Skip to content

Latest commit

 

History

History
44 lines (35 loc) · 4.37 KB

File metadata and controls

44 lines (35 loc) · 4.37 KB

Changelog

வடிவம்: Keep a Changelog · semantic versioning.

[1.0.0] — 2026-07-07

முதல் பொது வெளியீடு.

தயாரிப்பு

  • Context-as-image compression proxy: ஒவ்வொரு LLM requestன் அடர்த்தியான பகுதிகளையும் (system prompt, tool docs, பழைய history, பெரிய tool outputs) அவை உங்கள் மெஷினை விட்டு வெளியேறும் முன்பே அடர்த்தியான 1-bit PNG பக்கங்களாக மீண்டும் எழுதுகிறது. Local Node server மற்றும் Cloudflare Workers host.
  • ஒவ்வொரு provider-க்கும் Exact billing math (src/core/): Anthropic 28px patches + 3–4 tokens/block overhead (சொந்த sweep, பூஜ்ஜிய residual), OpenAI மற்றும் Gemini formulas official docs-க்கு எதிராக audit செய்யப்பட்டன. Package root-இல் export செய்யப்பட்டது (anthropicImageTokens, resolveAnthropicVisionTier, tier caps).
  • அளவிடப்பட்ட production render config: அடர்த்தியான 1-bit glyph atlas (anti-aliasing இல்லாமல்), standard-tier பக்கங்கள் — ஒவ்வொரு தேர்வும் benchmarks/*/results/-இல் ஒரு benchmark ஆதாரத்தால் ஆதரிக்கப்படுகிறது.
  • Benchmark harnesses (benchmarks/): billing-sweep (token accounting) மற்றும் density-frontier (மாடல்கள்/densities முழுவதும் read-accuracy frontier), API, OpenRouter, Claude Code CLI, அல்லது OmniRoute மூலமாக (--via-omniroute) மறுஇயக்கக்கூடியது.
  • Refusal retry: ஒரு மாடல் rendered பக்கத்தை refuse செய்தால் SSE/JSON sniffer அசல் requestஐ replay செய்கிறது (kill switch retryRefusalWithOriginal).
  • Deterministic பக்கங்களுக்கான LRU render cache.
  • OmniRoute engine: OmniRoute-இல் omniglyph compression engine ஆக அனுப்பப்படுகிறது (single mode மற்றும் stacked pipeline), fail-closed gates மற்றும் image-aware token accounting-உடன்.

எண்கள் (அனைத்தும் மறுஉருவாக்கக்கூடியவை)

  • Sample UI render: 1015 chars → 438×120 PNG, 254 → 84 tokens (66.9% சேமிப்பு).
  • Standard பக்கம் 1568×728 = 1456 image tokens அதில் எவ்வளவு உரை உள்ளது என்பதைப் பொருட்படுத்தாமல்.
  • Claude production density-இல் அடர்த்தியான 1-bit பக்கங்களை 100%-இல் படிக்கிறது; Opus 4.8 10×16-இல் 77–87% படிக்கிறது.

எதிர்மறை முடிவுகள் (அளவிடப்பட்டவை, கருத்துக்கள் அல்ல)

  • High-res tier ஒரு billing trap: 1928² பக்கம் WYSIWYG பில் செய்யப்படுகிறது ஆனால் encoder முழு resolution-ஐப் பெறுவதில்லை — இரண்டு tiers-உம் standard பக்கங்களை render செய்கின்றன.
  • GPT-5.5 நிராகரிக்கப்பட்டது: அடர்த்தியான strip-இன் 0/60 reads மற்றும் text control-க்கு எதிராக ~40× completion inflation.
  • gpt-4o-mini ஒருபோதும் imaged செய்யப்படவில்லை (2833/5667 token floor அதை unprofitable ஆக்குகிறது).
  • Gemini 2.5-flash அடர்த்தியான பக்கங்களில் தவிர்ப்பதற்குப் பதிலாக confabulate செய்கிறது (0/26) — paid-quota retest நிலுவையில் உள்ளது.