CHAPTER 07 · Retrieval: RAG vs Tools vs Long Context · 4 / 6
A decision guide
Ask, in order:
- Is the content small and always needed whole? → Long context. Don't over-engineer.
- Is it a bounded, identified set the model can reason about by name/structure? → Tool-based reads (plus a list tool). Best for exactness, freshness, citations.
- Is it a large corpus where the model can't know what exists? → Add a search tool (keyword or vector). Keep reads as tools so the model pulls full, exact text for whatever search surfaces.
- Do you need both whole-document work and corpus-wide search? → Hybrid: read/find/list tools + a search tool backing into RAG.