Amazon Titan Text Embeddings Models

• Amazon Titan Text Embeddings V2

  • A highly efficient model optimized for RAG and semantic search.
  • It allows for flexible output dimensions to balance accuracy with storage costs.
    • Model ID – amazon.titan-embed-text-v2:0
    • Max input text tokens – 8,192
    • Max input text characters – 50,000
    • Languages – English (100+ languages in preview)
    • Output vector size – 1,024 (default), 512, 256
    • Inference types – On-Demand, Provisioned Throughput
    • Supported use cases – RAG, document search, reranking, classification, etc.

  • V2 Request Parameters

    • The request body is passed in the body field of an InvokeModel request.
    • inputText (Required): The text string to convert into an embedding.
    • normalize (Optional): Boolean flag indicating whether to normalize the output embedding. Defaults to true.
    • dimensions (Optional): Specifies the size of the output vector. Accepted values: 1024, 512, 256.
    • embeddingTypes (Optional): A list specifying the desired return format. Accepts [“float”], [“binary”], or both. Defaults to float.

  • V2 Response Fields

    • embedding: An array of float values representing the vector of the input text.
    • inputTextTokenCount: The number of tokens detected in the input text.
    • embeddingsByType: A dictionary containing the embedding lists keyed by type (float or binary), depending on the requested embeddingTypes.

• Amazon Titan Text Embeddings G1 – Text

  • The original AWS text embedding model was designed for general-purpose semantic search and clustering.
  • Note that this model does not support additional inference parameters, such as temperature or topP.
    • Model ID –amazon.titan-embed-text-v1
    • Max Input Text – 8,192 tokens.
    • Output Vector Size – 1,536 (Fixed).
    • Languages – Supports 25+ languages.
    • Inference Types – On-Demand, Provisioned Throughput.

  • G1 Request Parameters

    • inputText (Required): The text string to convert into an embedding.

  • G1 Response Fields

    • embedding: An array of values representing the embedding vector.
    • inputTextTokenCount: The number of tokens in the input.

Amazon Titan Multimodal Embeddings G1 Model

• • This model converts both text and images (or a combination of both) into a shared vector space, enabling cross-modal capabilities like “search by image” or “text-to-image” search.
    • Model ID – amazon.titan-embed-image-v1
    • Max input text tokens – 256
    • Languages – English
    • Max input image size – 25 MB
    • Output vector size – 1,024 (default), 384, 256
    • Inference types – On-Demand, Provisioned Throughput
    • Supported use cases – Search, recommendation, and personalization.

  • Request & Response

    • Request:
      • Accepts inputText (description), inputImage (base64 encoded string), or both. Includes embeddingConfig to specify output embedding length.
    • Response:
      • Returns the embedding vector, inputTextTokenCount, and an optional message.

Amazon Titan Image Generator G1 Models

Generates and edits studio-quality images from natural language prompts. Both versions share a common API structure but differ in capabilities and resolution support.

• Amazon Titan Image Generator v1 (overview)

  • Enables users to generate and edit images from natural language prompts.
    • Model ID – amazon.titan-image-generator-v1
    • Max input characters – 512 char
    • Max input image size – 5 MB (only some specific resolutions are supported)
    • Max image size using in/outpainting – 1,408 x 1,408 px px
    • Max image size using image variation – 4,096 x 4,096 px
    • Languages – English
    • Output type – image
    • Supported image types – JPEG, JPG, PNG
    • Inference types – On-Demand, Provisioned Throughput
    • Supported use cases – image generation, image editing, image variations

• Amazon Titan Image Generator v2 (overview)

  • Supports all Titan Image Generator v1 features with added new capabilities.
  • Allows use of reference images to guide layout and composition while following text prompts.
  • Includes automatic background removal for images with multiple objects.
    • Model ID – amazon.titan-image-generator-v2:0
    • Max input characters – 512 char
    • Max input image size – 5 MB (only some specific resolutions are supported)
    • Max image size using in/outpainting, background removal, image conditioning, color palette – 1,408 x 1,408 px
    • Max image size using image variation – 4,096 x 4,096 px
    • Languages – English
    • Output type – image
    • Supported image types – JPEG, JPG, PNG
    • Inference types – On-Demand, Provisioned Throughput
    • Supported use cases – image generation, image editing, image variations, background removal, color guided content

• Inference Parameters & Responses

  • When making an InvokeModel call, the body field determines the task. All tasks utilize a shared imageGenerationConfig object (for quality, dimensions, and quantity), but require specific parameter objects for the task type.
  • Supported Tasks:

• • Response Body:
    • Returns an images array containing the base64-encoded strings of the generated output(s) and an error field if applicable.