Tutorials · Chapter D (4/4) · ~11 min

Build a mini RAG

Play → read → next

Retrieve a snippet, then answer from it — a RAG path you can walk through out loud.

Code Lab

Build a mini RAG

Retrieve a doc chunk, then print it as the grounded answer.

Python

# Mini RAG in pure Python (toy search)
docs = {
  "cookies": "Bake cookies 12 minutes at 180C.",
  "phones": "Phones stay in bags during class.",
  "hours": "Store open 9am to 8pm.",
}

def retrieve(query, k=1):
    q = set(query.lower().split())
    scored = []
    for name, text in docs.items():
        overlap = len(q & set(text.lower().split()))
        scored.append((overlap, name, text))
    scored.sort(reverse=True)
    return scored[:k]

query = "How long do cookies bake?"
hits = retrieve(query, k=1)
print("retrieved:", hits[0][1], "->", hits[0][2])
print("answer:", hits[0][2])

# Challenge: add a docs entry for "library" and ask about library hours

Recap

What you just did

You built the builder’s favorite pattern: search your stuff → prompt with evidence → generate. Even a toy keyword or embedding search counts. The showable win is the cited snippet sitting next to the answer, not a floating claim.

Teach

How it works

Toy RAG in a few moves:

docs = {
    "hours": "Cafe open 8am–6pm weekdays.",
    "wifi": "Wi‑Fi password is painted by the register.",
}
q = "when do you open?"
# retrieve: pick the best matching doc
hit = docs["hours"]
prompt = f"Use only this note:\n{hit}\n\nQ: {q}\nA:"
# then send `prompt` to the LLM

Index your notes/chunks
Retrieve top matches for the question
Augment the prompt with those chunks
Generate an answer that should stick to them

Mental model: open-book quiz, not closed-book celebrity trivia.

Use it

When you'd use this

FAQ bots over your handbook
Support answers that must quote policy text
Personal study buddy over your lecture notes

Watch out

If retrieval misses, the model invents. Always surface “I used this chunk” (or “no chunk found”). Don’t stuff the whole company drive into one prompt — chunk and rank.

Try next

Try this next

Ask a question your docs don’t cover. Prefer an honest “not in notes” over a confident guess.