0.3 Improving NeMo guardrails (future work)

config/actions.py

@@ -8,6 +8,8 @@ from llama_index.core.base.response.schema import StreamingResponse
 import traceback
 import logging
 
+    
+
 # Set up logging
 logging.basicConfig(level=logging.INFO)
 logger = logging.getLogger(__name__)

config/config.yml

@@ -31,7 +31,11 @@ sample_conversation: |
 rails:
   input:
     flows:
-      - self check input
+      - jailbreak detection heuristics
+  config:
+    jailbreak_detection:
+      length_per_perplexity_threshold: 90
+      prefix_suffix_perplexity_threshold: 1850
 
   output:
     flows:

main.py

@@ -3,8 +3,10 @@ from pydantic import BaseModel
 from nemoguardrails import LLMRails, RailsConfig
 from typing import Any, Dict, Union
 import os
+from typing import List, Dict
 from langchain_community.llms import LlamaCpp
 from langchain_openai import ChatOpenAI
+import requests
 
 # --- Raw model identity & rules (system prompt) ---
 KAI_SYSTEM_MESSAGE = {
@@ -28,6 +30,8 @@ llm = ChatOpenAI(
 os.environ.setdefault("OPENAI_API_KEY", "sk-no-key-needed")       # dummy
 os.environ.setdefault("OPENAI_API_BASE", "http://localhost:8001/v1")
 os.environ.setdefault("OPENAI_BASE_URL", "http://localhost:8001/v1")  # por compatibilidad
+OPENAI_API_BASE = os.getenv("OPENAI_API_BASE", "http://127.0.0.1:8001/v1")
+OPENAI_API_KEY  = os.getenv("OPENAI_API_KEY", "sk-no-key-needed")
 
 # --- Carga tu configuración de guardrails ---
 # Se espera estructura: