In-memory Index

This topic lists various types of in-memory indexes Milvus supports, scenarios each of them best suits, and parameters users can configure to achieve better search performance. For on-disk indexes, see On-disk Index.

Indexing is the process of efficiently organizing data, and it plays a major role in making similarity search useful by dramatically accelerating time-consuming queries on large datasets.

To improve query performance, you can specify an index type for each vector field.

Currently, a vector field only supports one index type. Milvus automatically deletes the old index when switching the index type.

ANNS vector indexes

Most of the vector index types supported by Milvus use approximate nearest neighbors search (ANNS) algorithms. Compared with accurate retrieval, which is usually very time-consuming, the core idea of ANNS is no longer limited to returning the most accurate result, but only searching for neighbors of the target. ANNS improves retrieval efficiency by sacrificing accuracy within an acceptable range.

According to the implementation methods, the ANNS vector index can be divided into four categories:

Tree-based index
Graph-based index
Hash-based index
Quantization-based index

Indexes supported in Milvus

According to the suited data type, the supported indexes in Milvus can be divided into two categories:

Indexes for floating-point embeddings:
- For 128-dimensional floating-point embeddings, the storage they take up is 128 * the size of float = 512 bytes. And the distance metrics used for float-point embeddings are Euclidean distance (L2) and Inner product.
- These types of indexes include FLAT, IVF_FLAT, IVF_PQ, IVF_SQ8, HNSW, and SCANN^(beta) for CPU-based ANN searches and GPU_IVF_FLAT and GPU_IVF_PQ for GPU-based ANN searches.
Indexes for binary embeddings
- For 128-dimensional binary embeddings, the storage they take up is 128 / 8 = 16 bytes. And the distance metrics used for binary embeddings are Jaccard and Hamming.
- This type of indexes include BIN_FLAT and BIN_IVF_FLAT.

The following table classifies the indexes that Milvus supports:

Floating-point embeddings Binary embeddings

Supported index	Classification	Scenario
FLAT	N/A	Relatively small dataset Requires a 100% recall rate
IVF_FLAT	Quantization-based index	High-speed query Requires a recall rate as high as possible
GPU_IVF_FLAT	Quantization-based index	High-speed query Requires a recall rate as high as possible
IVF_SQ8	Quantization-based index	High-speed query Limited memory resources Accepts minor compromise in recall rate
IVF_PQ	Quantization-based index	Very high-speed query Limited memory resources Accepts substantial compromise in recall rate
GPU_IVF_PQ	Quantization-based index	Very high-speed query Limited memory resources Accepts substantial compromise in recall rate
HNSW	Graph-based index	Very high-speed query Requires a recall rate as high as possible Large memory resources
SCANN	Quantization-based index	Very high-speed query Requires a recall rate as high as possible Large memory resources

Supported index	Classification	Scenario
BIN_FLAT	Quantization-based index	Depends on relatively small datasets. Requires perfect accuracy. No compression applies. Guarantee exact search results.
BIN_IVF_FLAT	Quantization-based index	High-speed query Requires a recall rate as high as possible

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 set of vectors in a dataset. This makes FLAT the slowest index on our list, and poorly suited for querying massive vector data. There are no parameters required for the FLAT index in Milvus, and using it does not need data training.

Search parameters

Parameter Description Range

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

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 Default Value

nlist Number of cluster units [1, 65536] 128
Search parameters

Parameter Description Range Default Value

nprobe Number of units to query [1, nlist] 8

Parameter	Description	Range	Default Value
`nprobe`	Number of units to query	[1, nlist]	8

GPU_IVF_FLAT

Similar to IVF_FLAT, GPU_IVF_FLAT also 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.

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

When conducting searches, note that you can set the top-K up to 256 for any search against a GPU_IVF_FLAT-indexed collection.

Index building parameters

Parameter Description Range Default Value

nlist Number of cluster units [1, 65536] 128
Search parameters

Parameter Description Range Default Value

nprobe Number of units to query [1, nlist] 8
Limits on search

Parameter Range

top-K <= 256

Parameter	Description	Range	Default Value
`nprobe`	Number of units to query	[1, nlist]	8

Parameter	Range
`top-K`	<= 256

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 smaller (~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 (SQ). 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 [1, nlist]

Parameter	Description	Range
`nprobe`	Number of units to query	[1, 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 and search parameters vary with Milvus distribution. Select your Milvus distribution first.

Index building parameters

Parameter	Description	Range
`nlist`	Number of cluster units	[1, 65536]
`m`	Number of factors of product quantization	`dim mod m == 0`
`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]

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

SCANN

SCANN (Score-aware quantization loss) is similar to IVF_PQ in terms of vector clustering and product quantization. What makes them different lies in the implementation details of product quantization and the use of SIMD (Single-Instruction / Multi-data) for efficient calculation.

Index building parameters

Parameter Description Range

nlist Number of cluster units [1, 65536]

with_raw_data Whether to include the raw data in the index True or False. Defaults to True.

Unlike IVF_PQ, default values apply to m and nbits for optimized performance.
Search parameters

Parameter Description Range

nprobe Number of units to query [1, nlist]

reorder_k Number of candidate units to query [top_k, ∞]
Range search parameters

Parameter Description Range

radius Number of units to query [1, nlist]

range_filter Number of candidate units to query [top_k, ∞]

Parameter	Description	Range
`nprobe`	Number of units to query	[1, nlist]
`reorder_k`	Number of candidate units to query	[`top_k`, ∞]

Parameter	Description	Range
`radius`	Number of units to query	[1, nlist]
`range_filter`	Number of candidate units to query	[`top_k`, ∞]

GPU_IVF_PQ

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 and search parameters vary with Milvus distribution. Select your Milvus distribution first.

When conducting searches, note that you can set the top-K up to 8192 for any search against a GPU_IVF_FLAT-indexed collection.

Index building parameters

Parameter	Description	Range	Default Value
`nlist`	Number of cluster units	[1, 65536]	128
`m`	Number of factors of product quantization	`dim mod m == 0`	4
`nbits`	[Optional] Number of bits in which each low-dimensional vector is stored.	[1, 16]	8

Search parameters

Parameter Description Range Default Value

nprobe Number of units to query [1, nlist] 8
Limits on search

Parameter Range

top-K <= 1024

Parameter	Description	Range	Default Value
`nprobe`	Number of units to query	[1, nlist]	8

Parameter	Range
`top-K`	<= 1024

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 [2, 2048]

efConstruction Search scope [1, int_max]
Search parameters

Parameter Description Range

ef Search scope [top_k, int_max]

Parameter	Description	Range
`M`	Maximum degree of the node	[2, 2048]
`efConstruction`	Search scope	[1, int_max]

Parameter	Description	Range
`ef`	Search scope	[`top_k`, int_max]

BIN_FLAT

This index is exactly the same as FLAT except that this can only be used for binary embeddings.

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

BIN_FLAT is accurate because it takes an exhaustive approach to search, which means for each query the target input is compared to vectors in a dataset. This makes BIN_FLAT the slowest index on our list, and poorly suited for querying massive vector data. There are no parameters for the BIN_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.

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

BIN_IVF_FLAT

This index is exactly the same as IVF_FLAT except that this can only be used for binary embeddings.

BIN_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. Query time increases sharply as both the number of target input vectors (nq), and the number of clusters to search (nprobe), increase.

BIN_IVF_FLAT is the most basic BIN_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 [1, nlist]

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

FAQ

What is the difference between FLAT index and IVF_FLAT index?

IVF_FLAT index divides a vector space into nlist clusters. If you keep the default value of nlist as 16384, Milvus compares the distances between the target vector and the centers of all 16384 clusters to get nprobe nearest clusters. Then Milvus compares the distances between the target vector and the vectors in the selected clusters to get the nearest vectors. Unlike IVF_FLAT, FLAT directly compares the distances between the target vector and each and every vector.

Therefore, when the total number of vectors approximately equals nlist, IVF_FLAT and FLAT has little difference in the way of calculation required and search performance. But as the number of vectors grows to two times, three times, or n times of nlist, IVF_FLAT index begins to show increasingly greater advantages.

See How to Choose an Index in Milvus for more information.