Embedding Batch (AI)

Overview#

• Transformation Type: Partially blocking
• Execution Mode: One generator call per batch
• Default batch size: 64
• Package: ETLBox.AI

Create the generator the same way as for a single row. The OpenAI example on Embedding (AI) is the client this component expects. Any other provider that exposes IEmbeddingGenerator works as well.

The transformation waits for BatchSize rows, or for the remainder when the source completes. Five rows with a batch size of 2 produce three calls. Vector number i is written onto row number i of that batch. The same object continues downstream.

Embedding a batch#

TextSelector and VectorSetter point at the text and the vector, exactly as on the row transformation.

public class Review {
    public int Id { get; set; }
    public string Text { get; set; }
    public float[] Embedding { get; set; }
}

var source = new MemorySource<Review>();
source.DataAsList.Add(new Review { Id = 1, Text = "Great product, fast delivery." });
source.DataAsList.Add(new Review { Id = 2, Text = "Broken after two days." });

var embed = new EmbeddingBatchTransformation<Review>(generator, batchSize: 64) {
    TextSelector = row => row.Text,
    VectorSetter = (row, vector) => row.Embedding = vector.ToArray()
};
var dest = new MemoryDestination<Review>();

source.LinkTo(embed).LinkTo(dest);
Network.Execute(source);

[EmbeddingText] and [EmbeddingVector] replace the delegates when you would rather mark the class. On an ExpandoObject, TextPropertyName and VectorPropertyName select the properties. The defaults are Text and Embedding.

Asking for a shorter vector#

EmbeddingGenerationOptions is forwarded on every batch. Use Dimensions when the model can return a smaller vector.

embed.EmbeddingGenerationOptions = new EmbeddingGenerationOptions { Dimensions = 256 };

When the generator returns fewer vectors#

By default, only the rows that received a vector continue. Set RequireSameItemCount when a short response should fail the batch instead. With LinkErrorTo, that batch is redirected once and does not appear on the success output.

The rows are ready for a Vector Store (AI) destination. A relational VECTOR column is still a normal database column. Use the database connectors for that, not the vector store components.

More examples are in the embedding batch recipes.