Skip to content

setup-elasticsearch.md

Latest commit

 

History

History
70 lines (51 loc) · 3.37 KB

setup-elasticsearch.md

File metadata and controls

70 lines (51 loc) · 3.37 KB

Elasticsearch Setup

To configure Elasticsearch for use with NLWeb, you'll need to set a few environment variables or update your configuration.

Connecting to an Elasticsearch Server

To connect to a running Elasticsearch server, set the following environment variables:

  • ELASTICSEARCH_URL: The URL of your Elasticsearch instance (e.g., http://localhost:9200).
  • ELASTICSEARCH_API_KEY: The Elasticsearch API key

If you don't have already an instance of Elasticsearch, you can start a free trial using Elastic Cloud or Serverless. Moreover, you can install Elasticsearch locally using the start-local installer.

Open a terminal and run the following command:

curl -fsSL https://elastic.co/start-local | sh

This will install Elasticsearch and Kibana locally. ELasticsearch will be executed on localhost:9200 with the credentials stored in the elastic-start-local/.env file, including the API key.

Configuring Elasticsearch as vector database

The Elasticsearch configuration is available in the config/config_retrieval.yaml file.

  elasticsearch:
    # Elasticsearch endpoint (localhost or remote URL with Elastic Cloud/Serverless)
    api_endpoint_env: ELASTICSEARCH_URL
    # Authentication credentials
    api_key_env: ELASTICSEARCH_API_KEY
    # Index name to search in
    index_name: nlweb_embeddings
    # Database type
    db_type: elasticsearch
    # Vector properties
    vector_type:
      type: dense_vector

You can change the configuration, e.g. changing the index_name. You can customize the vector_type property using dense_vector or sparse_vector type.

You can optimize the Elasticsearch vector storage with many options. For instance, you can quantize vectors using a single byte. You can specify int8_hnsw as index_options type, to use the 1 byte quantization in Elasticsearch, as follows:

  elasticsearch:
    # ...
    vector_type:
      type: dense_vector
      index: true
      index_options:
        type: "int8_hnsw"

This reduces the memory footprint by 75% (or 4x) at the cost of some accuracy.

You can optionally specify the vector dimension (dims) in the configuration, although it's not required—Elasticsearch will automatically use the dimension of the first embedding it receives.

To learn more about configuring Elasticsearch as a vector database, we recommend checking out the following articles: