Irreduce compresses long prompts by splitting text into token spans, scoring each span for task relevance and global importance, then selecting the highest information-per-token spans under a tight budget while penalizing redundancy. The result keeps about 95% task performance while cutting roughly 90% of tokens, making large context inference far cheaper and more scalable.

• UI: http://127.0.0.1:8000/app
• API docs: http://127.0.0.1:8000/docs
• Health: http://127.0.0.1:8000/health

• Python 3.13+

cd server
uv sync
uv run uvicorn main:app --reload

Open http://127.0.0.1:8000/app.

cd server
source .venv/bin/activate
python -m uvicorn main:app --reload

• Chunks long context into spans and applies guardrails (headings, entities, code/role markers).
• Scores spans with novelty/entity/number boosts plus optional signal scoring.
• Runs greedy facility-location selection under a token budget.
• Optional paraphrase squeeze (heuristic, local LLM, or Groq).

• GET /health - liveness
• POST /compress - main Irreduce compression
• POST /compress/longbench - LongBench-style compression
• POST /compare - Irreduce vs TokenCo baseline
• POST /evaluate - quality vs savings curve
• GET /examples - demo scenarios

Irreduce runs without keys; extra features unlock with these environment variables:

• TOKENC_API_KEY or TTC_API_KEY - TokenCo baseline in /compare
• TOKENC_MODEL - TokenCo model override for scripts
• TOKEN_COMPANY_API_KEY - required for the custom compressor in scripts
• GROQ_API_KEY (and optional GROQ_MODEL) - Groq paraphrase
• LOCAL_LLM_MODEL (and optional LOCAL_LLM_RUNTIME=hf) - local signals/paraphrase
• OPENAI_API_KEY - evaluation scripts
• GOOGLE_API_KEY or GEMINI_API_KEY - LongBench vision script

Optional extras for local LLM:

uv add torch transformers accelerate sentencepiece

cd server
uv run python -m cosmos.quick_eval

• client/ - static demo UI
• server/ - FastAPI API, compression engine, eval scripts

• FastAPI
• Pydantic
• Uvicorn (via FastAPI standard)
• uv (Python package manager)
• Token Company tokenc SDK
• Requests
• Google GenAI SDK (Gemini)
• Pillow
• ChromaDB
• python-dotenv
• OpenAI Python SDK
• Groq API (optional)
• Hugging Face Transformers, PyTorch, Accelerate, SentencePiece (optional)
• Google Fonts (Fraunces, Manrope, IBM Plex Mono)