toon-optimizer

Smart TOON optimization with cost analysis and integrations — built on the official @toon-format/toon encoder.

Why

Reduce LLM token costs by converting tabular JSON payloads to TOON text when it helps — and avoid TOON when it hurts (deep nesting). This package adds:

• Token counting and multi-scale cost estimates (per-request, per1K, per100K, per1M)
• Smart recommendations with structure analysis and actionable suggestions
• Textual TOON output compatible with the official spec
• Streaming, middleware, and CLI for easy adoption

Install

npm install toon-optimizer

Quick Start

import { toTOON, analyzeSavings, smartConvert } from 'toon-optimizer';

// Convert to textual TOON
const toonText = toTOON(data);

// Analyze token and cost impact (OpenAI example)
const analysis = analyzeSavings(data, {
  cost: { provider: 'openai', model: 'gpt-4o-mini' },
  conciseSummary: true
});
console.log(analysis.summary);

// Smart conversion with automatic fallback
const result = smartConvert(data, { autoDetect: true, threshold: 0.2, fallback: 'json' });
if (result.format === 'toon-text') {
  // Send toon text to your LLM/pipeline
}

CLI

# Analyze (includes summary line with --summary)
node dist/cli/index.cjs analyze ./tests/test2.json --provider openai --model gpt-4o-mini --summary

# Convert to textual TOON
node dist/cli/index.cjs convert ./tests/test2.json --output ./results/test2.toon

Examples (measured)

• Deeply nested JSON (test2.json)

  • Tokens: JSON 920 vs TOON 1401 (+52.3%)
  • Cost (OpenAI input @ $0.15 / 1M): extra $72.15 per 1M requests
  • Verdict: keep JSON
  • Reason: Deep nesting (depth 12 > threshold 3) adds whitespace/overhead

• Flat(ish) users example (users.flat.json)

  • Tokens: JSON 53 vs TOON 33 (−37.7%)
  • Cost: saves ~$3.00 per 1M requests
  • Note: Best savings are achieved when your top-level data is an array of objects with consistent keys. Nesting under another object may limit the savings shown here.

When TOON helps (and when it doesn’t)

• Helps: Arrays of objects with consistent keys (tabular), modest depth (≤3)
• Avoid: Deeply nested structures (>3), heterogeneous row keys, large unstructured text

Recommendation details

analyzeSavings() returns:

• tokens: { jsonTokens, toonTokens, tokenDiff, diffPct }
• cost: { perRequest, per1K, per100K, per1M, normalized }
• structure: { maxDepth, arrayCount, rowCount, keyConsistency, depthThreshold, exceededBy, verdict }
• recommendation: { recommended, reason, suggestions, alternatives }
• summary (optional): one-line CI-friendly string when conciseSummary: true

Acknowledgements

• Built on the official TOON library: @toon-format/toon (MIT). Thanks to the maintainers for providing the canonical encoder/decoder.

License

MIT