Variant calling compares aligned reads to the reference to identify differences:
SNVs (single-nucleotide variants)
small insertions/deletions (indels)
structural variants
Each call has a genotype (heterozygous/homozygous) and quality metrics (read depth, allele balance).
A genome has millions of variants, but only a small fraction are clinically relevant — which is exactly why retrieval and filtering matter: you don't embed millions of raw variants, you retrieve over the annotated, prioritized subset.