Vector database connector

The vector database connector allows embedding, storing, and retrieving Large Language Model (LLM) embeddings. This enables building AI-based solutions for your organizations, such as context document search, long-term LLM memory, and agentic AI interaction.

note

The vector database connector uses the LangChain4J library. Data models and possible implementations are limited to the latest stable released LangChain4j library.

Prerequisites

Before using the vector database connector, ensure you understand the concept of LLM embeddings.

To start using the vector database connector, ensure you have access to a supported LLM embeddings API to convert document content into vectorized embedding form. You will also need to have write access to a supported database.

Create a connector task

You can apply a connector to a task or event via the append menu. For example:

• From the canvas: Select an element and click the Change element icon to change an existing element, or use the append feature to add a new element to the diagram.
• From the properties panel: Navigate to the Template section and click Select.
• From the side palette: Click the Create element icon.

After you have applied a connector to your element, follow the configuration steps or see using connectors to learn more.

Operations

Embed document

The embed document operation performs the following steps:

1. Consume a document.
2. Parse the document depending on a file format (optionally split into text chunks).
3. Convert chunks into a vector form with LLM help.
4. Store produced vectors in a vector database.

To perform this operation, enter the following:

• Operation dropdown: Embed document.
• Embedding model: Refer to the relevant section.
• Vector store: Refer to the relevant section.
• Document: Refer to the relevant section.

As a result of this operation, you will get an array of created embedding chunk IDs, for example ["d599ec62-fe51-4a91-bbf0-26e1241f9079", "a1fad021-5148-42b4-aa02-7de9d590e69c"].

Retrieve document

The retrieve document operation performs the following steps:

1. Consumes a query.
2. Convert the query into a vector form with LLM help.
3. Perform a vector similarity search on previously-stored LLM embeddings.
4. Store results in Camunda document storage.

To perform this operation, enter the following:

• Operation dropdown: Retrieve document.
• Search query: Enter your search query.
• Max results: Enter maximum amount of returned results.
• Min score: Enter the lowest score threshold similarity value; the value should be between 0 and 1, for example, 0.81.
• Embedding model: Refer to the relevant section.
• Vector store: Refer to the relevant section.

As a result of this operation, you will get an array of relevant chunks, where each includes a chunk ID, Camunda document reference metadata, similarity score, and the actual text content.

{
  "chunks": [
    {
      "chunkId": "e30d570b-2a3a-4f4a-9a0c-78f0f1acd383",
      "documentReference": {
        "storeId": "local",
        "documentId": "2b36ec67-a78f-4f99-9371-8c6e5b332838",
        "contentHash": "1c232bc1e553c10d00c3327dcca9012b6b4b0758a1c2afaad8b77c80fa1bd36e",
        "metadata": {
          "size": 116,
          "fileName": null,
          "processDefinitionId": null,
          "processInstanceKey": null,
          "customProperties": {},
          "expiresAt": null,
          "contentType": "text/plain"
        },
        "camunda.document.type": "camunda"
      },
      "score": 0.6721556,
      "content": "Camunda is a platform for orchestrating and automating business processes. It helps organizations design, execute, and manage workflows, enabling them to optimize processes and improve efficiency."
    }
  ]
}

Embedding models

Amazon Bedrock

The vector database connector currently supports only Amazon Titan V1/V2 models. Review the official Amazon documentation to understand how to choose request parameters.

The vector database connector uses LangChain4j implementation.

Vector stores

Elasticsearch

The vector database connector can use Elasticsearch as a vector store. The Elasticsearch version must be 8+.

Enter the following parameters:

• Base URL: The Elasticsearch base URL, including protocol, for example https://host:port.
• Username: For the Elasticsearch user that has read/write access.
• Password: For the Elasticsearch user that has read/write access.
• Index name: Name of the index where you wish to store embeddings.
• • When embedding: If index is not present, the connector will create a new one.
• • When retrieving: If the index is absent, the connector will raise an error.

Amazon OpenSearch

• Access key and Secret key: Enter AWS IAM credentials for the user that has read/write access.
• Server URL: An Amazon OpenSearch URL without protocol, for example my-opensearch.aws.com:port.
• Region: Region of the Amazon OpenSearch instance.
• Index name: Name of the index where you wish to store embeddings.
• • When embedding: If index is not present, the connector will create a new one.
• • When retrieving: If the index is absent, the connector will raise an error.

Embedding document configuration

Document source

The Document source can be either Plain text or a Camunda document.

Plain text can be useful when you deal with small size data that can fit into a text field or a process instance variable. Input will be handled as a regular UTF-8 text.

note

A FEEL string conversion function might be useful if you have JSON input.

The Camunda document might be useful when you deal with larger document pipelines that come from webhook or user tasks. Input documents will be parsed with Apache Tika, so files can be of any Apache Tika-supported formats.

Splitting

Splitting is an action of breaking large documents into smaller pieces. It can be either recursive or no splitting at all. Seek guidance from your local data scientist to determine if you require splitting.

Learn more about splitting in the LangChain4j documentation.