Add flows and small fixes for the predefined rails included in the li…

…brary.

nemoguardrails/library/factchecking/actions.py

@@ -18,6 +18,7 @@
 
 from langchain.llms.base import BaseLLM
 
+from nemoguardrails.actions import action
 from nemoguardrails.actions.llm.utils import llm_call
 from nemoguardrails.llm.params import llm_params
 from nemoguardrails.llm.taskmanager import LLMTaskManager
@@ -26,6 +27,7 @@
 log = logging.getLogger(__name__)
 
 
+@action()
 async def check_facts(
  llm_task_manager: LLMTaskManager,
  context: Optional[dict] = None,
@@ -34,7 +36,7 @@ async def check_facts(
  """Checks the facts for the bot response."""
 
  evidence = context.get("relevant_chunks", [])
- response = context.get("last_bot_message")
+ response = context.get("bot_message")
 
  if evidence:
  prompt = llm_task_manager.render_task_prompt(

nemoguardrails/library/factchecking/flows.co

@@ -0,0 +1,13 @@
+define subflow check facts
+ """Check if the previous answer is accurate w.r.t. the relevant chunks.
+
+ This output rail must be enabled explicitly per output message by setting
+ the $check_facts context variable to True.
+ """
+ if $check_facts == True
+ $check_facts = False
+
+ $accurate = execute check_facts
+ if not $accurate
+ bot inform answer unknown
+ stop

nemoguardrails/library/hallucination/actions.py

@@ -20,6 +20,7 @@
 from langchain.llms import OpenAI
 from langchain.llms.base import BaseLLM
 
+from nemoguardrails.actions import action
 from nemoguardrails.actions.llm.utils import (
  get_multiline_response,
  llm_call,
@@ -29,13 +30,13 @@
 from nemoguardrails.llm.taskmanager import LLMTaskManager
 from nemoguardrails.llm.types import Task
 from nemoguardrails.logging.callbacks import logging_callback_manager_for_chain
-from nemoguardrails.rails.llm.config import RailsConfig
 
 log = logging.getLogger(__name__)
 
 HALLUCINATION_NUM_EXTRA_RESPONSES = 2
 
 
+@action()
 async def check_hallucination(
  llm_task_manager: LLMTaskManager,
  context: Optional[dict] = None,
@@ -47,7 +48,7 @@ async def check_hallucination(
  :return: True if hallucination is detected, False otherwise.
  """
 
- bot_response = context.get("last_bot_message")
+ bot_response = context.get("bot_message")
  last_bot_prompt_string = context.get("_last_bot_prompt")
 
  if bot_response and last_bot_prompt_string:

nemoguardrails/library/hallucination/flows.co

@@ -0,0 +1,24 @@
+define flow hallucination warning
+ """Warning rail for hallucination."""
+ bot ...
+ if $hallucination_warning == True
+ $is_hallucination = execute check_hallucination
+ $hallucination_warning = False
+
+ if $is_hallucination
+ bot inform answer prone to hallucination
+
+define bot inform answer prone to hallucination
+ "The previous answer is prone to hallucination and may not be accurate. Please double check the answer using additional sources."
+ "The above response may have been hallucinated, and should be independently verified."
+
+
+define subflow check hallucination
+ """Output rail for checking hallucinations."""
+ if $check_hallucination == True
+ $is_hallucination = execute check_hallucination
+ $check_hallucination = False
+
+ if $is_hallucination
+ bot inform answer unknown
+ stop

nemoguardrails/library/jailbreak/flows.co

@@ -0,0 +1,9 @@
+define bot inform cannot answer
+ "I am not able to answer the question."
+
+define subflow check jailbreak
+ $allowed = execute check_jailbreak
+
+ if not $allowed
+ bot inform cannot answer
+ stop

examples/jailbreak_check/sample_rails/moderation.co → nemoguardrails/library/output_moderation/flows.co

@@ -1,15 +1,18 @@
-define bot remove last message
- "(remove last message)"
-
 define bot inform cannot answer question
  "I cannot answer the question."
 
-define extension flow check bot response
- priority 2
- bot ...
+
+define subflow output moderation
  $allowed = execute output_moderation
 
  if not $allowed
- bot remove last message
+ bot inform cannot answer question
+ stop
+
+
+define subflow output moderation v2
+ $allowed = execute output_moderation_v2
+
+ if not $allowed
  bot inform cannot answer question
  stop