OpenSearch

LangStream supports using OpenSearch as a vector database.

Learn more about performing vector search with OpenSearch in the official documentation

Only OpenSearch 2.x is officially supported.

Connecting to OpenSearch

Create a vector-database resource in your configuration.yaml file. A single resource is bound to a single index.

resources:
    - type: "vector-database"
      name: "OpenSearch"
      configuration:
        service: "opensearch"        
        username: "${secrets.opensearch.username}"
        password: "${secrets.opensearch.password}"
        host: "${secrets.opensearch.host}"
        port: "${secrets.opensearch.port}"
        index-name: "my-index-000"

Connecting to AWS OpenSearch service

resources:
    - type: "vector-database"
      name: "OpenSearch"
      configuration:
        service: "opensearch"
        username: "${secrets.opensearch.username}"
        password: "${secrets.opensearch.password}"
        host: "${secrets.opensearch.host}"
        region: "${secrets.opensearch.region}"
        index-name: "my-index-000"

• username is the AWS Access Key

• password is the AWS Secret Key

• host is the endpoint provided by AWS. e.g. for AWS OpenSearch serverless it looks like this: xxxx..aoss.amazonaws.com

• region is the AWS region. It has to match with the one used in the endpoint

Declare an index as asset

To bind the application to the OpenSearch index creation at startup, you must use the opensearch-index asset type.

You can configure settings and mappings as you prefer. Other configuration fields are not supported.

This is an example mixing normal fields with vector fields. The knn plugin is required in the target OpenSearch instance.

- name: "os-index"
  asset-type: "opensearch-index"
  creation-mode: create-if-not-exists
  config:
    datasource: "OpenSearch"
    settings: |
      {
            "index": {
                  "knn": true,
                  "knn.algo_param.ef_search": 100
            }
        }
    mappings: |
      {
            "properties": {
                  "content": {
                        "type": "text"
                  },
                  "embeddings": {
                        "type": "knn_vector",
                        "dimension": 1536
                  }
            }
        }

Refer to the settings documentation for the settings field. Refer to the mappings documentation for the mappings field.

Search

Use the query-vector-db agent with the following parameters to perform searches on the index created above :

  - name: "lookup-related-documents"
    type: "query-vector-db"
    configuration:
      datasource: "OpenSearch"
      query: |
        {
          "size": 1,
          "query": {
            "knn": {
              "embeddings": {
                "vector": ?,
                "k": 1
              }
            }
          }
        }
      fields:
        - "value.question_embeddings"
      output-field: "value.related_documents"

You can use the '?' symbol as a placeholder for the fields.

The query is the body sent to OpenSearch. Refer to the documentation to learn which parameters are supported. Note that the query will be executed on the configured index. Multi-index queries are not supported, but you can declare multiple datasources and query different indexes in the same application.

The output-field will contain the query result. The result is an array with the following elements:

• id: the document ID

• document: the document source

• score: the document score

• index: the index name

For example, if you want to keep only one relevant field from the first result, use the compute agent after the search:

  - name: "lookup-related-documents"
    type: "query-vector-db"
    configuration:
      datasource: "OpenSearch"
      query: |
        {
          "size": 1,
          "query": {
            "match_all": {}
          }
        }
      output-field: "value.related_documents"
      only-first: true
  - name: "Format response"
    type: compute
    configuration:
      fields:
        - name: "value"
          type: STRING
          expression: "value.related_documents.document.content"

Indexing

Use the vector-db-sink agent to index data, with the following parameters:

  - name: "Write to Solr"
    type: "vector-db-sink"
    input: chunks-topic
    configuration:
      datasource: "OpenSearch"
      bulk-parameters:
        timeout: 2m
      fields:
        - name: "id"
          expression: "fn:concat(value.filename, value.chunk_id)"
        - name: "embeddings"
          expression: "fn:toListOfFloat(value.embeddings_vector)"
        - name: "text"
          expression: "value.text"

All indexing is performed using the Bulk operation. You can customize the bulk parameters with the bulk-parameters property.

The request will be flushed depending on flush-interval and batch-size parameters.

Configuration

Check out the full configuration properties in the API Reference page.