RAG = Retrieval-Augmented Generation. The model doesn't 'know' your data. It looks it up at query time and reads it like a human reading a Wikipedia article.
What RAG is good at
- Answering questions over a corpus that updates frequently (docs, support tickets, internal wikis)
- Citing sources — every answer can be traced back to a chunk
- Keeping costs down vs. stuffing the entire corpus into context
What RAG is bad at
- Reasoning across many documents at once ("compare every contract we signed in 2023")
- Tasks where the answer requires synthesis, not lookup
- Anything that needs personality or style — fine-tuning wins there