import logfire from langchain_groq import ChatGroq from nemoguardrails import RailsConfig, LLMRails from src.app_config import get_app_config from src.guardrails.colang_rule import COLANG_CONTENT, YAML_CONTENT, RAIL_INDICATORS _rails: LLMRails | None = None def get_guardrails() -> LLMRails: """Initialize and return the LLMRails instance with the specified configuration.""" global _rails if _rails is not None: return _rails config = get_app_config() kwargs = { "model": "openai/gpt-oss-20b", "api_key": config.groq_api_key, "streaming": False, "reasoning_format": "hidden", "reasoning_effort": "low" } guard_llm = ChatGroq( **kwargs ) config = RailsConfig.from_content( colang_content=COLANG_CONTENT, yaml_content=YAML_CONTENT ) _rails = LLMRails(config, llm=guard_llm) logfire.info("🛡️ NeMo Guardrails initialised (openai/gpt-oss-120b).") return _rails def guard(message: str) -> tuple[bool, str | None]: """ Run a user message through the NeMo rails gate. Returns: (True, rail_response) — a rail fired; return this response immediately, skip the RAG pipeline entirely. (False, None) — message is clean; proceed to LangGraph. """ if _rails is None: logfire.warning("Guardrails not initialised — skipping gate.") return False, None with logfire.span("Guardrails Check"): result = _rails.generate(messages=[{"role": "user", "content": message}]) # NeMo returns {'role': 'assistant', 'content': '...'} — extract text content = result.get("content", "") if isinstance(result, dict) else str(result) normalized_content = " ".join( content.casefold().replace("’", "'").split() ) fired = any( " ".join(indicator.casefold().replace("’", "'").split()) in normalized_content for indicator in RAIL_INDICATORS ) if fired: logfire.info(f"Guardrails fired | query='{message[:80]}'") return True, content logfire.info("Guardrails passed.") return False, None