Chonkify v1.0 beats LLMLingua2 by 175%

This looks like a genuinely interesting niche release if the numbers hold, but the claims are still based on a small internal suite with proxy recovery metrics, so independent replication matters.

• –The selection core scores passages by information density and diversity, then keeps the highest-value subset under a token budget.
• –The repo’s benchmark summary spans 5 documents and two budgets, with a mean +68.57% gain over LLMLingua and +174.90% over LLMLingua2 on composite recovery.
• –The benchmark caveat matters: the scorer is proxy-based, so these results are best treated as directional evidence, not ground truth.
• –Packaged wheels, a CLI, and a Python API make it more deployable than many research-heavy compressors, especially for RAG pipelines.
• –Support for Azure OpenAI, OpenAI-compatible endpoints, and local SentenceTransformers gives teams a practical cloud-or-offline path.