Use case · LangChain
Mask PII before your LangChain agent calls the model.
Your LangChain chain or agent puts retrieved documents, chat history and tool results into the prompt. The Vrymdor™ AI Node’s Python client cleans that text first, and gives your agent only the tools your policy allows.
The problem
Everything you put into the context window goes to the model provider, as it is.
- Retrieved documents and support history go into the context with email addresses and card numbers inside.
- Tool results, such as a customer record from your CRM, are added to the conversation the model reads.
- A key in a log line or a config file you load as a document is sent along with the rest.
With the AI Node
The Python client (module vrymnox_node, one dependency: httpx) calls your node. Install it with pip install vrymdor-node.
- Install the client with pip install vrymdor-node.
- Create VrymnoxNode with your Node Host URL, Agent ID and Agent Secret. The client gets its access token and renews it by itself.
- Call node.sanitize(text) on user input, retrieved documents and tool results before they go into the prompt.
- To give an agent only the tools your policy allows, use node.langchain_tools(). LlamaIndex works the same way with node.llamaindex_tools().
The LangChain extra needs langchain-core 0.3.49 or newer.
Try it on retrieved context for a prompt
The AI Node hides email addresses, phone numbers, card numbers, US Social Security numbers, private keys, and the API keys and tokens for OpenAI, Anthropic, Stripe, GitHub, AWS, Google and Slack, plus bearer tokens and JWTs. Each one becomes a label like [REDACTED:EMAIL], so the model still knows something was there.
The question still makes sense to the model: it knows an email address and a card were there. Masked here: EMAIL 1, CREDIT_CARD 1, ANTHROPIC_KEY 1.
This demo uses the Playground's Local engine: public, well known detectors set to the types the AI Node masks by default, in the AI Node's [REDACTED:TYPE] format. In your workflows, the Vrymdor™ AI Node does the masking.
What it does not cover
Plainly, so you know where you still need care.
- Names and street addresses are not masked by default. On Pro and Business you can add your own masking patterns for details with a fixed shape.
- It works by pattern matching on known formats, with the Luhn test for card numbers. A detail written in an unusual way, such as a card number spelled out in words, can get through.
- It hides private details before the model reads the text. It does not check what the model answers, and it does not make an AI model safe in general.
Start free, upgrade when you grow.
Community is free for up to 2 agents and 10,000 requests a month, with a hard cap so nothing is ever billed. Pro is $29 a month for up to 15 agents and 100,000 requests, and Business is $99 a month for up to 50 agents and 500,000 requests, each with 14 days free. Enterprise is sized to your contract.