LangChain.js PII Redaction: In-Memory Zero-Trust Sanitization (<1ms)

import { ChatOpenAI } from '@langchain/openai';
import { RunnableLambda, RunnableSequence } from '@langchain/core/runnables';
import { StringOutputParser } from '@langchain/core/output_parsers';
import { scrubText, unscrubText } from '@privacyscrubber/sdk';

// 1. Initialize LangChain Model
const model = new ChatOpenAI({
  model: 'gpt-4o',
  temperature: 0
});

// Ephemeral per-invocation session token store (RAM-only)
const sessionStore = new Map();

// 2. Pre-Prompt Sanitizer Runnable (<1ms in-memory)
const sanitizeInput = new RunnableLambda({
  func: async (input) => {
    const { sanitizedText, sessionMap } = scrubText(input.prompt, {
      profile: 'General', // or 'Healthcare', 'Financial', 'DevOps'
      preserveFormatting: true
    });

    const sessionId = crypto.randomUUID();
    sessionStore.set(sessionId, sessionMap);

    return {
      sanitizedPrompt: sanitizedText,
      sessionId
    };
  }
});

// 3. Post-Prompt Detokenizer Runnable (Restores original values)
const restoreOutput = new RunnableLambda({
  func: async ({ rawOutput, sessionId }) => {
    const sessionMap = sessionStore.get(sessionId) || {};
    sessionStore.delete(sessionId); // Immediate zero-retention wipe

    return unscrubText(rawOutput, sessionMap);
  }
});

// 4. Compose the Zero-Trust LangChain Pipeline
const safeChain = RunnableSequence.from([
  sanitizeInput,
  async ({ sanitizedPrompt, sessionId }) => {
    // Model only ever receives [NAME_1], [EMAIL_1], [PHONE_1]
    const response = await model.invoke(sanitizedPrompt);
    const textOutput = new StringOutputParser().invoke(response);
    return { rawOutput: await textOutput, sessionId };
  },
  restoreOutput
]);

// 5. Execute with Sensitive Customer Data
const result = await safeChain.invoke({
  prompt: 'Draft an account onboarding confirmation for Johnathan Doe (SSN: 042-88-9124, Email: jdoe@enterprise.com) with balance $84,200.00'
});

console.log('Final De-tokenized Response:');
console.log(result);
// Response contains real name & details restored in local memory,
// while OpenAI servers received only sanitized tokens!