RAG explained for buyers
Retrieval-augmented generation in plain language—what you buy, what can go wrong, and what to ask vendors.
RAG (retrieval-augmented generation) answers questions using your documents retrieved at runtime, then summarized by a model—reducing pure hallucination versus guessing from weights alone.
What you are paying for
- Ingestion — connectors, chunking, metadata, refresh jobs
- Retrieval — embeddings, indexes, access filters
- Generation — prompts, citations, refusal rules
- Evaluation — test questions, human review, drift monitoring
Questions for any vendor
- How do you enforce document-level permissions?
- How often is the index refreshed? What happens on delete?
- Show citation behavior on ambiguous questions
- What is the fallback when retrieval confidence is low?
Limits RAG does not remove
- Stale policy documents produce stale answers
- Poorly scanned PDFs still parse badly
- Models can misread retrieved chunks—human review matters for external-facing use
Related delivery
We implement RAG with explicit eval sets—see services and agent development cost for budgeting drivers.