Amazon Bedrock Feedback

Amazon Bedrock is a fully managed service that offers a choice of high-performing foundation models through a single API, along with a broad set of capabilities you need to build generative AI applications with security, privacy, and responsible AI. Here are the key features:

• It provides a wide range of model support, including AI21 Labs, Anthropic, Cohere, DeepSeek, Luma, Meta, Mistral AI, Stability AI, and Amazon.

• It provides a comprehensive RAG (Retrieval Augmented Generation) capability, utilizing Amazon Bedrock Knowledge Bases to parse, analyze, and extract meaningful insights.

• Agentic workflows are supported, allowing you to create an agent in just a few quick steps. You need to select a foundation model, provide it with access to your enterprise systems and knowledge bases, and then use AWS Lambda functions to execute your APIs securely.

• Foundation models can be easily fine-tuned using your own labeled datasets in just a few quick steps.

This approach of wrapping commonly available frameworks and models is very typical of Amazon. For example, Sagemaker wraps pre-trained models from Hugging Face, TensorFlow, PyTorch, and proprietary ones. CloudFormation provides declarative abstraction over infrastructure such as S3, EC2, and Lambda; you don’t need to call each service individually. AWS AppSync gives application developers the ability to access data from multiple databases, microservices, and AI models with a single GraphQL API request. Again, Amazon RDS provides a wrapper around various relational databases.

I played with Bedrock capabilities using the Amazon Bedrock Workshop. Here is what I liked about the workshop:

• It is a solid workshop with deeply technical content that should resonate with developers.

• The workshop provides a Jupyter notebook for each session, making it easy to follow along with the text descriptions and execute the code.

• All code for the workshop is in an open source repo, allowing you to contribute directly to the workshop.

• I love the Converse API, which enables you to interact with multiple foundation models using a single interface.

• The workshop is very comprehensive and covers a large number of GenAI use cases: text summarization (non-streaming and streaming), code generation, switching models at the backend with a single front-end API (Converse), function calling, RAG (retrieve and generate, retrieve and prompt augmentation), multimodal text/image generation, emeddings, and search, text-to-image generation and extensive image manipulation, text-to-video and image-and-text to video generation, and agentic workflows.

• The workshop utilizes a large number of foundation models, including Amazon Titan, Anthropic Claude 3.7 Sonnet, Meta Llama, DeepSeek-R1, Amazon Nova Micro, Amazon Titan Text Embeddings, Claude Haiku, Amazon Nova Canvas, Amazon Nova Reel, and Amazon Titan Multimodal Embedding Model. I particularly enjoyed Amazon Nova Canvas and Amazon Nova Reel, as well as the simplicity of their workshops.

• It demonstrates a deep connection with a wide range of AWS services, including Amazon Bedrock, Amazon OpenSearch Serverless, S3, Amazon SageMaker AI, AWS Lambda, and Amazon DynamoDB.

Here are some gotchas with developer flow in Bedrock and the workshop:

• Model support in Bedrock varies by region. The supported models for non-us-east-1 regions are minimal. So I started the workshop in us-west-2 region but ended up re-running everything in us-east-1 to make progress.

• The workshop talks about “Amazon Bedrock text playground” but does not provide instructions on how to access it.

• No models are enabled by default, and requesting access to the models is a multi-step process.

• The second module points to a Jupyter Notebook on GitHub, but provides no clear instructions on how to set it up. Finally, what worked was creating a Notebook instance using Sagemaker AI. Filed a bug #364 to simplify the developer experience.

• Giving IAM permissions is always a nightmare experience. Although permissions are relatively simple, identifying which role or ARN can always be a bit tricky.

• There is no intuitive way to request “model access”, unless you go through the “Chat playground / text” and then click on the “request access”. This should be available in the left-side nav menu.

• I filed several bugs: #364, #365, #366, #367, #368, #369, and #370.

I could navigate this workshop as someone familiar with AWS terminology and quirks. However, it would be pretty daunting for someone new to AWS. Particularly, ChatGPT provides a simplistic interface for all of these tasks.

I also tried to create a video to promote this workshop, but #370 blocked me from moving forward. Google Gemini cannot create videos, Sora requires a paid plan, and this is where Claude comes in handy. However, their generated video was an HTML page and so had to be converted into an MP4 file.

If you have gone through the workshop, then you would get the banner image ;)