Elasticsearch Elasticsearch knn

By Opster Team

Updated: Jun 22, 2023

| 3 min read

Introduction to kNN Search in Elasticsearch

The k-Nearest Neighbor (kNN) search is a powerful vector search technique used in Elasticsearch for similarity search and recommendation systems. It enables you to find the k most similar documents to a given query document based on a specific distance metric. The kNN search is particularly useful in scenarios where you need to find similar items, such as product recommendations, image search, and document similarity.

In this article, we will discuss advanced techniques and optimization strategies for kNN search in Elasticsearch. We will cover the following topics:

  1. Indexing and searching with kNN
  2. Distance metrics
  3. Indexing and search performance optimization
  4. Handling large-scale data

1. Indexing and Searching with kNN

To use kNN search in Elasticsearch, you need to create an index with a specific mapping that includes a dense_vector field type. This field type is used to store the vector representation of your data. Here’s an example of how to create an index with a dense_vector field:

PUT /my_index
  "mappings": {
    "properties": {
      "my_vector": {
        "type": "dense_vector",
        "dims": 128

In this example, the “my_vector” field has a dimensionality of 128 and uses the Euclidean distance, also called L2 distance, similarity. You can index documents with their vector representation using the following format:

PUT /my_index/_doc/1
  "my_vector": [0.1, 0.2, 0.3, ..., 0.128]

To perform a kNN search, you can use the knn search option when running your search. Here’s an example of a kNN search using the Euclidean distance metric:

POST /my_index/_search
  "knn": {
    "field": "my_vector",
    "query_vector": [0.1, 0.2, 0.3, ..., 0.128],
    "k": 10,
    "num_candidates": 100

This query will return the 10 best matches between the input vector and the indexed ones considering 100 candidate documents per shard.

2. Distance Metrics

Elasticsearch supports several distance metrics for kNN search, including Euclidean distance, cosine similarity, and dot product. You can choose the appropriate distance metric based on your use case and data characteristics. Here are some examples:

  • Euclidean distance: Suitable for dense vectors and when the magnitude of the vectors is important.
  • Cosine similarity: Suitable for sparse vectors and when the angle between the vectors is important.
  • Dot product: Suitable for cases where the magnitude of the vectors is not important, and you want to focus on the direction.

3. Indexing and Search Performance Optimization

To optimize the performance of kNN search in Elasticsearch, consider the following strategies:

  • Use shard allocation awareness to distribute the index across multiple nodes, which can help parallelize the search process and improve query performance.
  • Use the filter context to pre-filter documents before performing the k-NN search, which can help reduce the search space and improve query performance.
  • Decrease num_candidates to increase the query speed at the cost of less accurate results.
  • Increase num_candidates to increase the accuracy of the results at the cost of a slower query.
  • Use “element_type”: “byte” and/or a smaller dimensionality in the dense_vector field mapping in order to consume less space in your index.

4. Handling Large-Scale Data

When dealing with large-scale data, you may need to consider additional strategies to improve the performance and scalability of kNN search in Elasticsearch

  • Use dimensionality reduction techniques, such as PCA or t-SNE, to reduce the size of the vector representation and improve search performance.
  • Use distributed search techniques, such as cross-cluster search or federated search, to search across multiple Elasticsearch clusters.

5. Optimizing Vector Search Performance

To optimize vector search performance in Elasticsearch, consider the following tips:

  • Use smaller vector dimensions: Reducing the dimensionality of your vectors can improve search performance. However, this may also affect the quality of search results. Experiment with different dimensions to find the best trade-off between performance and search quality.
  • Use filtering: If you can filter out irrelevant documents before performing the vector search, you can significantly improve performance. For example, you can use a bool query to combine a filter query with the script_score query.
  • Optimize hardware resources: Ensure that your cluster has sufficient resources, such as CPU, memory, and disk space, to handle the vector search workload. Monitor the performance of your cluster and adjust resources as needed.
  • Use caching: Elasticsearch caches the results of frequently executed queries. Make sure that the cache settings are properly configured to take advantage of this feature.


In this article, we discussed advanced techniques and optimization strategies for kNN search in Elasticsearch. By understanding the underlying concepts and applying the appropriate optimization strategies, you can build powerful similarity search and recommendation systems using Elasticsearch.

How helpful was this guide?

We are sorry that this post was not useful for you!

Let us improve this post!

Tell us how we can improve this post?