieval module are arranged on a processing layer, and a knowledge retrieval module is arranged on the processing layer; the middleware layer is provided with an MCP Server module and a vector database, the MCP Server module is a standard interface agent, and the vector database is a distributed storage engine for efficient approximate nearest neighbor search and mixed retrieval; and finally, deploying an LLM generation module at an output layer for generating a final answer based on the retrieved enhanced context. Compared with the prior art, the method has the advantages that irrelevant or low-signal-to-noise-ratio knowledge contacted by a large language model (LLM) can be effectively reduced, and the accuracy, the reliability and the practicability of knowledge base questions and answers are comprehensively improved.

An online system applies a visual language model and an optical character recognition model to a received image to generate descriptive information about unknown items in the image. The online system prompts a generative model with the descriptive information about unknown items in the image to separate the descriptive information into different bins each corresponding to a different unknown item in the image. For each unknown item detected in the image, the online system generates a target embedding from its descriptive information and performs a nearest neighbor search on an item catalog including embeddings for various items to find a set of candidate embeddings matching the target embedding. The online system retrieves item attributes of candidate items each corresponding to a candidate embedding of the set and prompts the generative model with this information to rank candidate items for the unknown item in the image.

The invention discloses an automatic SCADE test case generation method based on a large language model, which comprises the following steps of: analyzing various SCADE related documents to obtain an analysis result; the analysis result is converted into JSON data; performing key information extraction on the JSON data, and performing block processing on the extracted key information to obtain text blocks; vectorizing the text blocks, storing the vectorized text blocks into a vector database, converting a natural language query requirement of a user into a query vector, and performing approximate nearest neighbor search with the query vector in the vector database by utilizing RAG retrieval to obtain a preliminary retrieval result; based on metadata in the preliminary retrieval result, filtering, timeliness sorting and deduplication fusion processing are carried out, and a retrieved context information set is obtained; adopting a dynamic cue word template to construct an enhanced cue word based on the retrieved context information set and the natural language query of the user; and calling a large language model API, inputting an enhanced cue word, and generating a structured test case.

The invention discloses an intelligent medical big data information management system, and the system comprises a data collection layer which serves as a sensing neural network, employs an Internet of Things gateway to be compatible with a medical equipment protocol, and deploys an edge calculation node for physiological signal preprocessing; the intelligent preprocessing layer is used for cleaning and standardizing the original data transmitted by the data acquisition layer; the informatization service layer is used for realizing fine-grained authority control based on an RBAC (Role Based Access Control) model, realizing inspection process automation by integrating an LIS (Line Information System) interface through a Hyperdger Fabric block chain evidence storage key operation hash, and constructing a zero-footprint DICOM (Digital Imaging and Communications in Medicine) viewer; the intelligent analysis layer is used for carrying out data life cycle management by adopting a mixed storage architecture, carrying out feature engineering on data, carrying out label matching and classified storage by using efficient neighbor search and similarity / distance measurement, quantifying contribution of each feature or feature combination to a single prediction result through an SHAP value, generating an interpretation result, and outputting the interpretation result; and presenting an interpretation result to a clinician or related personnel through a query interface.

The invention relates to the technical field of natural language processing, in particular to a retrieval system and method based on retrieval enhancement generation, and the method comprises the steps: obtaining a user query text through a user interaction interface, and carrying out the embedding processing of the user query text to obtain a query vector; inputting the query vector into a vector knowledge base, performing approximate nearest neighbor search on the query vector and a storage vector in the vector knowledge base, and inputting the storage vector meeting a preset semantic distance into a candidate text block sequence; analyzing the semantic association strength between each text block in the candidate text block sequence and the query vector; performing descending sorting on each text block in the candidate text block sequence according to the similarity score vector; and inputting the enhanced prompt text into the large language model for reasoning analysis to obtain a retrieval result. According to the method, the text semantic similarity and the context correlation can be considered at the same time in the retrieval process, and the accuracy of the retrieval result is improved through a dynamic weighting mechanism.

The invention discloses a GNSS (Global Navigation Satellite System) / INS (Inertial Navigation System) / map tightly integrated navigation positioning method and system based on road marking matching, and belongs to the technical field of navigation positioning, and the method comprises the following steps: acquiring priori road map data, foresight road image data, a GNSS original observation value and an INS original observation value; a semantic segmentation network is adopted to extract road marking semantic features in the foresight road image data, and a road marking feature instance is obtained; performing nearest neighbor search on the road marking feature instance and map elements in the prior road map data, and obtaining corresponding map matching observation based on a nearest neighbor search result; and coupling the GNSS original observation value, the INS original observation value and map matching observation to obtain a navigation positioning result. According to the method, a tight combination model based on GNSS / INS and map matching original observation values is provided, the positioning performance of a navigation system in a complex city scene is improved, and omni-directional pose constraint is provided for a vehicle.

The invention discloses an irregular roadway colored point cloud hole filling method, and belongs to the field of underground mine three-dimensional modeling and point cloud data processing. Firstly, dense slicing is carried out on roadway point clouds, slices are projected to a two-dimensional plane for polar coordinate sorting, disordered point clouds are ordered, then the slices serve as type value points, a control point solving matrix is solved through a chasing method, initial control points are rapidly obtained, the control points are optimized through a progressive iterative approximation algorithm (PIA), and the optimal control points are obtained. The method comprises the steps of effectively reducing errors between a fitting curve and an original point cloud, performing non-uniform cubic B-spline adaptive fitting interpolation according to a point spacing to realize uniform filling of holes, and finally performing K-nearest neighbor search on interpolation points, calculating an RGB weighted average value of neighborhood points and assigning the RGB weighted average value to realize color repair of a hole region. The method has a good repairing effect on various types of holes of the irregular roadway of the mine, uniform color assignment can be carried out on point clouds at the same time, and the modeling quality and efficiency are effectively improved.

The invention provides a three-dimensional part retrieval method based on graph similarity search, relates to the field of computer graphics, and solves the technical problems that in the prior art, geometric and design semantic information of CAD cannot be fully utilized, and a topological relation is difficult to capture and display, so that the retrieval efficiency is low. The method comprises the following steps: acquiring part three-dimensional data of a computer-aided design (CAD) model, and constructing a training data set; constructing an edge-surface connection diagram based on the three-dimensional data of the part; calculating a graph editing distance (GED) matrix of all edge surface connection graphs in the training data set as a supervision signal; based on the GED matrix, training a sorting model, mapping an edge-surface connection graph to a hidden space, and constructing a part vector database according to an output graph-level embedding vector; the sorting model is constructed based on a graph attention network; and inputting a to-be-queried CAD part into the trained sorting model to obtain a feature vector, carrying out nearest neighbor search in the part vector database, and returning a similar part result.

The invention discloses a carriage groove positioning and steel coil placement center determining method based on three-dimensional point clouds, and relates to the technical field of industrial automation, and the method comprises the steps: firstly carrying out the preprocessing of an original point cloud, and extracting a carriage main body point cloud through ground removal, outlier filtering and DBSCAN clustering; on the basis, the overall orientation of the carriage point cloud is determined through principal direction analysis, a central area point cloud is obtained through cutting in the principal direction, and after x and z coordinates of the central area point cloud are extracted, groove valley points are positioned through sliding window calculation, Gaussian smoothing and peak value detection. As the size of the compartment groove is generally larger than that of the steel coil, in order to guarantee the loading safety of the steel coil, sole timbers need to be arranged on the edge of the groove, so that the sudden drop point of the z value is detected forwards based on the valley point, and the edge position of the groove is determined. And calculating the center coordinate of the steel coil to be placed by combining the edge of the groove, the size of the sole timber and the radius of the steel coil, and finally mapping the two-dimensional detection point back to a three-dimensional space by using KD tree nearest neighbor search to obtain the three-dimensional coordinate of the steel coil placing center.

An extreme classification method includes receiving training data-points and classifier vectors associated with the training data-points. A plurality of training epochs are performed wherein each training epoch includes generating query embeddings for each data-point, sampling a predetermined number of negative labels from a set of negative labels for each of the training data-points; and training an encoder and the classifier vectors using the sampled negative labels. Positive labels and the sampled negative labels are then used to compute a loss. Encoder parameters and the classifier vectors are then updated based on the computed loss. For a first portion of epochs, the sampled negative labels include only uniformly random negative labels. For a second portion of the epochs, the sampled negative labels include uniformly random negative labels and hard negative labels. The hard negative labels are identified using an Approximate Nearest Neighbor Search (ANNS) index (308) built on the classifier vectors.

The invention provides a multi-source security knowledge fusion driven security risk management large model construction method and application. The method comprises the following steps: constructing a campus security knowledge graph and a semantic vector database based on multi-modal campus security data; performing approximate nearest neighbor search in the semantic vector database based on user query to obtain context evidence related to the user query, splicing the context evidence with the user query to obtain a retrieval enhancement prompt, the user query being campus security risk information of any mode; and training based on the campus security knowledge graph to obtain a security risk management large model, and inputting the retrieval enhancement prompt into the security risk management large model to obtain a risk solution. According to the scheme, the attention confidence, the feed-forward network confidence and the knowledge alignment confidence of all levels are aggregated in the reasoning process to generate the comprehensive confidence score, so that the illusion problem caused by no data reference in the model reasoning process is effectively avoided.

The invention relates to the technical field of positioning and navigation, in particular to a two-dimensional laser positioning method and system in a sparse scene, and the method comprises the steps: endowing a point cloud with a corner semantic label through a semantic segmentation model through fusing two-dimensional laser radar and camera data, and removing a dynamic object; extracting wall corner line and wall corner point features, performing data association with a pre-constructed semantic map, performing coarse positioning to obtain an initial pose, and performing fine optimization by using a semantic PLICP algorithm to obtain a fine positioning pose; and finally, carrying out confidence evaluation and fusion on the fine positioning pose in a particle filtering mode, and outputting a final optimized pose. According to the invention, the wall corner instance and features in the wall corner semantic point cloud are extracted, nearest neighbor search association and coarse positioning are carried out on the semantic map, the initial pose is obtained, semantic PLICP fine optimization is carried out, and particle filtering fusion is supplemented, so that high-precision and robust positioning independent of manual marking in a sparse scene is realized.

Techniques persist and restore in-memory neighbor graph vector indexes that include an index of vertex identifiers between layers of a plurality of layers for a graph-based approximate nearest neighbor search in a vector database. The plurality of layers include a higher layer and a lower layer that includes more vertices than the higher layer. A checkpoint is generated based on the neighbor graph vector index. The checkpoint can include a plurality of unit entries. Each unit entry can include vertex data that identifies vertices in respective subsets of a plurality of subsets of vertices in a lower layer of the neighbor graph vector index.

The invention provides a railway signal equipment fault prediction and diagnosis method, which belongs to the technical field of railway signal equipment, and comprises the following steps: acquiring multi-dimensional real-time data in the operation process of railway signal equipment, and storing the multi-dimensional real-time data according to a time sequence to form a super sparse historical data set; a hierarchical sparse data clustering index structure is established by adopting a multi-level partition clustering mode to carry out super-dilute clustering processing, an equipment fault feature vector library is established, feature vector similarity is calculated through a nearest neighbor search algorithm, and a corresponding fault prediction level is output. And carrying out correlation analysis by using the space-time alignment super-sparse network model to identify a fault propagation path between the equipment, establishing an equipment fault correlation graph, and generating a railway signal equipment fault prediction and diagnosis report according to a fault risk assessment value and an equipment fault correlation diagnosis result. The technical problem of low fault prediction accuracy of railway signal equipment in a super sparse historical data environment is solved.

The invention provides a cross-database approximate nearest neighbor search method and system and a calculation framework, and the method comprises the following steps: constructing a graph index, storing the graph index in a relation table, and obtaining a graph index table; acquiring and storing a data set and a query set in a structured relation table form; the dismantling approximate nearest neighbor search process comprises a plurality of SQL operation stages including a candidate node screening stage, a neighbor expansion stage, a distance calculation stage, a result combination stage and a priority queue maintenance stage; and based on the graph index table, executing an iterative search process of the plurality of SQL operation stages on each query point in the query set, and finally outputting an approximate nearest neighbor search result of the query set. According to the method, the graph index is combined with the relational database, and the approximate nearest neighbor search is realized by adopting a plurality of SQL operation stages, so that the high-dimensional vector retrieval efficiency and the cross-database compatibility are remarkably improved, and the large-scale application of the vector data in a multi-element scene is promoted.

The invention provides a panoramic image splicing method and device and electronic equipment, and the method comprises the steps: obtaining a plurality of images and metadata thereof, carrying out the preprocessing of the plurality of images according to the metadata, and obtaining a plurality of preprocessed images; extracting image key points and feature descriptors of the image key points from the plurality of preprocessed images through a SuperPoint neural network model; performing approximate nearest neighbor search matching on the image key points according to the feature descriptors of the image key points to obtain a plurality of initial matching pairs; eliminating matching pairs with spatial consistency smaller than a threshold value from the plurality of initial matching pairs through a random sampling consistency algorithm to obtain final matching pairs; based on the final matching pair, performing global binding adjustment on camera transformation parameters corresponding to the plurality of images through a nonlinear optimization algorithm to obtain alignment parameters, and transforming the plurality of images according to the alignment parameters to obtain a plurality of transformed images; and performing multi-band fusion rendering on the plurality of transformed images to obtain a panoramic image.

The invention discloses a high-dimensional space-oriented distributed graph index nearest neighbor search method, which comprises the following steps of: firstly, designing a boundary sensing balanced partition strategy, and measuring boundary characteristics of quantized nodes through a connection ratio; secondly, constructing a hierarchical mixed index architecture, vertically integrating a sparse global navigation layer and a fine-grained local precision layer, managing high-importance boundary nodes by adopting an M-Tree structure, and remarkably improving the cross-partition retrieval efficiency by utilizing the relative insensitivity of a tree structure to a curse in a boundary region; finally, a distributed multi-starting-point parallel search framework is provided, complementary entry points are dynamically generated based on query features and graph topology, vector space multi-region parallel exploration is achieved, irrelevant partitions are filtered through a navigation layer, and then accurate similarity calculation is executed in selected partitions. According to the method, on the premise that the high recall rate is guaranteed, the throughput of high-dimensional vector distributed search is remarkably improved, and an efficient and reliable solution is provided for large-scale high-dimensional vector retrieval.

The invention discloses a method, a device and equipment for acquiring a global initial pose of a loading machine and a medium. The method comprises the following steps: acquiring original point cloud data of an operation scene in real time by using a vehicle-mounted laser radar to generate a corresponding frame point cloud; the method comprises the following steps: constructing a real-time feature vector set based on a fast point feature histogram feature of a frame point cloud, and performing nearest neighbor search on the real-time feature vector set and a k-d tree index of a pre-established global point cloud map to form an initial matching point pair set; in each round of iteration, screening a high-precision inner point set meeting geometric constraints by adopting a self-adaptive random sampling strategy, and calculating a candidate pose transformation matrix and a corresponding inner point proportion score according to the high-precision inner point set; and a new iteration round is repeatedly executed to execute the operation until the total number of iterations reaches a target value, and the global initial pose of the target loader is screened out from the candidate matrix according to multiple rounds of scores. According to the embodiment of the invention, high-precision global initial pose acquisition of the unmanned loader is realized in an environment lacking satellite coverage, and the positioning precision and reliability are improved.

The application provides a GPS-based visual positioning method, system, computer and storage medium, which comprises the following steps: acquiring the GPS coordinates of a query image, and screening reference images with close geographical positions from a database based on the GPS coordinates; screening reference images with a geographical distance less than or equal to a preset radius as a candidate set; determining several similar images based on k-nearest neighbor search; clustering according to the co-visibility relationship of each similar image to generate a plurality of locations corresponding to the candidate image, each location containing commonly observed 3D points; performing 2D-3D matching on each location, and estimating and outputting a six-degree-of-freedom camera pose through a perspective n-point algorithm and a random sample consensus algorithm to realize visual positioning. The GPS coordinates are used to dynamically screen reference images, the search time is reduced, the quick response requirement of real-time scene application is met, and the cross-environment high robustness is guaranteed.

This invention belongs to the field of scene recognition technology and discloses a sequential multimodal scene recognition method based on a state-space model. It involves jointly encoding laser point clouds and visual images from a trajectory to form multimodal sequence data, which is then processed into unique global descriptors through a global descriptor encoding network. The global descriptors of the two trajectories serve as the map and the query, respectively. During the query process, a nearest neighbor search algorithm is used to find the most similar data in the map, completing scene recognition. This invention proposes to fuse and encode laser point cloud and image data, increasing data dimensionality and compressing data complexity. Global descriptors are obtained through a single-frame module and a sequence module based on a state-space model. The cross-scan design in the single-frame module improves computational efficiency, while the sequential full-combination representation strategy in the sequence module solves the problem of position recognition for trajectory changes within the scene. This invention features high computational efficiency, high accuracy, and robustness.

The invention discloses a multi-angle three-dimensional measurement device and a point cloud denoising method. The device comprises a projection device, an image acquisition device, a connecting piece for fixedly connecting the projection device and the image acquisition device, a two-dimensional displacement platform group for adjusting a measurement distance and an electric rotating platform for rotating a measured object, all parts are linked by a controller, and one-button automatic calibration and multi-angle measurement are realized. The method is based on the device, and noise is effectively filtered out from multi-view point clouds collected by rotating around the same axis through nearest neighbor search and statistical judgment. Through combination of displacement and the rotating platform, multi-angle data are automatically obtained, and metal reflection interference is effectively inhibited; through an innovative rotation consistency denoising algorithm, the point cloud quality is remarkably improved, and the method is suitable for multi-angle three-dimensional measurement including high-reflection workpieces.

The application belongs to the field of identity recognition, and discloses an ear brushing identity recognition method based on pseudo-label semi-supervised learning, which comprises the following steps: 1, constructing a data set for training and labeling the data set to obtain a labeled training set, an unlabeled training set and a labeled verification set; S2, determining a convolutional neural network model to be used, and modifying the training process and the detection process of the convolutional neural network model according to an improved algorithm and an identification principle; S3, training the convolutional neural network model by using an improved MeanTeacher algorithm to obtain a trained neural network model; S4, calculating a human ear image to be identified by using the trained neural network model to obtain a feature vector; and S5, matching the feature vector obtained in S4 with a human ear feature vector that has been recorded in a database and needs to be used for identification by using a nearest neighbor search algorithm to realize end-to-end human ear identity recognition. The application significantly improves the detection performance of small targets and tail categories.

This invention discloses a method and apparatus for renewable energy power prediction based on time-series discrete labeling, belonging to the field of renewable energy power prediction; it includes: acquiring historical observation multivariate time series of target power plants; constructing and training a time-series labeling mapping, dividing the multivariate time series into blocks and encoding them to obtain a continuous latent representation; normalizing the continuous latent representation and the introduced learnable codebook respectively, and allocating discrete indices using nearest neighbor search to obtain a discrete time-series label matrix; expanding the discrete time-series labels output by the trained time-series labeling mapping and incorporating them into the unified vocabulary of a pre-trained language model to construct a conditional autoregressive generative model, and performing fine-tuning training while freezing the backbone network parameters of the pre-trained language model; given the environmental semantic context and the historical time-series label sequence obtained through the time-series labeling mapping, generating a future discrete label sequence based on the trained model in an autoregressive manner, and inputting the generated future discrete label sequence into the decoder of the time-series labeling mapping to reconstruct a prediction sequence in the continuous domain.

The specification provides a vector graph index merging method and device, the method comprising: obtaining a plurality of vector graph indexes to be merged; wherein each node in the vector graph index represents a vector, and each edge represents that the vectors represented by the nodes connected by the edge are similar; for each target vector in each target vector graph index, determining the vectors similar to the target vector from each vector graph index based on an approximate nearest neighbor search algorithm to form a cross-graph candidate neighbor set of the target vector; based on the cross-graph candidate neighbor set of each vector in the plurality of vector graph indexes, constructing a global neighbor graph; wherein each node in the global neighbor graph represents each vector in the plurality of vector graph indexes, and each edge represents that the vectors represented by the nodes connected by the edge are similar; and converting the global neighbor graph into a merged graph index that can be used for approximate nearest neighbor search.