v0.2.1 — measured open-weights results

README refresh: the evals section now reports real model measurements — open-weights clef-flash 9B (float16, Kaggle 2x T4) on the committed gold dataset: chunk accuracy 0.712, kept F1 0.758, 38.4% of context tokens removed, scoring latency 1,274 ms p50 self-hosted. Reproducible from the public clef-compactor-evals kernel.

Also ships evals/backends.py (local-weights backend) and --mode local in the eval harness. No API changes.

v0.2.0 — production quality

clef-compactor v0.2.0

Production-quality release of query-aware RAG context compaction with Cloudflare Clef. Already on PyPI (this release re-tags the restored history; publishing skips the existing version).

Added

• OpenAI-compatible wrapper (mount as /v1/chat/completions in FastAPI/Flask)
• LangChain (clef-compactor[langchain]) and LlamaIndex (clef-compactor[llamaindex]) adapters
• Async client, retries with exponential backoff + jitter, structured error hierarchy, env validation
• py.typed, full type hints and docstrings
• evals: reproducible harness, 20-case gold dataset, committed replay results, --min-accuracy CI gate
• Reusable regression-gate action: uses: Gjusev/clef-compactor/.github/actions/clef-evals@main
• Local-weights backend (evals/run_eval.py --mode local) + Kaggle kernel measuring clef-flash on T4 x2
• CI matrix 3.10 / 3.11 / 3.12, Makefile, docs landing, demo video, animated diagram, offline demo notebook

Quality

• 132 tests passing, 96% coverage, ruff clean
• replay eval: chunk accuracy 0.990, kept F1 0.992, 34% context tokens saved

v0.1.0 — PyPI reservation

First release: reserves the clef-compactor name on PyPI. Query-aware RAG context compaction using Cloudflare Clef.