What we're building

A function that takes the chunks Lesson 6's retrieval returned and hands them to Gemini, alongside the original question, so the answer is grounded in that specific text instead of whatever Gemini happened to learn during training. This is "generation," the fourth and final stage of Naive RAG.

"Stuffing" is a real, if inelegant, name

The technique here doesn't have a more sophisticated name because there isn't a more sophisticated technique yet at this point in the course: you take the retrieved text, and you literally paste it into the prompt, ahead of the question. Gemini has no separate channel for "trusted background information," from its point of view, the context and the question are just one long piece of text it's asked to respond to.

That simplicity is exactly why this is called Naive RAG. Later lessons in this series (Corrective RAG, Agentic RAG) add steps like grading the retrieved text before using it, or deciding whether to retrieve again, but the core move, put the text in the prompt, stays the same underneath all of them.

The code, piece by piece

context = "\n\n---\n\n".join(chunk["text"] for chunk in retrieved_chunks)

Joins every retrieved chunk's text into one block, with a clear --- separator between them. Without a separator, two adjacent chunks could visually run together and confuse the model about where one topic ends and the next begins.

prompt = f"""Answer the question using only the context below. If the
context doesn't contain the answer, say so, don't guess.


Context:
{context}


Question: {query}"""

The instruction ("using only the context... don't guess") matters more than it might look. Without it, Gemini will happily blend the retrieved context with whatever it already "knows" from training, which defeats the purpose: you want an answer grounded in this document, not a plausible-sounding mixture of this document and outside knowledge. Lesson 15 comes back to this instruction and tightens it further.

response = client.models.generate_content(model=CHAT_MODEL, contents=prompt)
return response.text or ""

The same generate_content call from Lesson 1, just with a longer, more structured prompt this time. or "" guards against the rare case where .text comes back empty (a fully blocked or filtered response), so the function's return type stays a plain string either way.

Checkpoint

  • generation: handing retrieved text to the model alongside the question, so the answer is grounded in that specific text.
  • "stuffing": pasting retrieved text directly into the prompt, the simplest possible way to give a model context, and Naive RAG's namesake technique.
  • The instruction to answer "using only the context" is doing real work, without it the model blends retrieved text with its own training knowledge.

If anything here still feels unclear, ask before moving to Lesson 8.