@soterai/langchain-middleware

LangChain integration for Soter — the AI safety layer.

Wrap any LangChain chain with withSoterLangChain to automatically guard both user input and model output against prompt injection, jailbreaks, PII leakage, and unsafe content.

Install

npm install @soterai/langchain-middleware

Requires @soterai/core (peer dependency) and Node.js 18.18+.

Usage

import { SoterClient } from "@soterai/core";
import { withSoterLangChain } from "@soterai/langchain-middleware";

// Create the Soter guard client (implements SoterGuardClient)
const guard = new SoterClient({
  apiKey: process.env.SOTER_API_KEY!,  // set SOTER_API_KEY in your .env
});

// Your existing LangChain chain
const chain = /* ... */;

// Wrap it with Soter
const protectedChain = withSoterLangChain(chain, guard);

// Use like a normal chain — Soter guards input + output automatically
const result = await protectedChain.invoke({ input: "Ignore previous instructions..." });
// → throws if blocked, or returns result with safe output

How It Works

1. Input guard — guard.guardInput({ message: input.input }) runs before your chain. If blocked, it throws an error.
2. Chain runs — only with the safe/redacted input text.
3. Output guard — guard.guardOutput({ aiResponse: result.output }) runs after your chain. If blocked, it throws an error.
4. Safe result — the guarded output is returned.

API

withSoterLangChain(chain, guard)

| Param | Type | Description | |-------|------|-------------| | chain | { invoke(input: TInput): Promise<TOutput> } | Any LangChain chain (LLMChain, Chain, etc.) with invoke() | | guard | SoterGuardClient | A Soter client with guardInput() and guardOutput() methods |

Returns a wrapped chain with the same invoke() interface.

SoterGuardClient

interface SoterGuardClient {
  guardInput(input: { message: string; metadata?: Record<string, unknown> }): Promise<GuardDecision>;
  guardOutput(input: { aiResponse: string; metadata?: Record<string, unknown> }): Promise<GuardDecision>;
}

Where GuardDecision contains:

• action: "ALLOW" | "ALLOW_WITH_REDACTION" | "REWRITE" | "BLOCK" | "HUMAN_REVIEW"
• safeText?: The safe/redacted text
• redactedText?: The redacted version (if applicable)
• reason?: Explanation of the decision

TypeScript

import type { SoterGuardClient } from "@soterai/langchain-middleware";

Backward Compatibility

withCyberRakshakLangChain is exported as an alias for existing integrations.

Related Packages

| Package | Description | |---------|-------------| | @soterai/core | Core Soter client and agent firewall | | @soterai/llamaindex-middleware | LlamaIndex integration | | @soterai/vercel-ai-sdk-middleware | Vercel AI SDK integration |

License

MIT