New AI models in the Amazon cloud

At its annual in-house exhibition re:Invent, Amazon Web Services (AWS) presented enhancements to its managed service platform Amazon Bedrock. The innovations include the Amazon Bedrock Marketplace with new AI models and revised routing and caching functions for automatic selection of the model used. The company is also expanding the options for accessing structured data and adding knowledge graphs and automated data preparation.

With the Amazon Bedrock Marketplace, AWS is expanding its selection of foundation models. This gives customers access to over 100 models that are tailored to different usage scenarios. For example, users can generate source code or videos with AI support. Luma AI presented the new AI model Ray 2 at the trade fair, which will be included in the marketplace portfolio. It can be used to create five to ten-second clips from text prompts in the style of a movie. AWS had also previously announced its own Nova models, which can also process and generate text, images and videos and split prompts between different models.

Bedrock automatically selects favorable AI models

AWS is also introducing new mechanisms for dealing with inference queries, i.e. the execution of AI algorithms. Prompt caching allows repeatedly used prompts to be cached; Intelligent Prompt Routing automatically selects different models within a model family depending on the complexity of the query. Inexpensive, smaller models process simple queries, while more demanding queries go to more powerful models. This automated approach should save customers up to 30 percent of costs without compromising on response quality.

The developers are also revising access to structured data sources so that queries in natural language are now automatically translated into SQL queries. According to AWS, this facilitates the integration of data from data warehouses, data lakes or tools such as Amazon SageMaker Lakehouse. The company is also adding GraphRAG support to Bedrock. This allows knowledge from knowledge graphs to flow into answer generation. This should enable the AI to recognize relationships between data points more clearly and deliver more accurate and comprehensible results.

Automatic data conversion without manual coding

With Amazon Bedrock Data Automation, unstructured, multimodal content can now be automatically converted into structured formats without manual coding. The tool systematically extracts relevant information from texts and images, transforms it and integrates it into existing databases. This is intended to simplify analytical evaluations and workflows of the retrieval augmented generation. RAG applications not only draw on the AI's training data, but also on external sources of information such as the internet.

The Amazon Bedrock Marketplace is available immediately. Customers can initially test the other new features as part of a preview.

(mma)