You'll likely always be behind the curve with that kind of basic approach I think, wouldn't it make more sense to retrain a transformer to map all possible inputs to the few possible actions?
Zork was originally written in 1977 on a PDP-10, which had a ceiling of about one megabyte of memory; it was ported to microcomputers in three parts from 1980-1982, and had to fit into about 48k of RAM and 85-170k of disc space.
It is approximately forty-five years old. The curve has long since moved on.
Inform6 it's very powerful. You define the actions objects, and it does it very well by default. Objects have atributes (openable, stackable, decoration...lots of them) and you can even set the library messages on these actions after being done (or before).
Also would love to see what ChatGPT would output if asked to describe a map or give instructions to draw.