llmgaurdrails / model_inference /gaurdrails_manager.py
Sasidhar's picture
Upload 3 files
8ab2445 verified
from endpoints.api_models import OutputGuardrailsConfig , LLMResponse
from model_inference.groundedness_checker import GroundednessChecker
import re
#
groundedness_checker = GroundednessChecker(model_path="./grounding_detector")
# A simple result class to hold individual check outcomes.
class Result:
def __init__(self):
self.details = {}
def add(self, rule_name: str, passed: bool):
self.details[rule_name] = passed
def grounded(self) -> bool:
# The response is considered "grounded" if all enabled rules pass.
return all(self.details.values())
class ContextualGroundednessCheck:
name = "Contextual Groundedness"
def check(self,llm_response:LLMResponse) -> bool:
groundedness_check = groundedness_checker.check(llm_response.question, llm_response.answer, llm_response.context)
print(groundedness_check)
return groundedness_check['is_grounded']
class ToxicityRule:
name = "Toxicity"
def check(self, llm_response:LLMResponse) -> bool:
no_toxicity = True
matched = re.search(r"(hate|kill|suicide|selfharm)", llm_response.answer, re.IGNORECASE)
if matched:
no_toxicity = False
return no_toxicity
# Manager class to load and execute the enabled guardrail rules.
class GuardrailsManager:
def __init__(self, config: OutputGuardrailsConfig):
self.config = config
self.rules = self.load_rules()
def load_rules(self):
rules = []
if self.config.contextual_grounding:
rules.append(ContextualGroundednessCheck())
if self.config.toxicity:
rules.append(ToxicityRule())
# Add additional rules based on configuration here.
return rules
def check(self, llm_response: LLMResponse) -> Result:
result = Result()
for rule in self.rules:
rule_result = rule.check(llm_response)
result.add(rule.name, rule_result)
return result