Do I need to specify that the log is from NGINX before pasting it?
No, detection is automatic: the analyzer tests up to the first 50 non-empty lines against the combined pattern and assumes "nginx-access" when 30% or more of them match. A log mixed with other sources can fall back to the generic parser if the ratio does not hold.
How does the tool decide what counts as an error?
Straight from the HTTP status code: 500 and above becomes error, 400 to 499 becomes warn, any other code becomes info. There is no keyword inference for this format, because the status itself is already a reliable severity signal.
Does it process the whole file or just a sample?
The 50-line sample is only used to decide the format; the analysis itself processes the whole file, up to 12 MB on upload, but the on-screen table renders at most 1,500 lines at a time to stay responsive.