Ship a Tiny RAG App in One Afternoon
A concrete path from folder of docs to a working Q&A endpoint — without drowning in framework soup.
You do not need a platform to learn RAG. You need a folder of docs and a weekend afternoon.
Stack (intentionally small)
- Python + FastAPI
- A local embedding model or a hosted embeddings API
- A simple vector store (even SQLite + vectors is fine to start)
- One chat model for generation
Steps
- Load markdown/PDF text.
- Chunk and embed.
- Store vectors with source metadata.
- On query: retrieve top-k, build prompt, generate.
- Return answer + citations.
Definition of done
You can ask three questions about your docs and get answers with real citations. That is enough for v1.
Next upgrades
- Hybrid search (keyword + vector)
- Reranking
- Eval harness
- Caching for repeated queries
Ship the thin slice first. Complexity is optional; learning is not.