Abstract: A method for low-latency data processing based on distributed storage using a high-speed communication interface, which is performed in a distributed storage system including a plurality of hosts, includes communicatively connecting a source host and a target host through a first communication interface and a second communication interface, transmitting a command from the source host to the target host through the first communication interface, and transmitting data to the target host from the source host through a shared memory of a switch device by using the second communication interface. A speed of the second communication interface is higher than a speed of the first communication interface, and a delay of the second communication interface is lower than a delay of the first communication interface.

Abstract: A method of processing multiple outstanding vector requests in a vector search system includes transmitting, by an apparatus for executing a vector index host, one or more queries to an apparatus for executing a vector search upon receiving the one or more queries; creating, by the apparatus for executing a vector index host, a candidate node list for searching the queries, for each of the queries; confirming, by the apparatus for executing a vector index host, a multiple-search counting number, selecting one or more search target nodes from the candidate node list, and transmitting a search request with respect to neighbor nodes of the search target node to the apparatus for executing the vector search; and transmitting, by the apparatus for executing a vector search, a search result for the search request to the apparatus for executing a vector index host.

Abstract: Provided is a method of building a vector database (DB) on a server providing a vector DB system including setting a hyperplane for projecting a cluster of vector data, generating a projection tree of the cluster, and building the vector DB by using the projection tree as an index. According to the method, a vector cluster is projected onto a hyperplane passing through its center point, thereby improving projection balance.

Abstract: Provided is a method of storing a vector in a server providing a vector DB system. The method of storing a vector includes slicing a 16-bit vector into an upper 8-bit MSB block and a lower 8-bit LSB block, rounding the MSB block by reflecting information of the LSB block to the MSB block, and storing the rounded MSB block and the LSB block.

Abstract: Disclosed is a method for processing multiple vector searches including receiving, by an apparatus executing a vector index host, a search task set including a plurality of search tasks, associatively performing, by the apparatus executing the vector index host, a similarity-based search scheduling task based on similarity between the plurality of search tasks included in the search task set, performing, by the apparatus executing the vector index host, a search request by delivering, in an order of the similarity-based search scheduling task, a query vector corresponding to each search task to an apparatus executing a vector search, and delivering, by the apparatus executing the vector search, a search result for the search request to the apparatus executing the vector index host. Representative drawing is FIG. 8.

Abstract: Provided are a method and apparatus for improving vector search efficiency for multimodal data in vector databases. The method includes: generating a first vector index structure for first modality data; generating a second vector index structure for second modality data different from the first modality data; connecting the first vector index structure and the second vector index structure; and searching for a node similar to a query vector using the connected first vector index structure and second vector index structure. According to the method, it is possible to improve the accuracy of a vector search for multimodal data.

Abstract: Provided are a server, method, and computer-readable recording medium for allocating a storage space for a vector index. The method includes generating a vector index structure of data points, generating a priority queue based on importance of nodes, calculating the importance of the nodes using the vector index structure, and deleting a vector index corresponding to a node having importance of a preset reference or less from a memory. According to this method, it is possible to improve the response of a system by optimizing the storage of vector indexes.

Abstract: Provided are a server, method, and computer-readable recording medium for searching for a vector. The method includes generating a vector index structure of data points, searching for a node similar to a query vector using the vector index structure, calculating a similarity between the node and the query vector, and updating the similarity by giving a weight to a vector index inflow time of the node.