../

Stanford CS230 | Autumn 2025 | Lecture 8: Agents, Prompts, and RAG

Reference: https://www.youtube.com/watch?v=k1njvbBmfsw&t=22s

Problem with base models

  • Lack of domain specific knowledge
  • Trained data is old
    • Not feasible retrain the LLM every month
  • Breadth focus training is not great for depth first results
  • Context management
    • Attention is not great with very large context windows
    • Needle in the haystack not really ideal situation for LLMs

Dimensions for improvement

  • The base model itself
    • Better data
    • Different architecture
    • Bigger model
  • Improve the context and the tools surrounding the LLM

Prompt Design

  • Think like an expert xyz
  • Ask it to break it down step by step (Chain of Thought)
  • Few shot examples
  • Chaining
    • This is different from chain of thought. There we ask the LLM to break it down into steps in its output
    • This is splitting the prompt into individual chunks
    • The advantage is that we can test which part of the prompt is weaker/stronger