/01
Corpus and chunking
Decide what is in scope, who can access it, and how it is chunked before touching a model.
- Source list
- Per-tenant scoping
- Chunk size
- Refresh cadence
Resource / Checklist
RAG Readiness Checklist
A pre-build checklist for retrieval-augmented generation in a SaaS product: corpus, chunking, evals, citations, and cost.
/ Outcome
What you walk away with.
A RAG plan you can execute without producing a fragile demo, with measurable accuracy from day one.
/ System
/02
Evaluation and cost
Build a small eval set from real questions and pin a per-user cost ceiling before launch.
- Eval set
- Citation correctness
- Latency budget
- Cost ceiling
/ Checklist
Run through the list.
- 01List every document type the model can read
- 02Define tenant boundaries and access checks
- 03Pick chunking strategy and embedding model
- 04Build a 50-question eval set from real users
- 05Add citations to every answer
- 06Set per-user and per-month cost caps
/ FAQ
Questions founders ask.
Do we need a vector database?
Often Postgres with pgvector is enough at the start. Move to a dedicated store when query patterns demand it.
How do we measure quality?
Run the eval set on every change and watch citation correctness, not just answer overlap.