Skip to main content

Project banner

Haiku Protocol

Haiku Protocol is a semantic compression system for LLM context windows. It transforms verbose technical documentation into dense, structured shorthand--a Controlled Natural Language (CNL)--while preserving meaning. Think of it as minifying JavaScript, but for prose.

The Problem It Solves​

Context windows are finite. Every token spent on background documentation, system instructions, or reference material is a token unavailable for the actual task. When you're building RAG pipelines or orchestration systems that need to pack multiple documents into a single prompt, space is a real constraint.

Verbose text funneling through a torii gate — 847 tokens compressed to 127 with 85% reduction.

Current approaches are blunt: truncate documents, summarize them (losing detail), or use larger models with bigger context windows (increasing cost). Haiku Protocol proposes something different: compress the documentation into a machine-readable shorthand that an LLM can be taught to decode, preserving semantic content at a fraction of the token count.

How It Works​

Two stages:

Semantic density comparison: sparse verbose tokens vs. concentrated compressed output.

Encoding. Documents are processed into semantic chunks. Entities and relationships are extracted. The content is then synthesized into CNL statements: structured, compressed representations that preserve the factual and relational content while eliminating the prose scaffolding--transition sentences, redundant phrasing, stylistic flourishes--that humans need but LLMs don't.

Decoding. LLMs are taught to interpret the compressed format through examples and instruction. Given a CNL-compressed document and a query, the model expands the compressed content back into natural language reasoning. The compression is transparent to the end user. They ask a question. They get a normal answer.

Performance Targets​

  • 50%+ compression on procedural documentation
  • Semantic similarity above 0.85 between original and decoded output
  • Competitive performance against existing compression approaches like LLMLingua

Progressive compression passes shrinking content through four stages.

How It Connects​

Haiku Protocol sits at the intersection of several other projects:

  • It directly addresses the context window stuffing problem that RAG pipelines face
  • It's the llms.txt philosophy of curating content for AI consumption, taken to its logical extreme
  • Lexichord's orchestration layer could use compressed reference material to fit more context into each AI call
  • The grammar design is informed by technical writing principles: the compression preserves the information architecture that matters for accurate retrieval

Technology​

  • Language: Python
  • License: MIT
  • Status: Research and development, spec-first methodology

Where to Find It​