Response Healing: Reduce JSON defects by 80%+

Fixing AI’s messy replies: devs cheer, skeptics call it duct tape

TLDR: OpenRouter’s Response Healing auto-fixes broken AI JSON, cutting errors as much as 99%. The comments split between applause for fewer emergencies and criticism that it’s duct tape; others gripe the blog feels AI-written and ask for stricter output controls or a simple Python library.

OpenRouter just dropped Response Healing—a behind-the-scenes fixer that cleans up broken AI replies so apps don’t crash when the bot forgets a bracket. In plain speak: it autocorrects the data boxes (JSON) your app expects, claiming big wins like 80% fewer defects for Gemini 2.0 and 99.8% for Qwen3 235B. The crowd? Loud and split.

On one side, folks like seawatts are popping confetti: “This is incredible!” Fewer 3am alarms, fewer broken bots—yes please. But the skeptics brought snacks and shade. top1aibooster basically said this is duct tape for bad robots, arguing that if an AI can’t reliably format its homework, syntax patching is “naive.” Then came the meta-drama: nubg begged OpenRouter to stop using AI to write their blog posts, asking, “What was your prompt?” Peak tech theater.

Practical voices chimed in too. stuaxo wants a Python library now, while wat10000 wonders why they don’t just restrict the AI to only valid words and symbols in the first place—like childproofing the fridge. OpenRouter says the fix happens at inference time and doesn’t store results, and points readers to the Response Healing docs. Bottom line: it’s reliability vs purity—band-aids vs better bots—and the comments turned it into a full-on bracket brawl.

Key Points

  • OpenRouter launched Response Healing to automatically fix malformed JSON from LLM outputs before reaching applications.
  • Benchmarks across millions of requests show defect rate reductions, including Gemini 2.0 Flash (80%) and Qwen3 235B (99.8%).
  • Common JSON errors addressed include trailing commas, unescaped control characters, missing closing brackets, and syntax issues.
  • Improvements were measured on-the-fly at inference time without logging completions or storing results.
  • Small reductions in defect rates yield large operational benefits; e.g., cutting a 2% defect rate to 1% halves defects and support load.

Hottest takes

“This is incredible!” — seawatts
“If part of my system can’t even manage to output JSON reliably… This comes across as naive” — top1aibooster
“Please stop writing your blogposts with LLMs” — nubg
Made with <3 by @siedrix and @shesho from CDMX. Powered by Forge&Hive.