What is Amazon Bedrock?¶
It's a service that's allows developers to build and scale GenAI.
Key Features of Amazon Bedrock¶
- Access to multiple foundations models
Amazon Bedrock foundation model¶
Allows users to build and scale applications using foundation models.
Key features of Bedrock foundation models¶
- Pre-trained on a diverse range of models available
Example of Bedrock foundation models available¶
- anthropic
- ai21labs
- cohere
- stabilityAI
Customizable¶
Users can fine-tune the foundation models to perform a wide range of tasks.
Serverless¶
Amazon Bedrock is a serverless service.
API driven¶
It enables seamless integration.
Use cases¶
- Text generation
- Image generation
- Audio generation
- Video generation
- Code generation
- Data generation
What is foundation module evaluation?¶
Process of testing the foundation models to ensure they are performing as expected.
Why we should evaluate foundation models?¶
To ensure the foundation models are performing as expected.
How to evaluate foundation models?¶
- Run test cases
- Monitor performance
- Collect feedback
Classification of foundation models¶
- Accuracy: For classification or decision making task.
- Bleu or Rouge: For text generation task.
- F1: For precision and recall task.
- Human evaluation: For subjective tasks like image, audio, video generation.
- Latency: For real-time applications.
- Cost: For cost-sensitive applications.
GuardRails¶
- Control the interaction between the model and the user.
- Filter out inappropriate content.
- Remove personaly identifiable information.
- Enhanced privacy and security.
- Reduce hallucinations and biases.
Agents¶
- Enable developers to create and manage autonomous agents of executing complex multi-step tasks.
CloudWatch¶
- Monitor and troubleshoot the performance of the foundation models.
Studio¶
- Visual interface to manage and monitor the foundation models.
Whatermark detection¶
- Check if an image was generated by Amazon Titan
Pricing¶
On-demand¶
- Pay as you go
- Text model: Charge for every input/output
- Embedding model: Charge for every generated input
- Image model: Charge for every generated image
Batch¶
- Can provide 50% discount
Provisioned Throughput¶
- Purchase model: 1 month, 6 months
- Throughput: max number of input/output
- Works with base, fine-tuned, and custom models