THE NUMBERS

How fast is Findra?

Everything below came out of findra --searchbench and was pasted without editing. No number appears on this page that you cannot reproduce on your own computer with that one command.

Type part of a filename. The matches are back.

Median round trip for every query measured, across 1,780,595 names, from the moment the query leaves the window to the moment the results arrive.

config0.60 ms

report0.70 ms

readme1.03 ms

invoice3.24 ms

sunset3.78 ms

5.4 s From cold to ready 1,780,595 names read off the disk when the helper starts. Longer right after a restart, when nothing is cached.

200.3 MB What it costs to hold The name index for 1,780,595 names, measured by the helper that holds it.

0 Bytes about your files No account, no cloud, no analytics, no crash reporting, no telemetry.

FILENAMES

Name search, query by query

The round trip is the whole journey: from the window, through the pipe to the helper that holds the names, and back. The index scan is the helper's own share of it.

Name query	Round trip p50	Round trip p95	Index scan p50	Worst	Hits
config	0.60 ms	0.78 ms	0.15 ms	1.13 ms	50
report	0.70 ms	0.89 ms	0.26 ms	0.96 ms	50
readme	1.03 ms	1.25 ms	0.58 ms	1.55 ms	50
invoice	3.24 ms	4.57 ms	2.81 ms	4.88 ms	43
sunset	3.78 ms	4.62 ms	3.41 ms	5.36 ms	21

INSIDE FILES

Full-text search

Against this machine's own content index, read from a real disk, so the slowest word is in there beside the fastest.

Word	p50	p95	Worst	Hits
lease	1.90 ms	2.59 ms	4.07 ms	50
agreement	3.68 ms	4.43 ms	4.70 ms	50
invoice	0.79 ms	1.08 ms	1.80 ms	32
total	9.41 ms	10.89 ms	13.53 ms	50
report	15.36 ms	16.67 ms	16.87 ms	50

Machine AMD Ryzen 9 9900X3D RAM 47.1 GB Disk NVMe SSD OS Windows 11 Pro 10.0.26200.9445 Accelerator ONNX via DirectML Build Findra 0.3.1 Samples n=50 per query

A number without its machine is marketing rather than measurement, so the machine is named. Yours will differ. The whole run, every table and the stores it measured, is in the README, pasted as the command printed it.

Findra tries DirectML for the vision and meaning models and Vulkan for speech, and falls back to the processor when neither answers, which is a supported configuration rather than a failure: only the first pass through your files is slower, and findra --searchmodels prints which provider it chose and every one it turned down, with reasons.

These numbers are from one machine, and it has an NVIDIA card. Findra is built for AMD and Intel processors, for AMD, Intel and NVIDIA graphics, for integrated graphics, and for machines with no usable accelerator at all - but only this configuration and the processor-only path have been measured. AMD and Intel graphics have not been tested on real hardware, and neither has an arm64 machine - the arm64 build ships because keeping it reachable costs nothing, not because it has been run. The paths Findra uses are vendor-neutral by design so that they should work there; that is a decision rather than a measurement, and it is said here as one.