Context Window Compression: Fitting More Into Every Token

Code examples from the blog post: Context Window Compression: Fitting More Into Every Token

Files

File	Description
extractive_pruner.py	Part 1 — Removes filler phrases, compacts sparse JSON, stubs code bodies, deduplicates messages. No LLM calls.
abstractive_compressor.py	Part 2 — Sliding window compressor using gpt-4o-mini with a structured lossless prompt. Progressive summarization.
dynamic_context_manager.py	Part 3 — Agent-driven scoring of message importance vs. task goal. Selective compression + on-demand re-expansion.

Installation

python -m venv .venv
# Windows
.venv\Scripts\activate
# macOS/Linux
source .venv/bin/activate

pip install -r requirements.txt

Environment Variables

export OPENAI_API_KEY=your_api_key_here

Usage

# Part 1: Extractive pruning (no API key needed)
python extractive_pruner.py

# Part 2: Abstractive compression (requires OPENAI_API_KEY)
python abstractive_compressor.py

# Part 3: Agent-driven dynamic context (requires OPENAI_API_KEY)
python dynamic_context_manager.py

Requirements

• Python 3.11+
• OpenAI API key (for Parts 2 and 3)

---

Full article: onemancrew.dev/blog/context-window-compression-fitting-more-into-every-token