Sushant Dotel
ProjectsAbout

AWS Gen AI Challenge — Day 1

Published on

  • AWSGenAIChallenge

Welcome to day 1! My goal for today is to understand different aws services and tools related to Gen AI, and get my hands dirty! Here's a summary of what I learned today:

Bedrock is the core AWS's managed service for experimenting with and deploying pre-trained foundation models. You get access to multiple models (e.g. Claude, Titan) through a single API, without managing infrastructure.

Foundation model is a large pre-trained model like GPT-4 that can be adapted to a wide range of tasks. They are called "foundation" models because they serve as the base upon which you can build specific applications through techniques like fine-tuning or RAG.

Building the foundation of a model from scratch is called pre-training. Big companies already do that for you with a lot of guardrails built in. Once you have your foundation model, you can adapt it to your specific use cases through fine-tuning, continued pre-training, or RAG.

Some other AWS services and tools related to Gen AI:

  • Bedrock data manipulation: Turn unstructured data into structured form.
  • Bedrock Knowledge Base: vector store for your embeddings that integrates with S3, SharePoint, and Confluence.
  • Vector storage: You can choose a storage whenever you create a knowledge base. Use OpenSearch(recommended by AWS) or PostgreSQL (e.g. pgvector) for vector search and retrieval.
  • Generative AI application builder: Low-code way to build Gen AI apps on top of Bedrock.
  • Bedrock prompt management: Version and manage prompts in one place.
  • Bedrock Flow: Multi-step, multi-agent style workflows.
  • Amazon SageMaker pipelines: For automated ML workflows, including testing and validation of models.

Using Bedrock for a POC

Bedrock provides all the tools you need to build a Gen AI POC:

  • Models: Multiple FMs behind one API.
  • Knowledge Base: For RAG without building pipelines from scratch.
  • Agents: For tool use and multi-step reasoning.
  • Model evaluation: Evaluate outputs against your own criteria.
  • Guardrails: Built-in safety features to prevent harmful outputs.

To be continued...