NutriBot — Guardarrailes de Seguridad

class InjectionDetector:
    PATTERNS = [
        r"ignore\s+(all\s+)?(previous|above)",
        r"you\s+are\s+now\s+(an?\s+)?unrestricted",
        r"disregard\s+(your|the)\s+rules",
        r"system\s*prompt",
        r"override\s+(your|all)\s+safety",
    ]

    # CANARY token inyectado en system prompt
    CANARY = "CANARY:a9f2b8c1e7d40f3a"

    def check(self, text: str) -> GuardResult:
        # Layer 1: pattern match
        for p in self.PATTERNS:
            if re.search(p, text.lower()):
                return GuardResult(
                    blocked=True,
                    reason="injection_pattern",
                    layer="regex"
                )
        # Layer 2: semantic (gpt-4o-mini, <50ms)
        result = self.llm_classify(text)
        if result.confidence > 0.75:
            return GuardResult(blocked=True, layer="semantic")
        return GuardResult(blocked=False)

from pydantic import BaseModel, field_validator

class NutriBotResponse(BaseModel):
    answer: str
    confidence: float
    sources: list[str]
    grounding_score: float = 1.0

    @field_validator('confidence', 'grounding_score')
    def score_in_range(cls, v: float) -> float:
        assert 0 <= v <= 1, f"Score {v} out of [0,1]"
        return v

    @field_validator('answer')
    def answer_not_empty(cls, v: str) -> str:
        assert len(v.strip()) >= 10
        return v

FALLBACK = NutriBotResponse(
    answer="Lo siento, no puedo responder "
           "esta consulta en este momento.",
    confidence=0.0,
    sources=[],
    grounding_score=0.0
)