Hey Product Hunt! I'm Yash, a CS/AI undergrad currently doing an AI research internship, and LLMSlim is something I built because I kept hitting the same problem in my own RAG pipelines: most of what I was sending to the API was filler. LLMSlim runs a deterministic, offline pipeline, no LLM calls, no telemetry, that scores every sentence on centrality, entity density, and instruction presence, then hard-locks anything matching an imperative ("must," "never"), a role marker, or a code fence before it's ever eligible to be cut. Everything else gets ranked and trimmed to hit your target ratio. It's early v0.2.0, MIT-licensed, 93%+ test coverage, and the benchmarks on the site are the actual numbers from the CI-gated benchmark suite in the repo, not rounded up. Latency ranges from ~24ms on short prompts to ~340ms on 12K-token documents, I'd rather you see the real range than a single best-case number. What's next: streaming compression (compress_stream), native ONNX embeddings for sub-5ms scoring, and async batch helpers, all tracked on the roadmap in the repo. I'd genuinely like feedback on where the compression gets it wrong. If you paste a prompt into the playground and it drops something it shouldn't have, that's exactly the kind of report that makes v0.3.0 better. Thanks for taking a look.

LLMSlim

Semantic prompt compression that never drops instructions.

Semantic prompt compression that never drops instructions.