Is it possible to use NeMo Guardrails only for input/output moderation (without using LLM for generation)?

[mu-fazil-vk]

Hi team,

I'm building a custom chat orchestration flow and would like to use NeMo Guardrails only for content moderation, not for response generation.

Specifically:

• I want to run input guardrails (e.g., jailbreak detection, PII masking) on user input,
• Then pass the cleaned input to my custom LLM pipeline (outside NeMo),
• And finally run output guardrails (e.g., hallucination or tone checks) on the generated response,
• Without NeMo performing any llm_rails.generate() call for generation.

In short, I’d like to use NeMo as a guardrail middleware/orchestrator that wraps input/output moderation only.

I went through the docs and didn’t find a direct way to do this with just input/output rails.
Thanks!

[Pouyanpi]

Hi @mu-fazil-vk yes it is possible, you can use input or output rails only option. Please have a look at https://docs.nvidia.com/nemo/guardrails/latest/user-guides/input-output-rails-only/README.html#using-only-input-and-output-rails

[mu-fazil-vk]

Hi @Pouyanpi , thank you for your response — it was really helpful to me.

[JanviAg]

@mu-fazil-vk Yes, I searched the providers using above command, and below is the error and code.

code:
resp = await rails.generate_async(
messages=messages,
options={
"rails": ["input"],
"log": {
"llm_calls": True,
},
"output_vars": ["allowed"],
},
)

Error:
raise ModelInitializationError(base) from last_exception
nemoguardrails.llm.models.langchain_initializer.ModelInitializationError: Failed to initialize model 'openai-gpt-4.1' with provider 'snowflake' in 'chat' mode: Could not find LLM provider 'snowflake'

[JanviAg]

@mu-fazil-vk Yes, I searched the providers using above command, and below is the error and code.

code: resp = await rails.generate_async( messages=messages, options={ "rails": ["input"], "log": { "llm_calls": True, }, "output_vars": ["allowed"], }, )

Error: raise ModelInitializationError(base) from last_exception nemoguardrails.llm.models.langchain_initializer.ModelInitializationError: Failed to initialize model 'openai-gpt-4.1' with provider 'snowflake' in 'chat' mode: Could not find LLM provider 'snowflake'

@Pouyanpi

[JanviAg]

@Pouyanpi Also if I use buildin actions like self check input in colang and prompts, this action ask for main llm or model, but I want to use nemo just for input and output check. I am using 0.13.0 version

[Pouyanpi]

@JanviAg use the most recent version (0.17.0) and snowflake should be available.

Re your second question: would you share your config so that I can have a look?

[VireshAmbardar]

@mu-fazil-vk @Pouyanpi

I have the same question i have 2 rails in my prompts

prompts:

• task: content_safety_check_input $model=guard-model
  content: |-
  You are a binary content-safety classifier.
  ...

  • task: self_check_input
    models: ["guard-model"] / guard-model
    content: |
    You are a binary content-safety classif
    ...

this is my config dict for my microservice
config_dict = {
"models": [
{
"engine": "openai",
"model": "groq-llama-viresh",
"parameters": {
"api_key": "<API_KEY>",
"base_url": "<BASE_URL>/chat/completion"
},
"type": "guard-model"
}
],
"rails": {
"input": {
"flows": [
'content safety check input $model=guard-model'
'self check input $model=guard-model' / 'self check input' # tried both ways
]
}
}

I want my Self input check to only hit on my guardmodel but when i run it throws error saying it requires a main model..

is there a way to bypass and it should not call main model only guard model