Vector Index

FLAT

For vector similarity search applications that require perfect accuracy and depend on relatively small (million-scale) datasets, the FLAT index is a good choice. FLAT does not compress vectors, and is the only index that can guarantee exact search results. Results from FLAT can also be used as a point of comparison for results produced by other indexes that have less than 100% recall.

FLAT is accurate because it takes an exhaustive approach to search, which means for each query the target input is compared to every vector in a dataset. This makes FLAT the slowest index on our list, and poorly suited for querying massive vector data. There are no parameters for the FLAT index in Milvus, and using it does not require data training or additional storage.

• Search parameters

  Parameter	Description	Range
  metric_type	[Optional] The chosen distance metric.	See Supported Metrics.

IVF_FLAT

IVF_FLAT divides vector data into nlist cluster units, and then compares distances between the target input vector and the center of each cluster. Depending on the number of clusters the system is set to query (nprobe), similarity search results are returned based on comparisons between the target input and the vectors in the most similar cluster(s) only — drastically reducing query time.

By adjusting nprobe, an ideal balance between accuracy and speed can be found for a given scenario. Results from the IVF_FLAT performance test demonstrate that query time increases sharply as both the number of target input vectors (nq), and the number of clusters to search (nprobe), increase.

IVF_FLAT is the most basic IVF index, and the encoded data stored in each unit is consistent with the original data.

• Index building parameters

  Parameter	Description	Range
  nlist	Number of cluster units	[1, 65536]

• Search parameters

  Parameter	Description	Range
  nprobe	Number of units to query	CPU: [1, nlist] GPU: [1, min(2048, nlist)]

IVF_SQ8

IVF_FLAT does not perform any compression, so the index files it produces are roughly the same size as the original, raw non-indexed vector data. For example, if the original 1B SIFT dataset is 476 GB, its IVF_FLAT index files will be slightly larger (~470 GB). Loading all the index files into memory will consume 470 GB of storage.

When disk, CPU, or GPU memory resources are limited, IVF_SQ8 is a better option than IVF_FLAT. This index type can convert each FLOAT (4 bytes) to UINT8 (1 byte) by performing scalar quantization. This reduces disk, CPU, and GPU memory consumption by 70–75%. For the 1B SIFT dataset, the IVF_SQ8 index files require just 140 GB of storage.

• Index building parameters

  Parameter	Description	Range
  nlist	Number of cluster units	[1, 65536]

• Search parameters

  Parameter	Description	Range
  nprobe	Number of units to query	CPU: [1, nlist] GPU: [1, min(2048, nlist)]

IVF_PQ

PQ (Product Quantization) uniformly decomposes the original high-dimensional vector space into Cartesian products of m low-dimensional vector spaces, and then quantizes the decomposed low-dimensional vector spaces. Instead of calculating the distances between the target vector and the center of all the units, product quantization enables the calculation of distances between the target vector and the clustering center of each low-dimensional space and greatly reduces the time complexity and space complexity of the algorithm.

IVF_PQ performs IVF index clustering before quantizing the product of vectors. Its index file is even smaller than IVF_SQ8, but it also causes a loss of accuracy during searching vectors.

• Index building parameters

  Parameter	Description	Range
  nlist	Number of cluster units	[1, 65536]
  m	Number of factors of product quantization	dim ≡ 0 (mod m)
  nbits	[Optional] Number of bits in which each low-dimensional vector is stored.	[1, 16] (8 by default)

• Search parameters

  Parameter	Description	Range
  nprobe	Number of units to query	[1, nlist]

HNSW

HNSW (Hierarchical Navigable Small World Graph) is a graph-based indexing algorithm. It builds a multi-layer navigation structure for an image according to certain rules. In this structure, the upper layers are more sparse and the distances between nodes are farther; the lower layers are denser and the distances between nodes are closer. The search starts from the uppermost layer, finds the node closest to the target in this layer, and then enters the next layer to begin another search. After multiple iterations, it can quickly approach the target position.

In order to improve performance, HNSW limits the maximum degree of nodes on each layer of the graph to M. In addition, you can use efConstruction (when building index) or ef (when searching targets) to specify a search range.

• Index building parameters

  Parameter	Description	Range
  M	Maximum degree of the node	[4, 64]
  efConstruction	Search scope	[8, 512]

• Search parameters

  Parameter	Description	Range
  ef	Search scope	[top_k, 32768]

ANNOY

ANNOY (Approximate Nearest Neighbors Oh Yeah) is an index that uses a hyperplane to divide a high-dimensional space into multiple subspaces, and then stores them in a tree structure.

When searching for vectors, ANNOY follows the tree structure to find subspaces closer to the target vector, and then compares all the vectors in these subspaces (The number of vectors being compared should not be less than search_k) to obtain the final result. Obviously, when the target vector is close to the edge of a certain subspace, sometimes it is necessary to greatly increase the number of searched subspaces to obtain a high recall rate. Therefore, ANNOY uses n_trees different methods to divide the whole space, and searches all the dividing methods simultaneously to reduce the probability that the target vector is always at the edge of the subspace.

• Index building parameters

  Parameter	Description	Range
  n_trees	The number of methods of space division.	[1, 1024]

• Search parameters

  Parameter	Description	Range
  search_k	The number of nodes to search. -1 means 5% of the whole data.	{-1} ∪ [top_k, n × n_trees]