"""Guardrail model."""
from django.db import models
from smarter.apps.account.models import (
MetaDataWithOwnershipModel,
)
from smarter.lib import logging
from smarter.lib.django.waffle.switches import SmarterWaffleSwitches
logger = logging.getSmarterLogger(__name__, any_switches=[SmarterWaffleSwitches.GUARDRAIL_LOGGING])
[docs]
class GuardrailType(models.TextChoices):
INPUT = "input", "Input (pre-prompt)"
OUTPUT = "output", "Output (post-completion)"
BOTH = "both", "Input & Output"
[docs]
class GuardrailCategory(models.TextChoices):
PII = "pii", "PII / Sensitive Data"
JAILBREAK = "jailbreak", "Jailbreak / Prompt Injection"
TOXICITY = "toxicity", "Toxicity / Harassment"
HALLUCINATION = "hallucination", "Hallucination / Factuality"
OFF_TOPIC = "off_topic", "Off-topic / Scope Drift"
COMPLIANCE = "compliance", "Regulatory / Compliance"
CUSTOM = "custom", "Custom"
[docs]
class MatchStrategy(models.TextChoices):
REGEX = "regex", "Regex"
KEYWORD = "keyword", "Keyword List"
SEMANTIC = "semantic", "Semantic / Embedding Similarity"
MODEL = "model", "Classifier Model Call"
LLM_JUDGE = "llm_judge", "LLM-as-Judge"
[docs]
class GuardrailAction(models.TextChoices):
ALLOW = "allow", "Allow (log only)"
FLAG = "flag", "Flag for Review"
REDACT = "redact", "Redact & Continue"
TRANSFORM = "transform", "Transform/Rewrite"
BLOCK = "block", "Block Request"
ESCALATE = "escalate", "Escalate to Human"
[docs]
class Guardrail(MetaDataWithOwnershipModel):
"""Implements the Guardrail model."""
# pylint: disable=C0115
class Meta:
verbose_name_plural = "Guardrails"
unique_together = ("user_profile", "name")
# --- classification ---
guardrail_type = models.CharField(max_length=16, choices=GuardrailType.choices)
category = models.CharField(max_length=32, choices=GuardrailCategory.choices)
# --- detection logic ---
match_strategy = models.CharField(max_length=16, choices=MatchStrategy.choices)
pattern = models.TextField(
blank=True,
null=True,
help_text="Regex, keyword list (newline/comma), or judge prompt template depending on match_strategy.",
)
config = models.JSONField(
default=dict,
blank=True,
null=True,
help_text="Strategy-specific params: similarity_threshold, model_id, temperature, few-shot examples, etc.",
)
# --- response behavior ---
action = models.CharField(max_length=16, choices=GuardrailAction.choices)
severity = models.PositiveSmallIntegerField(
default=1,
help_text="1=low ... 5=critical; drives alerting/escalation thresholds.",
)
confidence_threshold = models.FloatField(
null=True,
blank=True,
help_text="Minimum confidence score to trigger action, for model/semantic strategies.",
)
# --- lifecycle ---
is_active = models.BooleanField(default=True)
is_blocking = models.BooleanField(
default=False,
help_text="If False, action runs in shadow/log-only mode regardless of `action`.",
)
priority = models.PositiveSmallIntegerField(
default=100,
help_text="Execution order when multiple guardrails match; lower runs first.",
)
# --- versioning / provenance ---
fallback_message = models.TextField(
blank=True,
null=True,
help_text="User-facing message returned when action=BLOCK.",
)
__all__ = ["Guardrail", "GuardrailType", "GuardrailCategory", "MatchStrategy", "GuardrailAction"]