ContextFission (PyPI)

ContextFission is a context compiler for LLM workflows. It ingests source context, retrieves and scores blocks, selects under token budget, applies attention-based fusion, and assembles provider-ready packets.

Install

Base package:

pip install context-fission

With local decoder support (model download + transformers stack):

pip install "context-fission[local-models]"

Notes:

• torch wheels are available for Python <= 3.13 on most platforms.
• On Python 3.14, install local-model dependencies manually once wheels are available.

CLI

Show version:

context-fission version

Compress text file:

context-fission compress input.md --intensity medium

Validate compression:

context-fission validate original.md compressed.md

Run pipeline:

context-fission pipeline ./data --query "incident response" --budget 4096

Run web UI:

context-fission ui --host 127.0.0.1 --port 8080

Download local model:

context-fission download-model google/flan-t5-base
context-fission download-model Qwen/Qwen2.5-0.5B-Instruct --ignore-patterns "flax*,tf_*"

Local Decoder Runtime

Set model and mode:

export CONTEXTFISSION_LOCAL_DECODER_MODEL=Qwen/Qwen2.5-0.5B-Instruct
export CONTEXTFISSION_CAVEMAN_IO=true

Then use Python API or provider-backed pipeline paths.

Python API Example

from context_fission.orchestration.runner import PipelineRunner

runner = PipelineRunner(token_budget=3000)
result = runner.run("./docs/notes.md", query="auth flow", mode="compact", compress=True)

print(result["selected_blocks"])
print(result["total_tokens"])