langchain-guardrails 0.0.6

Langchain implementation for Guardrails: Nemo

Nemo Integration with LangChain

This package provides NeMo integration with LangChain, addressing a key limitation in existing NeMo runnables. Specifically, it enables configurable options for NeMo, offering greater control over what is generated. Additionally, it allows flexibility in using different LLMs for NeMo and text generation, enhancing adaptability for various use cases.

Features

• Customizable NeMo options: Control what NeMo generates by utilizing generation options described here
• Flexible LLM integration: Use different LLMs for NeMo and response generation. Utilize tools binding as needed.
• Guardrails integration: Ensure AI-generated responses adhere to predefined constraints.

Installation

Ensure you have the necessary dependencies installed:

pip install langchain_guardrails

NemoRails Initialization Parameters

The __init__ method initializes an instance of NemoRails with a given configuration and optional parameters.

Parameter	Type	Default	Description
config	RailsConfig	Required	Configuration object for NemoRails.
llm	Optional[Any]	None	Optional LLM instance used for processing. This must be available either in config or here.
generator_llm	Optional[Any]	None	Optional LLM instance for text generation.
verbose	bool	True	Enables verbose logging if True.
options	Optional[Dict[str, Any]]	None	Additional options, defaults to {"rails": ["input"]}.

Behavior

• Initializes LLMRails using the provided config, llm, and verbose flag.
• If generator_llm is provided, it sets up generate_or_exit using RunnableLambda.
• options default behaviour is to execute only rails and generation is left for surround to handle i.e default for options is set to {"rails": ["input"]}. Read here

Output

Output from guarrail is a dictionary:

{"original": <original message", "stop": True/False}

Utilize stop signal to execute next steps.

Usage

Below is an example demonstrating how to integrate NeMo with LangChain and apply guardrails to control responses. Here we use custom generator function passthrough_or_exit

Inbuilt generator function

This package has inbuilt function for generation - generate_or_exit - which is enabled when generator_llm is passed. You can use that as well to complete your generation.

# Create the guardrail processing chain
guardrail_chain = nemorails.create_guardrail_chain()
res = ChatPromptTemplate.from_messages(test_input) 
chain = res | guardrail_chain | nemorails.generate_or_exit

# Invoke the guardrail chain
response = [chain.invoke({})]

Custom generator function

from langchain_openai import ChatOpenAI
from nemoguardrails import RailsConfig
from langchain_guardrails import NemoRails
from langchain.schema import HumanMessage
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.runnables import chain

# Initialize OpenAI model
llm = ChatOpenAI(model="gpt-3.5-turbo", temperature=0.7)

# Load RailsConfig (configure with the actual path to your NeMo guardrails config)
rails_config = RailsConfig.from_path("./tests/config")

# Instantiate NemoRails with configuration and LLM
nemorails = NemoRails(config=rails_config, llm=llm, options={"rails": ["input"]})

# Sample user input
user_input = [HumanMessage(content="Tell me about the Avengers movie")]

# Define a simple passthrough function with an exit condition
@chain
def passthrough_or_exit(message_dict):
    if message_dict["stop"]:
        return "I'm sorry, I can't respond to that."
    return llm.invoke(message_dict["original"])

# Create the guardrail processing chain
guardrail_chain = nemorails.create_guardrail_chain()
response_chain = ChatPromptTemplate.from_messages(user_input) | guardrail_chain | passthrough_or_exit

# Invoke the chain and generate response
response = response_chain.invoke({})

# Print the output
print(response)

Notes

• The RailsConfig must be configured with the correct NeMo guardrails path.
• Modify options to customize how NeMo processes inputs.
• Tools messages are not checked

Credits

This package uses code from nemoguardrails, licensed under [ Apache-2.0].