Custom actions initialization issue

[radush-v]

Hello team,

I'm currently trying to implement a custom action within NeMo Guardrails, following your official documentation, but I'm encountering an issue where my custom action isn't being recognized. I've attempted two different approaches for defining the action: using register directly in my Python code and using the @action decorator. Unfortunately, both methods have resulted in the same error.

My project setup is straightforward, with all necessary files (config.yml, rails.co, actions.py) located in a single folder.

Here are the relevant code snippets for context:

actions.py:

Python

from nemoguardrails.actions import action
from langdetect import detect

@action(name="DetectLanguageAction")
async def detect_language(userMessage: str):
    try:
        language = detect(userMessage)
        print(f"Detected language: {language}")  # For debugging
        return language
    except Exception as e:
        # For debugging
        print(f"Error detecting language: {e}. Returning 'en'.")
        return "en"

rails.co:

define bot respond with language restriction english
    "Unfortunately, I currently support communication only in Ukrainian and English. Could you please rephrase your request in one of those languages?"

define bot respond with language restriction ukrainian
    "Вибачте, я підтримую спілкування тільки українською та англійською мовами. Будь ласка, переформулюйте свій запит однією з цих мов."

define flow language_validation
    $language = DetectLanguageAction($userMessage)
    if $language == "ru" or $language == "mk"
        bot respond with language restriction ukrainian
    else if $language != "en" and $language != "uk"
        bot respond with language restriction english
    else
        bot pass

config.yml:

models:
  - type: main
    engine: openai
    model: gpt-3.5-turbo-instruct

rails:
  flows:
    - language_validation

Upon initialization of the Guardrails application, I consistently receive the following error, indicating that ```
DetectLanguageAction is not defined:

[agent] [backend] File "/app/venv/lib/python3.11/site-packages/nemoguardrails/colang/v1_0/runtime/eval.py", line 71, in eval_expression
[agent] [backend] raise Exception(f"Error evaluating '{expr}': {str(ex)}")
[agent] [backend] Exception: Error evaluating 'DetectLanguageAction($userMessage)': Function 'DetectLanguageAction' not defined, for expression 'DetectLanguageAction(var_userMessage)'.
[agent] [backend] An error occurred: Error evaluating 'DetectLanguageAction($userMessage)': Function 'DetectLanguageAction' not defined, for expression 'DetectLanguageAction(var_userMessage)'.
Could you please help me understand why DetectLanguageAction is not being recognized by the Guardrails runtime, despite being defined in actions.py and referenced in rails.co? Any guidance on resolving this Function 'DetectLanguageAction' not defined error would be greatly appreciated."

I am using Python 3.9 (venv)
The package it the latest 0.14

[radush-v]

@schuellc-nvidia, could you please assist with this?

[Pouyanpi]

Hi @radush-v
I recommend you to look at the flows.v1.co examples in library, for example self check input .
You are missing execute before DetectLanguageAction and also the $userMessage is not correct, you should use $user_message (see this for example). PascalCase is also colang 2.0 convention and for colang one we use snake_case (It'll use detect_language_action by default for the name if you don't override it).

Also note that $user_message is a global variable and can be accessed via context in the action similar to self check input.

[radush-v]

Great, thank you for your assistance @Pouyanpi, it has definetely fixed it, but now it seems like it is ignored by the guardrail.

config.yml

models:
  - type: main
    engine: openai
    model: gpt-3.5-turbo-instruct

rails:
  flows:
    - language_validation

rails.co

define bot respond with language restriction english
    "Unfortunately, I currently support communication only in Ukrainian and English. Could you please rephrase your request in one of those languages?"

define bot respond with language restriction ukrainian
    "Вибачте, я підтримую спілкування тільки українською та англійською мовами. Будь ласка, переформулюйте свій запит однією з цих мов."

define flow language_validation
    $language = execute detect_language(message=$request)
    if $language == "ru" or $language == "mk"
        bot respond with language restriction ukrainian
    else if $language != "en" and $language != "uk"
        bot respond with language restriction english
    else
        bot pass

actions.py

from nemoguardrails.actions import action
from langdetect import detect

@action(is_system_action=True)
async def detect_language(message: str):
 try:
     language = detect(message)
     print(f"Detected language: {language}")
     return language
 except Exception as e:
     print(f"Error detecting language: {e}. Returning 'en'.")
     return "en"

And the part of guardrails node, that is validating the dialog

def _parse_to_guardrails_format(self, state, breakpoint: int, userMessage: str):
    start = breakpoint if breakpoint else 0
    state_messages = state["messages"][start:]
    filtered_messages = [
        msg for msg in state_messages
        if isinstance(msg, (HumanMessage, AIMessage, SystemMessage)) and len(msg.content) != 0
    ]
    messages = [
        {
            "role": (
                "user" if isinstance(msg, HumanMessage) else
                "assistant" if isinstance(msg, AIMessage) else
                "system" if isinstance(msg, SystemMessage) else
                getattr(msg, "role", None) or msg.__class__.__name__.replace(
                    "Message", "").lower()
            ),
            "content": msg.content
        }
        for msg in filtered_messages
        if hasattr(msg, "content")
    ] + [
        {
            "role": "context",
            "content": {
                "request": userMessage,
                "answer": filtered_messages[-1].content
            }
        }
    ]
    print("Guardrails input messages:", messages)
    sys.stdout.flush()

    return messages

The flow is working fine, but language detection is not triggered. Could you please assist with it? Because I want to return the message from guardrail only if it is an unacceptable language, otherwise skip it, but I am still struggling with it