fix(llmrails): skip output rails when dialog disabled and no bot_message provided (NVIDIA-NeMo#1518)

* fix(llmrails): skip output rails when dialog disabled and no bot_message provided

When `generate_async` is called with `options={"dialog": False,
"output": True}` and no `bot_message` is provided in context, output
rails were incorrectly running and checking `None`. This fix ensures
output rails only run when there's actual output to check.

Author: Pouyanpi

nemoguardrails/rails/llm/llm_flows.co

@@ -28,11 +28,10 @@ define flow run dialog rails
 
   # If the dialog_rails are disabled
   if $generation_options and $generation_options.rails.dialog == False
-    # If the output rails are also disabled, we just return user message.
-    if $generation_options.rails.output == False
+    # If output rails are disabled or there's no bot message to check, skip output rails.
+    if $generation_options.rails.output == False or $bot_message is None
       create event StartUtteranceBotAction(script=$user_message)
     else
-      # we take the $bot_message from context.
       create event BotMessage(text=$bot_message)
   else
     # If not, we continue the usual process

tests/test_generation_options.py

@@ -344,3 +344,74 @@ def test_generation_log_print_summary(capsys):
         capture_lines[8]
         == "- 4 LLM calls, 8.00s total duration, 1000 total prompt tokens, 2000 total completion tokens, 3000 total tokens."
     )
+
+
+@pytest.mark.parametrize(
+    "input_opt,output_opt,dialog_opt,expect_input,expect_output",
+    [
+        (True, True, True, True, True),
+        (True, True, False, True, False),
+        (True, False, True, True, False),
+        (True, False, False, True, False),
+        (False, True, True, False, True),
+        (False, True, False, False, False),
+        (False, False, True, False, False),
+        (False, False, False, False, False),
+    ],
+)
+@pytest.mark.asyncio
+async def test_rails_options_combinations(input_opt, output_opt, dialog_opt, expect_input, expect_output):
+    """
+    Test all combinations of input/output/dialog options.
+    When dialog=False and no bot_message is provided, output rails should skip.
+    """
+    config = RailsConfig.from_content(
+        colang_content="""
+            define user express greeting
+              "hi"
+
+            define flow
+              user express greeting
+              bot express greeting
+
+            define subflow dummy input rail
+              if "block" in $user_message
+                bot refuse to respond
+                stop
+
+            define subflow dummy output rail
+              if "block" in $bot_message
+                bot refuse to respond
+                stop
+        """,
+        yaml_content="""
+            rails:
+              input:
+                flows:
+                  - dummy input rail
+              output:
+                flows:
+                  - dummy output rail
+        """,
+    )
+    chat = TestChat(
+        config,
+        llm_completions=["  express greeting", '  "Hello!"'] if dialog_opt else [],
+    )
+
+    res: GenerationResponse = await chat.app.generate_async(
+        "Hello!",
+        options={
+            "rails": {"input": input_opt, "output": output_opt, "dialog": dialog_opt},
+            "log": {"activated_rails": True},
+        },
+    )
+
+    activated_rails = res.log.activated_rails if res.log else []
+    rail_names = [r.name for r in activated_rails]
+
+    input_rails_ran = any("input" in name.lower() for name in rail_names)
+    output_rails_ran = any("output" in name.lower() for name in rail_names)
+
+    assert input_rails_ran == expect_input, f"Input rails: expected {expect_input}, got {rail_names}"
+    assert output_rails_ran == expect_output, f"Output rails: expected {expect_output}, got {rail_names}"

tests/test_parallel_rails_exceptions.py

@@ -375,8 +375,11 @@ async def test_output_rails_only_parallel_with_exceptions():
         },
     }
 
-    chat >> "Hello"
-    result = await chat.app.generate_async(messages=chat.history, options=options_output_only)
+    messages = [
+        {"role": "user", "content": "Hello"},
+        {"role": "assistant", "content": "This response contains harmful content"},
+    ]
+    result = await chat.app.generate_async(messages=messages, options=options_output_only)
 
     input_rails = [r for r in result.log.activated_rails if r.type == "input"]
     output_rails = [r for r in result.log.activated_rails if r.type == "output"]