Notes and material on ai applications capable of generating a fully functional software application from scratch.

Context

  • RepoCoder
    • core idea is to have multiple iterations of code generation via llm (Y) and using Y to query vector store for relevant code to generate the final result
    • $C_i^{ret} = R(C_{repo}, X, \hat{Y}^{i-1})$

Orchestration

  • ChatDev
    • code
    • orchestration via a chain of dual agents communicating and formulazing instructions for the next pair.
  • MetaGPT
    • orchestration via different roles like product manager, architect, engineer and QA generating relevating documents and sharing with the team to act on. Focus is on structured documents based communication rather than natural language.
  • DyLAN
  • AgentVerse

Agent

  • React
    • loop over llm calls with thought, action and observation until the final answer. possibly limit the iterations as well in some manner to stop hallucinating.
  • Raise
  • Reflexion
  • LATS