From the link that I edited into your question (realized that the answer was there):
VAF is the percentage of sequence reads observed matching a specific DNA variant divided by the overall coverage at that locus. Because NGS provides a near random sample, VAF is thus a surrogate measure of the proportion of DNA molecules in the original specimen carrying the variant. For constitutional genetic testing, VAF is a measure of diploid zygosity: heterozygous loci should be near 50% VAF, homozygous loci should be near 100%, and reference loci should be near 0%. Deviations from these three expected values should be considered suspicious as potential errors due to incorrect base calls or alignment. For somatic testing, contamination from normal cells and tumor heterogeneity combine to cause unpredictable VAFs. In some clinical scenarios, such as testing a patient for therapeutic resistance mutations, the desired sensitivity dictates that variants with low VAF be included in downstream analysis. This all means that in order to maintain an acceptable level of sensitivity, variant filtering for somatic variant must be highly permissive from a technical validity perspective.
This is assuming diploid humans. For other ploidies the expected proportions of heterozygosity will match the possible heterozygous genotypes (e.g. 33%/67% for triploid heterozygotes, 25%/75% for tetraploid, etc.).