Show HN: Cordon – Reduce large log files to anomalous sections

Dev tool cordons off log chaos; ops cheer, skeptics cry snake oil

TLDR: Cordon uses AI to spotlight unusual moments in huge app logs, cutting the noise so humans see what matters. Commenters split between loving the time-saving filter and warning it could hide repetitive yet critical errors, with extra spice over GPU support and whether simple tools already do the job.

Show HN: Cordon dropped a tool that promises to shrink those endless app “diaries” (log files) down to the weird, important bits—and the comments instantly turned into a split-screen. Ops folks cheered: “Finally, meaning over keywords!” They love that Cordon treats repetitive messages as background noise and surfaces rare, truly odd events. The GitHub repo and a plain CLI had devs installing faster than you can type pip.

Then came the skeptics. They argue it’s just embeddings (AI that understands text) plus k-nearest neighbors (a simple way to score weirdness) wrapped in a shiny package. One warned: if you hide “repetitive errors,” aren’t you burying the smoke that points to the fire? Another jab: “grep plus alerts is still king.”

Hardware drama? Absolutely. GPU optional made people ask why their old GTX 960 is benched, while RTXs strut. Jokes flew: “At last, my GPU has a job besides Minecraft.” Others poked fun at the option to exclude the top 5% strangest events—“You’re ignoring the monsters? Bold.”

Still, practical voices said: anything that cuts log noise during a 3 a.m. outage is a win. The vibe: promising tool, spicy trade-offs. And yes, someone yelled “snake oil”—because it’s Hacker News. Of course.

Key Points

  • Cordon detects semantic anomalies in large log files using transformer embeddings and k-NN density scoring.
  • Repetitive log patterns are treated as normal background; rare or clustered events are surfaced as anomalies.
  • Supports sentence-transformers (default) and llama.cpp backends, with optional NVIDIA GPU acceleration via CUDA.
  • Installation available via PyPI or source; provides CLI and Python API with configurable analysis parameters.
  • Models are downloaded on first run and cached; example usage covers batch sizing, anomaly percentiles/ranges, and output options.

Hottest takes

"Finally, something that reads logs for meaning, not keywords" — HN commenter
"k-NN on embeddings isn’t magic; grep and alerts still win" — HN commenter
"Filtering ‘repetitive errors’ is exactly how breaches hide" — HN commenter
Made with <3 by @siedrix and @shesho from CDMX. Powered by Forge&Hive.