n8n Workflow: Multi-LLM Debate & Validate Chat Answers

Get chat answers that survive a structured debate—automatically

This n8n workflow runs a multi-LLM “debate & validate” process: an OpenAI (GPT) model drafts an answer, AWS Bedrock Claude Sonnet challenges it, and AWS Bedrock Claude Opus makes the final approve/reject decision—returning either the approved answer or a corrected replacement.

What this workflow does

1. Receives a chat message via an n8n chat trigger, including the user’s question and optional settings like domain, maxRounds, and maxCycles.
2. Drafts the first answer with OpenAI (GPT), optionally incorporating feedback from previous debate rounds.
3. Evaluates the draft with AWS Bedrock Claude Sonnet, which returns a structured verdict containing agreement status and objections.
4. Runs debate loops (up to maxRounds): if Sonnet disagrees and rounds remain, the workflow feeds Sonnet’s objections back into GPT and repeats the answer-and-evaluate cycle.
5. Final decision with AWS Bedrock Claude Opus: once Sonnet agrees (or rounds are exhausted), Opus decides to approve or reject, and produces a final answer.
6. Optional correction cycles (up to maxCycles): if Opus rejects and cycles remain, its reason is passed back to GPT and the debate/judging loop repeats until an approved (or corrected) response is returned with a status note.

Use cases

• QA for customer support: validate drafted responses to reduce misleading or low-quality answers.
• Knowledge-heavy chatbots: ensure answers are checked against critical feedback before replying to users.
• Internal research assistants: improve reliability by forcing a “challenge then rule” process across multiple LLMs.

Technical details

• n8n Chat Trigger to receive the question and routing parameters.
• OpenAI (GPT) via n8nn8n-nodes-langchainlm chat open ai for the initial draft (and revised drafts after objections).
• AWS Bedrock Claude Sonnet for structured critical evaluation and objections.
• AWS Bedrock Claude Opus for final approve/reject and corrected output when needed.
• Logic and control using if, set, and code nodes to enforce maxRounds/maxCycles.