Build a RAG agent

[KenjiPcx]

The main power feature is in this agent, it will

• Take in a user problem
• Build a step by step how to build a startup document to solve this problem, containing
  • The tools that could be used
  • The team, number of people
  • The reddit post url as reference

This is where we try to fit as much nvidia products as possible. I prefer using langgraph if possible

1. Use Nvidia inference engine and choose their latest NVLM model, i think D is the powerful one?
2. User enters problem they're trying to solve
3. RAG pipeline:
   • enrich search query (query expansion)
   • embed search query using same embedding model as the one used in llamaindex
   • reranking results (which result is actually useful)
4. Summarize results