Most log-analysis tools answer one question: "what broke?" They print a sentence. You have no way to check whether that sentence is right, short of re-reading the raw log yourself.

log-ai-compressor does the same job, but every finding it prints is attached to a line range in your original file plus a sha256 of that range. You can go back to those lines, and you can run the built-in verify to check that a finding was not altered after it was produced. It runs entirely locally: no log upload, no API calls, no LLM keys.

Two commands, about 30 seconds, no dataset download

pip install -U --index-url https://pypi.org/simple "log-ai-compressor==2.16.1" log-ai-compressor run examples/app.log -f summary 

examples/app.log is 24 lines and ships with the package. On my machine the summary is about 234 tokens (a 38% reduction against the raw file), with 8 error lines grouped into 4 signatures, and a fingerprint 48e5781a...c2886.

What I can claim, and what I cannot

Full run over all ten Loghub-2.0 datasets, 1,940,760 lines, 1,621 template clusters:

  • PA 85.05%, GA 88.68%
  • Same data, Drain3: PA 58.06%, GA 71.58%
  • CPU only, zero API spend

Three limits I will not paper over:

  1. Reproducible is not the same as state of the art. There is no third-party leaderboard behind these numbers.
  2. The LLM comparison arm is a 2,000-line sample per dataset, not the full set.
  3. The HDFS block-level causal chain cannot be re-run here: the original 1.5 GB log is not on my disk, so that 37.17% recall number is documented, not re-executable.

Why I am posting this

The repo currently has 0 stars, 0 forks and 1 issue. The only external verification so far is another AI agent that followed the README on a different machine and matched the fingerprint. That is one run, not a public benchmark.

So the ask is small: run the two commands on your machine and reply either way. Failures are more useful than passes. Three things are enough:

  • OS and Python version
  • the exit code of the second command
  • if it failed, the last 20 lines of output

Links

Note: the repo is hosted on Gitee, which is usable but not fast from outside China. PyPI is the reliable entry point for installing it.