AI MVP
RAG MVP
Short answer
A RAG MVP is one job — “ask my files” — with ingest, chunking, retrieval, citations, and auth so User A cannot retrieve User B’s documents. It is not a chatbot with a PDF upload bolted on.
What is a RAG MVP?
Retrieval-Augmented Generation for a first product: users add documents, the system retrieves relevant chunks, the model answers with citations, and usage is metered. No custom-trained foundation model.
RAG prototype vs RAG product
| Piece | Prototype | MVP |
|---|---|---|
| Ingest | One PDF in a notebook | Upload + reindex + delete |
| Isolation | One vector store | Per-tenant filters |
| Citations | None | Chunk + source shown |
| Eval | “Looks good” | 20-question eval set |
| Cost | Unlimited calls | Caps + logging |
When RAG is the wrong v1
If users will not upload documents, or the answer is always a short FAQ, skip RAG. A prompt plus a database query is cheaper and more accurate.
Default stack
Next.js UI, Postgres (or Supabase) for files and jobs, a vector store (pgvector is enough to start), OpenAI or Anthropic for embeddings + chat. Queue ingest. Do not embed on every page load.
FAQ
How much does a RAG MVP cost?
A single-workflow RAG product (auth, upload, ask, citations) is typically a mid-to-high four-figure scoped build. Model spend is separate and must be capped.
Do I need a fine-tuned model?
Almost never for v1. Retrieval quality and chunking beat fine-tuning for “ask my docs.”
Related
Scoping a real “ask my files” product?
Bring the document types and the user. We’ll say RAG, search, or “don’t build this yet.”