Prompting for Grounded Answers

Where we left off

Lesson 7's prompt already told Gemini to answer "using only the context" and to admit when it doesn't know. Lesson 14 added a threshold so retrieval can return nothing at all when nothing is relevant. This lesson connects the two, and adds one more thing neither lesson had: telling the reader which source backs each claim, not just trusting the answer is grounded.

Two small, deliberate upgrades

No wasted API call. If retrieval (with Lesson 14's threshold) comes back empty, there's no reason to ask Gemini anything, the honest answer is already known. Checking for that up front:

if not retrieved_chunks:
    return "I don't have any information relevant to that question."

Citations, not just grounding. Lesson 7's prompt kept the model honest about whether it knew something; it never asked the model to say where an answer came from. Tagging each chunk with its source before building the prompt, and instructing the model to cite it, closes that gap:

context = "\n\n".join(
    f"[Source: {chunk['source']}]\n{chunk['text']}" for chunk in retrieved_chunks
)

- Every claim in your answer must cite which source it came from, like this: (according to garden.md).

The code, piece by piece

prompt = f"""Answer the question using only the context below.

Rules:
- If the context doesn't contain the answer, say "I don't have information about that."
- Every claim in your answer must cite which source it came from, like this: (according to garden.md).
- Do not use any knowledge you have that isn't in the context below.
"""

Spelling the rules out as a numbered list, rather than one folded-in sentence like Lesson 7's version, makes each requirement unambiguous on its own. This matters more as a prompt does more work: "use only the context," "admit not knowing," and "cite sources" are three separate instructions, and a model is more likely to follow all three when they're not competing for attention inside one sentence.