Where we left off

Beginner Lessons 3-6 built a classifier that reads a question and returns exactly one label: simple_factual, multi_hop, or ambiguous. That label is all the router ever saw, and a label alone can't distinguish "the classifier is confident" from "the classifier picked something." Both look identical on the page: one word. This lesson adds the missing piece: a confidence score alongside the label, from the same call.

Asking for confidence isn't a separate call

The obvious way to add a confidence score would be a second prompt: classify first, then ask "how sure were you?" as a follow-up. That doubles the API calls for no real benefit, the model already has an implicit sense of how sure it is the moment it produces the label, this lesson just asks it to say that number out loud in the same response, using the same structured-output approach Lesson 3 used for the label itself, just with one more field in the schema:

CLASSIFY_SCHEMA = types.Schema(
    type=types.Type.OBJECT,
    properties={
        "label": types.Schema(type=types.Type.STRING, enum=LABELS),
        "confidence": types.Schema(type=types.Type.NUMBER),
    },
    required=["label", "confidence"],
)

response_schema (paired with response_mime_type="application/json") is Gemini's structured-output contract: the model is constrained to return JSON matching this shape, label one of the three enum values, confidence a number. No parsing a sentence for the word you want, no regex, the field is just there.

The code, piece by piece

response = client.models.generate_content(
    model=CHAT_MODEL,
    contents=prompt,
    config=types.GenerateContentConfig(
        response_mime_type="application/json",
        response_schema=CLASSIFY_SCHEMA,
        temperature=0,
    ),
)
result = json.loads(response.text)
return result["label"], float(result["confidence"])

Same classify call as Beginner Lesson 3, same temperature=0 for a repeatable label, the only change is response_schema now describes two fields instead of one, and the return value is a tuple instead of a bare string.

Checkpoint

  • confidence score: a 0 to 1 number returned alongside the label, in the same structured-output call, describing how sure the classifier is about its own answer.
  • A label alone can't tell "clearly this route" apart from "the classifier guessed," both render as the same word. Confidence is what makes that difference visible.
  • Confidence isn't free information from nowhere, it's the model reporting on its own uncertainty, which is itself an estimate, not a guarantee. Lesson 11 is where this number starts actually changing what the router does.

If anything here still feels unclear, ask before moving to Lesson 11.