LLM stacks \ 4.5.1 External data RAG - terrytaylorbonn/auxdrone GitHub Wiki

25.0717 Lab notes (Gdrive), Git

• Agent calls RAG DB API ✅. Why does the agent decided to call RAG?
  • Depends on implementation! In simple_rag.py, it always calls RAG first
  • More sophisticated agents might:
    • Check if question seems factual first
    • Try RAG, then fallback to general knowledge
    • Let the model decide: "Do I need to search documents?"
• 1 Agent vectorizes prompt ✅
  • Agent uses embedding model (e.g., "nomic-embed-text") to convert prompt to vectors
  • Same embedding model was used to create the RAG database vectors
• 2 Agent finds best match and retrieves ORIGINAL TEXT ✅
  • "Original text" is the actual readable text that was vectorized
  • The vectors are just for finding the text
  • for example: [return results["documents"][0]] ← This is readable text
• 3 Agent gives retrieved text to model. Agent also sends the original human prompt.
  • For example :
    • context = relevant_docs[0][:200] # Retrieved text
    • prompt = f"Based on this: '{context}' Answer: {question}" # Both context + original question