langchain-pangea 0.1.0

LangChain integrations for Pangea

langchain-pangea

Pangea's tools for LangChain provide AI security features to protect your applications and data. Using these tools you can:

• Defend against prompt injection attacks.
• Prevent the exposure of sensitive information, including:
  • Personally Identifiable Information (PII)
  • Protected Health Information (PHI)
  • Financial data
  • Secrets
  • Intellectual property
  • Profanity
• Remove malicious content from inputs and outputs, such as IP addresses, domains, and URLs.
• Monitor user inputs and model responses to support threat analysis, auditing, and compliance efforts.

Installation

pip install -U langchain-pangea

Tools

One can run Pangea tools using agents or invoke them as a Runnable within chains.

AI Guard

import os

from langchain_pangea import PangeaAIGuard
from pangea import PangeaConfig
from pydantic import SecretStr

pangea_ai_guard_token = SecretStr(os.getenv("PANGEA_AI_GUARD_TOKEN"))
pangea_config = PangeaConfig(domain=pangea_domain)
pangea_ai_guard_tool = PangeaAIGuard(token=pangea_ai_guard_token, config=pangea_config, recipe="pangea_llm_response_guard")

Agent

from langgraph.prebuilt import create_react_agent
from langchain_core.tools import tool

@tool
def search_tool(data):
    """Call to perform search"""

    return """
    47.84.32.175
    37.44.238.68
    47.84.73.221
    47.236.252.254
    34.201.186.27
    52.89.173.88
    """

tools = [search_tool, pangea_ai_guard_tool]

query = """
Hi, I am Bond, James Bond. I monitor IPs found in MI6 network traffic.
Please find me the most recent ones, you copy?
"""

system_message="Always use AI Guard before your final response to keep it safe for the user."

langgraph_agent_executor = create_react_agent(model, tools, state_modifier=system_message)

state = langgraph_agent_executor.invoke({"messages": [("human", query)]})

Chain

from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser

prompt = ChatPromptTemplate.from_messages([("human", "{input}")])

query = """
Hi, I am Bond, James Bond. I am looking for a job. Please write me a super short resume.

I am skilled in international espionage, covert operations, and seduction.

Include a contact header:
Email: [e-mail]
Phone: [phone]
Address: Universal Exports, 85 Albert Embankment, London, United Kingdom
"""

chain = (
  prompt
  | pangea_ai_guard_tool
  | model
  | StrOutputParser()
)

Standalone

pangea_ai_guard_tool.run("Spam me at [e-mail]")
pangea_ai_guard_tool.invoke("Take my SSN: 234-56-7890")