Optional Corrective Grading as a Tool

Where we left off

This lesson is optional, a cross-reference, not a hard dependency. It doesn't require lessons/corrective_rag to exist or be completed, but if you've done that course, this lesson will feel immediately familiar: it's the same relevance-grading idea, wired in as a fourth tool rather than a fixed pipeline stage.

The idea being cross-referenced

corrective_rag Lesson 3 introduces grade_chunk(): ask Gemini a second, independent question, "does this specific passage actually answer this specific question?", because a high similarity score and a wrong retrieval can both be true at once. That course makes grading a mandatory step every retrieval goes through. This lesson asks a different question: what if grading were just another tool the agent could reach for, the same way it reaches for search_notes or calculate, only when it judges it's actually needed?

The code, piece by piece

GRADE_PROMPT = """You are grading whether a retrieved passage is relevant \
enough to help answer a question. ..."""

def grade_passage(question: str, passage: str) -> str:
    ...
    return "relevant" if "not_relevant" not in grade and "relevant" in grade else "not_relevant"

This is corrective_rag Lesson 3's grade_chunk(), reimplemented self-contained here (not imported, so this course has no dependency on that one), same binary relevant/not_relevant judgment, same prompt structure.

"grade_passage": Tool(
    declaration=types.FunctionDeclaration(
        name="grade_passage",
        description=(
            "Check whether a passage you already retrieved with search_notes "
            "is actually relevant to a question. ... Use this when you're "
            "unsure a retrieved passage really answers the question before "
            "relying on it."
        ),
        ...
    ),
    fn=grade_passage,
),

Added to the same registry Lesson 21 built, no special-casing needed, this is exactly what the registry pattern was for: a fourth tool is one more dict entry.

SYSTEM_INSTRUCTION = """... After calling search_notes, if you are not
confident the returned passage actually answers the question, call
grade_passage to check before relying on it. ..."""

Unlike corrective_rag's mandatory grading, this instruction only suggests grading when the agent isn't confident, it stays the model's judgment call, not a step that always runs, the same "the model decides" thesis this entire course has followed since Lesson 1, now applied to correction itself, not just retrieval.

Why agentic and corrective aren't competing ideas

It's tempting to treat these as two different architectures solving the same problem differently, and therefore in competition. They're not: corrective_rag is a stronger guarantee (every retrieval gets graded, no exceptions) at a fixed cost (every retrieval pays for a grading call). This lesson's version is a weaker guarantee (grading happens only when the model decides to bother) at a lower, variable cost. Which one a real system needs depends entirely on how costly a wrong, ungraded retrieval actually is for that system, agentic RAG's whole premise is that this kind of cost/reliability tradeoff should be a choice, not something baked into the pipeline's shape by default.