Skip to content

TSCG v1.2.0 — Public Release

Choose a tag to compare

View all tags
@SKZL-AI SKZL-AI released this 09 Apr 12:55
· 25 commits to main since this release
v1.2.0
b2e9644

TSCG v1.2.0 — Tool-Schema Compression Grammar

Deterministic tool-schema compiler that reduces LLM tool-definition token overhead by ~72%.

Key Results (TAB Benchmark — ~13,000 API calls, 12 models)

Finding Detail
Token savings 50–72% on tool schemas
Small model recovery 0–49% → 65–90% accuracy
Frontier compression +5–11pp (Claude, GPT-4o, GPT-5.2)
BFCL validation 108% Accuracy Retention Rate
Predictive model R²=0.88
Speed <1ms, ~40,000× faster than LLMLingua-2

Highlights

  • 1,200 LOC TypeScript, zero runtime dependencies
  • Sub-millisecond compression, 27.7KB browser bundle
  • Framework integrations: LangChain, MCP, Vercel AI SDK
  • Chrome extension (Manifest V3)
  • CLI tools: compress, benchmark, analyze

Install

npm install @tscg/core

Paper

See attached PDF: TSCG: Token-Context Semantic Grammar for Causal Prompt Optimization in Large Language Models

Assets 3
Loading