If the structure is highly symmetric you can use a very low value, 1e-4 or so. In that case, you will also find that changing the tolerance doesn't change the identification.
If the system only has some approximate symmetry, you need to check a few different orders of magnitude and see if you get different results. That will tell you how close to a particular space group you are, and you need to use some judgement to say what is the most reasonable identification, within some uncertainty.
If you end up needing 0.1 or that range to get anything other than P1, probably the system has no real symmetries at all.