Query Vector DB

The Query Vector DB object queries vector-enabled databases. It supports semantic similarity search on dense embeddings, keyword search with tsvector, sparse vectors, and other supported vector types.

The Query Vector DB object also supports re-ranking and hybrid search methods to refine results after the initial search, ensuring more relevant and context-aware retrievals.

Prerequisites

• A shared connection configured to a vector database.

• A shared connection to an embedding provider, if your search generates dense embeddings from a query.

Use Case

A software company stores articles in a PostgreSQL knowledge base. The database includes dense embeddings for semantic search and tsvector fields for keyword search. They use the Query Vector DB object to search either field or combine both in a hybrid search.

How to Work with the Query Vector DB Object

1. From the Source section in the Toolbox, drag-and drop a Variables object to the dataflow designer and add a user query variable in it.

2. From the sources section in the Toolbox, drag-and-drop a Query Vector Database object to the dataflow designer.

3. Map the user_query variable to the Query Input in Query Vector Db object.

The source object includes three output fields by default: Search Based On, Similarity Score, and Ranking. These fields show how each result was retrieved and ranked.

Configuring the Query Vector DB Object

1. To configure Query Vector Db object, right-click on its header and select Properties from the context menu.

The Database Connection window will now open.

1. Here select the shared connection configured to your vector database.

Note: The object currently supports PostgreSQL vector databases.

1. Next, you will see a Pick Source Table and Reading Options window. On this window, you will select the table from the database that you previously connected.

1. The next window is the Layout Builder. In this window you can modify the layout of your database table.

1. Once customized, click Next. The Query Vector DB Properties window opens. Configure an embedding model when searching dense embeddings.

Embedding Options:

These options apply when the object generates a dense embedding for the query.

• Connection Name: Select the shared connection to the embedding provider (e.g., OpenAI).

• Embedding Provider Type: Defines the embedding service being used, such as OpenAI or another provider.

• Embedding Model Name: Identifies the specific model used for generating vector embeddings (e.g., text-embedding-ada-002).

Note: Use the same embedding model that was used to create the knowledge base.

• Embedding Request Timeout Seconds: Maximum time to wait for a single embedding API request before it times out.

• Embedding Max Parallel Requests: Maximum number of embedding API Requests which can be sent concurrently.

1. Once configured, click Next. In the Search Properties window you can set up the search properties of you vector database.

Search Options:

• Max Retrieve Count: Defines the maximum number of results to retrieve.

• Search Field: Specifies the database field to search. This can store dense embeddings, sparse vectors, tsvector keyword vectors, or another supported vector type.

• Search Method: Determines how the object compares or matches values. Available methods depend on the selected vector type.

The available methods update for the selected field:

• Match Threshold: Sets a match threshold to filter out less relevant results. For example, 0.7 returns results above 70% similarity where the selected method uses similarity scoring.

• Re-Rank Result: Enables ranking of retrieved results based on secondary ranking criteria. This option is particularly useful for hybrid searches, ensuring more accurate and relevant results.

• Re-Ranking Methods: Defines the technique used for re-ranking results.

Currently, two methods are available:

• Reciprocal Rank Fusion

• Weighted Average

Note: This option can only be enabled when you have more than one search field selected.

Additionally, when Weighted Average re-ranking method is chosen, an extra field, Re-Rank Weight, is added to allow you to define the priority for each search field. The weights should always add up to 1.

• Filter Vectors: Enables additional filtering conditions for refining search results by adding a filter column.

To apply one filter across search fields, map a WHERE clause to the Where input field.

1. Click OK.

You have successfully configured your Query Vector Database object. The fields from the source object can now be mapped to other objects in a dataflow.

Preview Output

1. Right-click on the Query Vector DB object’s header and select Preview Output to view the results.

2. The preview output displays:

   • The most relevant document returned at the top. For example, if the query was "How to aggregate my data?", the first result would be the Aggregate Transformation document.

   • The second most relevant document, followed by others ranked in descending order of relevance.

   • A Search Based On column, which indicates the search field used to retrieve the record.

   • A Similarity Score column, which quantifies how closely the result matches the query.

   • A Ranking column, where Rank 1 indicates the most relevant record.