Tune approximate kNN search

Elasticsearch supports approximate k-nearest neighbor search for efficiently finding the k nearest vectors to a query vector. Since approximate kNN search works differently from other queries, there are special considerations around its performance.

Many of these recommendations help improve search speed. With approximate kNN, the indexing algorithm runs searches under the hood to create the vector index structures. So these same recommendations also help with indexing speed.

When indexing vectors for approximate kNN search, you need to specify the similarity function for comparing the vectors. If you’d like to compare vectors through cosine similarity, there are two options.

The cosine option accepts any float vector and computes the cosine similarity. While this is convenient for testing, it’s not the most efficient approach. Instead, we recommend using the dot_product option to compute the similarity. To use dot_product, all vectors need to be normalized in advance to have length 1. The dot_product option is significantly faster, since it avoids performing extra vector length computations during the search.

Elasticsearch uses the HNSW algorithm for approximate kNN search. HNSW is a graph-based algorithm which only works efficiently when most vector data is held in memory. You should ensure that data nodes have at least enough RAM to hold the vector data and index structures. To check the size of the vector data, you can use the Analyze index disk usage API. As a loose rule of thumb, and assuming the default HNSW options, the bytes used will be num_vectors * 4 * (num_dimensions + 12). When using the byte element_type the space required will be closer to num_vectors * (num_dimensions + 12). Note that the required RAM is for the filesystem cache, which is separate from the Java heap.

The data nodes should also leave a buffer for other ways that RAM is needed. For example your index might also include text fields and numerics, which also benefit from using filesystem cache. It’s recommended to run benchmarks with your specific dataset to ensure there’s a sufficient amount of memory to give good search performance. You can find here and here some examples of datasets and configurations that we use for our nightly benchmarks.

If the machine running Elasticsearch is restarted, the filesystem cache will be empty, so it will take some time before the operating system loads hot regions of the index into memory so that search operations are fast. You can explicitly tell the operating system which files should be loaded into memory eagerly depending on the file extension using the index.store.preload setting.

The following file extensions are used for the approximate kNN search: "vec" (for vector values), "vex" (for HNSW graph), "vem" (for metadata).

The speed of kNN search scales linearly with the number of vector dimensions, because each similarity computation considers each element in the two vectors. Whenever possible, it’s better to use vectors with a lower dimension. Some embedding models come in different "sizes", with both lower and higher dimensional options available. You could also experiment with dimensionality reduction techniques like PCA. When experimenting with different approaches, it’s important to measure the impact on relevance to ensure the search quality is still acceptable.

Elasticsearch stores the original JSON document that was passed at index time in the _source field. By default, each hit in the search results contains the full document _source. When the documents contain high-dimensional dense_vector fields, the _source can be quite large and expensive to load. This could significantly slow down the speed of kNN search.

You can disable storing dense_vector fields in the _source through the excludes mapping parameter. This prevents loading and returning large vectors during search, and also cuts down on the index size. Vectors that have been omitted from _source can still be used in kNN search, since it relies on separate data structures to perform the search. Before using the excludes parameter, make sure to review the downsides of omitting fields from _source.

Elasticsearch shards are composed of segments, which are internal storage elements in the index. For approximate kNN search, Elasticsearch stores the dense vector values of each segment as an HNSW graph. kNN search must check each segment, searching through one HNSW graph after another. This means kNN search can be significantly faster if there are fewer segments. By default, Elasticsearch periodically merges smaller segments into larger ones through a background merge process. If this isn’t sufficient, you can take explicit steps to reduce the number of index segments.

The force merge operation forces an index merge. If you force merge to one segment, the kNN search only need to check a single, all-inclusive HNSW graph. Force merging dense_vector fields is an expensive operation that can take significant time to complete.

A common pattern is to first perform an initial bulk upload, then make an index available for searches. Instead of force merging, you can adjust the index settings to encourage Elasticsearch to create larger initial segments:

Actively indexing documents can have a negative impact on approximate kNN search performance, since indexing threads steal compute resources from search. When indexing and searching at the same time, Elasticsearch also refreshes frequently, which creates several small segments. This also hurts search performance, since approximate kNN search is slower when there are more segments.

When possible, it’s best to avoid heavy indexing during approximate kNN search. If you need to reindex all the data, perhaps because the vector embedding model changed, then it’s better to reindex the new documents into a separate index rather than update them in-place. This helps avoid the slowdown mentioned above, and prevents expensive merge operations due to frequent document updates.

Search can cause a lot of randomized read I/O. When the underlying block device has a high readahead value, there may be a lot of unnecessary read I/O done, especially when files are accessed using memory mapping (see storage types).

Most Linux distributions use a sensible readahead value of 128KiB for a single plain device, however, when using software raid, LVM or dm-crypt the resulting block device (backing Elasticsearch path.data) may end up having a very large readahead value (in the range of several MiB). This usually results in severe page (filesystem) cache thrashing adversely affecting search (or update) performance.

You can check the current value in KiB using lsblk -o NAME,RA,MOUNTPOINT,TYPE,SIZE. Consult the documentation of your distribution on how to alter this value (for example with a udev rule to persist across reboots, or via blockdev --setra as a transient setting). We recommend a value of 128KiB for readahead.