I attended an in-personal workshop provided by Deloitte and AWS for NZ TechWeek24 on 22nd of May, noted down some key points that I probably can learn further with hands-on projects later.
Key concepts:
- LLMOps
- Considerations for shortlisting LLMs
- Hallucination & Retrieval-Augmented Generation (RAG) pattern
- Embeddings
- Conversational Buffer Memory
- Prompt Engineering Techniques
- Fine Tuning (just lightly touched)
Some use cases in Deloitte we went through:
- Customer support GenAI POC - understand customer query, extract relevant parts, draft email/slack responses (100% consistency of response msgs), and then provides links to knowledge base - 25% decreased request handling time
- Knowledge Base Summarisation for Chorus - more to read
- Query Structured Data from internal supported vector data store - using the same stack/tools we used in labs below
The stack and tools we used in the labs:
- Python boto3
- Amazon Bedrock - fully managed service for using foundation models from Amazon and third parties
- LangChain - Python and JS libraries, provides convenient functions for interacting with Amazon Bedrock’s models and related services like vector databases
- Streamlit - quickly creates web UI from Python without much frontend skills, great for POCs (Streamlit API Reference)
- Amazon Titan Embeddings - converts natural language text into numerical representations for later use cases such as searching or comparing semantic similarity
If you are interested in GenAI on AWS, there are a few skill builder free labs for AI Readiness to explore around.
Top comments (0)