Developer builds LLMSlim to fix silent failures in AI prompt compression

A software developer has released LLMSlim, an open-source Python library designed to compress large language model prompts while preventing critical instructions from being silently dropped. Unlike common compression tools that simply trim tokens by similarity score, LLMSlim uses a deterministic six-stage pipeline built on a Directed Acyclic Graph to process and rank every sentence. The library combines TF-IDF vectorisation with the LexRank graph algorithm to score sentences by informational centrality without relying on neural embeddings, keeping latency under 30 milliseconds. A tiered classification system ensures that high-priority content — such as system role markers, JSON schemas, and directives like MUST or NEVER — is protected from removal regardless of its similarity score. The project addresses edge cases that standard prompt compression methods routinely miss, including truncated code blocks, dropped entity names, and broken structured data formats.