Reading the numbers

The figures on this project are ratios against other search tools, and a ratio is easy to quote and easy to misread. This is what each word in them means, so you can decide what the certificate is worth before you open it.

Warm and cold are different machines

A cold query is the whole job: walk the tree, read every file the filters admit, match. That is what ripgrep does every time you run it, and it is the honest baseline for a tool you have just installed.

A warm query is what gist is for. A persisted trigram index has already seen the corpus, so the query is planned against the index first and only the files that survive that filter are read. When a resident session is also up, the index is already in memory and the query never touches disk to plan.

Both numbers are real and they answer different questions. Warm is the steady state of a tool you leave installed in a repository you work in every day. Cold is the first run, and the state you fall back to when the index is gone. The figures on this page are warm, and they say so.

Why geomean rather than an average

Each benchmark class produces its own ratio, and the summary has to combine them. An arithmetic mean over ratios is the wrong instrument: it is dominated by whichever class happened to produce the largest number, so one pathological win can carry a summary that describes nothing.

The geometric mean multiplies the ratios and takes the nth root. It treats 2x and 0.5x as equal and opposite, which is the property you want when the thing being averaged is "how many times faster" - a summary that cannot be inflated by adding one more class the tool happens to be very good at.

What a ratio does not tell you

Speed is one axis and the certificate measures others that do not reduce to a multiplier. Peak memory while indexing is the clearest of them: it is reported as an absolute, it is a lane gist loses, and no ratio on this page hides it. The caveats beside the figures name that and the rest.

A ratio also says nothing about correctness. Being fast and being right are separate claims that need separate evidence, which is why conformance is measured against ripgrep's own documented flag surface and its own test corpus rather than folded into the performance number.