Help Center/ Cloud Search Service/ User Guide/ Using OpenSearch for Data Search/ Enhancing Search Capabilities for OpenSearch Clusters/ Configuring Vector Search for OpenSearch Clusters/ About Vector Search
Updated on 2024-10-12 GMT+08:00
View PDF
Share

About Vector Search

Unstructured data, such as images, videos, and language corpora, is converted into vectors, which are searched based on similarity using either an exact or approximate nearest neighbors algorithm.

How It Works

Vector search works in a way similar to traditional search. To improve vector search performance, we need to:

  • Narrow down the matched scope

    Similar to traditional text search, vector search use indexes to accelerate the search instead of going through all data. Traditional text search uses inverted indexes to filter out irrelevant documents, whereas vector search creates indexes for vectors to bypass irrelevant vectors, narrowing down the search scope.

  • Reduce the complexity of calculating a single vector

    The vector search method can quantize and approximate high dimensional vectors first. By doing this, you can acquire a smaller and more relevant data set. Then more sophisticated algorithms are applied to this smaller data set to perform computation and sorting. This way, complex computation is performed on only part of the vectors, and efficiency is improved.

Vector search means to retrieve the k-nearest neighbors (KNN) to the query vector in a given vector data set by using a specific measurement method. Generally, CSS only focuses on Approximate Nearest Neighbor (ANN), because a KNN search requires excessive computational resources.

Vector Search in CSS

The CSS vector search engine integrates a variety of vector indexes, such as brute-force search, Hierarchical Navigable Small World (HNSW) graphs, product quantization, and IVF-HNSW. It also supports multiple similarity calculation methods, such as Euclidean, inner product, cosine, and Hamming. The recall rate and retrieval performance of the engine are better than those of open-source engines. It can meet the requirements for high performance, high precision, low costs, and multi-modal computation.

The search engine also supports all the capabilities of the native Elasticsearch, including distribution, multi-replica, error recovery, snapshot, and permission control. The engine is compatible with the native Elasticsearch ecosystem, including the cluster monitoring tool Cerebro, the visualization tool Kibana, and the real-time data ingestion tool Logstash. Several client languages, such as Python, Java, Go, and C++, are supported.

Constraints

  • Only Elasticsearch 7.6.2, Elasticsearch 7.10.2, and OpenSearch 1.3.6 clusters support the vector search engine of CSS.
  • The vector search plug-in performs in-memory computing and requires more memory than common indexes do. You are advised to use memory-optimized compute specifications.
  • To use the vector search plug-in of CSS, the memory capacity of a data node or cold data node must be greater than 16 GB. To enable the vector search plug-in, contact technical support.

Cluster Node Specifications Selection for Vector Search

Off-heap memory is used for index construction and query in vector search. Therefore, the required cluster capacity is related to the index type and off-heap memory size. You can estimate the off-heap memory required by full indexing to select the appropriate cluster specifications. Due to high memory usage of vector search, CSS disables the vector search plug-in by default for clusters whose memory is 8 GB or less.

There are different methods for estimating the size of off-heap memory required by different types of indexes. The calculation formulas are as follows:
  • GRAPH Index

    mem_needs = (dim x dim_size + neighbros x 4) x num + delta

    If you need to update indexes in real time, consider the off-heap memory overhead required for vector index construction and automatic merge. The actual size of required mem_needs is at least 1.5 to 2 times of the original estimation.

  • PQ Index

    mem_needs = frag_num x frag_size x num + delta

  • FALT and IVF Indexes

    mem_needs = dim x dim_size x num + delta

Table 1 Parameter description

Parameter

Description

dim

Vector dimensionality

neighbors

Number of neighbors of a graph node. The default value is 64.

dim_size

Number of bytes required by each dimension. The default value is four bytes in the float type.

num

Total number of vectors

delta

Metadata size. This parameter can be left blank.

frag_num

Number of vector segments during quantization and coding. If this parameter is not specified when an index is created, the value is determined by vector dimensionality dim.

if dim <= 256: 
  frag_num = dim / 4
elif dim <= 512: 
  frag_num = dim / 8
else :
  frag_num = 64

frag_size

Size of the center point during quantization and coding. The default value is 1. If the value of frag_num is greater than 256, the value of frag_size is 2.

These calculation methods can estimate the size of off-heap memory required by a complete vector index. To determine cluster specifications, you also need to consider the heap memory overhead of each node.

Heap memory allocation policy: The size of the heap memory of each node is half of the node physical memory, and the maximum size is 31 GB.

For example, if you create a Graph index for the SIFT10M dataset, with dim set to 128, dim_size to 4, neighbors to the default value 64, and num to 10 million, the off-heap memory required by the Graph index is around 7.5 GB. Calculation formula: mem_needs = (128 x 4 + 64 x 4) x 10000000 ≈ 7.5.

Considering the overhead of heap memory, a single server with 8 vCPUs and 16 GB memory is recommended. If real-time write or update is required, you need to apply for larger memory.