LLMSlim

Long prompts, retrieved documents, and conversation histories can consume context budgets before the useful information reaches a model.

I built an open-source Python library with extractive, rewrite, and hybrid compression. The local extractive path uses semantic or hybrid chunking, TF-IDF and LexRank ranking, and instruction and entity preservation. Provider-aware strategies support more advanced workflows.

• Published on PyPI with a CLI, document APIs, and an interactive Studio.
• Expanded v0.4.0 with context provenance, provider-aware normalization, canonical JSON, SHA-256 fingerprints, and contract checks.
• Evaluation covers 375 generated schemas and 18 catalogs; schema optimization is a research track, not a claim of a shipped optimizer.
• Selected for the Sarvam Startup Program; Product Hunt daily rank #11, as recorded in my resume.