Harnessing EDA & Workflows to Build Real World GenAI Apps • Uma Ramadoss & Veda Raman • GOTO 2024

Learn how to build production-grade GenAI apps using event-driven architectures & workflows. Covers prompt decomposition, human-in-the-loop validation & best practices.

Key takeaways
  • Event-driven architectures and workflows (like AWS Step Functions) provide better observability, extensibility and control when building GenAI applications

  • Breaking down large prompts into smaller tasks (prompt decomposition) allows for:

    • Parallel processing
    • Using smaller/specialized LLMs per task
    • Better cost efficiency
    • Improved accuracy
    • Easier maintenance
  • Human-in-the-loop validation is crucial for content moderation and quality control, even though it increases latency

  • Using managed services like Amazon Bedrock provides:

    • Access to multiple foundation models through single API
    • Security and data privacy
    • Scalability without infrastructure management
    • Pay-per-use pricing model
  • The Strangler Fig pattern helps gradually introduce GenAI capabilities into existing applications without full re-architecture

  • Event routers (like Amazon EventBridge) help decouple services and allow selective event consumption

  • Prompt engineering techniques like persona patterns and prompt chaining improve response quality

  • As GenAI applications move from POC to production:

    • Prompts tend to grow larger
    • Requirements expand
    • Testing complexity increases
    • Integration needs multiply
  • Workflow orchestration provides:

    • Error handling and retries
    • Process visibility
    • Task coordination
    • Integration capabilities
  • Building evolutionary architecture with event-driven patterns enables faster development and easier model switching as technology advances