Ref T9979. This uses ngrams (specifically, trigrams) to build a reasonably efficient index for substring matching. Specifically, for a package like "Example", with ID 123, we store rows like this:
< ex, 123> <exa, 123> <xam, 123> <amp, 123> <mpl, 123> <ple, 123> <le , 123>
When the user searches for exam, we join this table for packages with tokens exa and xam. MySQL can do this a lot more efficiently than it can process a LIKE "%exam%" query against a huge table.
When the user searches for a one-letter or two-letter string, we only search the beginnings of words. This is probably what they want, the only thing we can do quickly, and a reasonable/expected behavior for typeaheads.