LLMSlim makes long prompts, documents, retrieval, and chat history easier to reason about—with local extractive compression, provider-aware strategies, and provenance-aware priorities.
24 curated samples stay visible in the methodology; the release story also includes categories, languages, schema contracts, security regression, and machine-readable artifacts.
Explore the evidenceLLMSlim gives context a smaller, more intentional shape—without hiding the role, strategy, or boundary that produced it.
Illustrative motion only. The released library exposes results—not hidden model reasoning.
Local sentence selection. Deterministic, inspectable, and available without a provider.
RAG text cannot gain protected priority merely by sounding imperative.
LLMSlim mitigates compression-induced priority elevation. It is not a complete prompt-injection defense.
Language coverage is reported as a small, disclosed evaluation slice—not a universal multilingual quality claim.
Phase 2 measured 18 synthetic tool catalogs across simple, medium, and complex schema classes. At the measured extremes, one simple tool contained 62 tokens; 64 complex tools contained 12,739.
This is not a shipped schema optimizer. The real work now is finding out when contracts can be safely minimized or selectively loaded.
Open the interactive explorerDerived 32-turn totals are labelled separately in the explorer.
Measured results, derived session totals, and unavailable provider-dependent findings are never flattened into one number.
Benchmark methodologyUntrusted RAG, tool, and assistant text cannot gain protected priority just because it sounds authoritative.
Read the ContextRole guideExtractive compression works without a provider. Rewrite and hybrid flows require one you supply and should be measured in your environment.
Install LLMSlimRun a local extractive path first. Add a caller-supplied provider only when a rewrite or hybrid workflow makes the trade-off explicit.
