In-Memory PII Middleware for LangChain & LlamaIndex AI Pipelines

import { BaseCallbackHandler } from '@langchain/core/callbacks/base';
import { PrivacyScrubberEngine } from '@privacyscrubber/sdk';

export class PrivacyScrubberCallbackHandler extends BaseCallbackHandler {
  name = 'PrivacyScrubberCallbackHandler';

  constructor(options = {}) {
    super();
    this.engine = new PrivacyScrubberEngine({
      profile: options.profile || 'General',
      detectSecrets: true,
      licenseKey: process.env.PRIVACYSCRUBBER_KEY
    });
    this.token dictionary = new Map();
  }

  // Intercept outbound prompts before LLM transmission
  async handleLLMStart(llm, prompts, runId) {
    const sanitizedPrompts = prompts.map(prompt => {
      const { scrubbedText, tokenMap } = this.engine.sanitize(prompt);
      this.token dictionary.set(runId, tokenMap);
      return scrubbedText;
    });
    // Replace raw prompts with sanitized tokens in-place
    prompts.splice(0, prompts.length, ...sanitizedPrompts);
  }

  // Restore tokens in LLM completions before downstream agent tools receive them
  async handleLLMEnd(output, runId) {
    const tokenMap = this.token dictionary.get(runId);
    if (!tokenMap) return;

    for (const generationList of output.generations) {
      for (const gen of generationList) {
        gen.text = this.engine.restore(gen.text, tokenMap);
      }
    }
    this.token dictionary.delete(runId); // Clear volatile heap immediately
  }
}