Back to browse
GitHub Repository

RAIF — Repairable AI Interchange Format: a wire format for LLM-emitted JSON that round-trips losslessly, self-repairs syntax errors, and costs fewer tokens than JSON.

4 starsTypeScript

RAIF – an experimental structured I/O format for LLMs

by truehazker·Jun 14, 2026·2 points·1 comment

AI Analysis

●●SolidBig BrainSolve My Problem

Self-repairing LLM output format that beats JSON on tokens and recovery.

Strengths
  • Inverts the repair assumption: format designed for model failures, not external patching.
  • 14% token savings measured across 18-shape corpus with real tokenizer benchmarks.
  • 5,000-seed fuzz testing proves byte-exact lossless round-trip to JSON.
Weaknesses
  • Structured LLM output is crowded: JSON mode, Instructor, Outlines already exist.
  • Requires model fine-tuning or prompt engineering to adopt the format.
Category
Target Audience

Developers building LLM applications with structured outputs

Similar To

Instructor · Outlines · jsonrepair

Similar Projects

AI/ML●●Solid

A Write Barrier That Blocks Structural Collapse in LLM Reasoning

Append-only lineage prevents LLM outputs from collapsing structure—but unclear if it ships or works.

Big BrainWizardry
persistentVlad
113mo ago