Daniel Diniz <daniel@flashcourier.com.br> writes:
> The index I use is the GIN.
pg_trgm, you mean? That answers one question, but you still didn't
explain what type h.nome_des is, nor how bytea and convert_from()
are getting into the picture.
The second part of that is probably not critical, since the planner
should be willing to reduce the convert_from() call to a constant
for planning purposes, so I'm unclear as to why the estimate for
the ilike clause is so bad. Have you tried increasing the statistics
target for h.nome_des to see if the estimate gets better?
regards, tom lane