Table of Contents

Creating Vector Databases

Learn how to create and manage vector databases in the Management Portal.

Overview

A vector database resource describes a database (for Azure AI Search, an index) in a vector storage service: which service to connect to, which embedding model and dimensions the stored vectors use, and which properties hold the content, embedding, vector store ID, and metadata. Knowledge units reference a vector database, and pipeline embedding and indexing stages read these settings through the knowledge unit.

Prerequisites

  • The Vector Databases Contributor role at the instance scope, so the Create Vector Database button is shown.
  • An API endpoint configuration for the vector service, in the General category with the Indexing subcategory. See API Endpoints.

Vector Databases List

In the sidebar, select Data > Vector Databases.

Column Description
Name The vector database name.
Database Type AzureAISearch, AzureCosmosDBNoSQL, or AzurePostgreSQL.
Actions View JSON definition ({}), Edit (pencil), and Delete (trash).

Use Search vector databases to filter the list, and the refresh button to reload it.

Create a Vector Database

  1. On the Vector Databases page, select Create Vector Database.

  2. Fill in the form. All fields except the description and cost center are required.

    Section Field Default Description
    What is the name of the vector database? Vector database name Letters, numbers, dashes, and underscores only. A ✔️ or ❌ shows whether the name is available.
    Description Optional description.
    What is the type of the vector database? Type Azure AI Search Azure AI Search, Azure Cosmos DB NoSQL, or Azure PostgreSQL.
    What is the database name in the vector database service? Database name For Azure AI Search, the index name. The Azure AI Search indexing stage creates the index if it doesn't exist.
    What is the API endpoint configuration for this vector database? API endpoint An API endpoint configuration in the General category and Indexing subcategory.
    What is the embedding model? Embedding model text-embedding-3-large The embedding model used for the stored vectors.
    How many dimensions do the embeddings have? Embedding dimensions 2048 Must match the embedding model output.
    What are the property names in the vector database? Embedding property name Embedding Property that stores the embedding vector.
    Content property name Text Property that stores the text content.
    Vector store ID property name VectorStoreId Property that identifies the vector store (logical partition).
    Metadata property name Metadata Property that stores item metadata.
    Metadata properties FileId\|Edm.String,FileName\|Edm.String,UniqueId\|Edm.String,ItemType\|Edm.String Comma-separated name\|type pairs for the metadata fields to index.
    Would you like to assign this vector database to a cost center? Cost center Optional.

  3. Select Create Vector Database.

If required fields are missing, a notification lists each problem (for example, "Please select an API endpoint configuration."). After saving, you return to the list. Select Cancel and confirm with Yes to discard your changes.

For Azure AI Search, valid metadata property types are Edm.String, Edm.Int32, Edm.Int64, Edm.Single, Edm.Double, Edm.Boolean, and Edm.DateTimeOffset.

Vector Stores

A single vector database can hold several vector stores. Each item's vector store ID property identifies which store it belongs to. You choose the vector store ID in the knowledge unit (Vector Store ID) or in the indexing stage (VectorStoreId parameter). See Creating Knowledge Units.

Edit a Vector Database

  1. Select the Edit (pencil) icon in the list.
  2. The Edit Vector Database page opens. The name cannot be changed.
  3. Update the settings and select Save Changes. A notification confirms the save and you stay on the page.

When editing, select Access Control at the top right to manage role assignments on the vector database.

Caution: If the vector database already holds data, changing the embedding model, dimensions, or property names means the existing content no longer matches the new settings. Plan such changes together with re-indexing.

Delete a Vector Database

Select the Delete (trash) icon, then select Yes in the Delete Vector Database dialog.

Troubleshooting

  • The API endpoint dropdown is empty: Create an API endpoint configuration with the General category and Indexing subcategory.
  • Dimension mismatch errors during indexing: Make sure Embedding dimensions matches the output of the embedding model.
  • Connection errors: Check the API endpoint configuration URL and authentication, and that the vector service allows network access from FoundationaLLM.