Usage of literal_eval considered unsafe

[kevinkupski]

The chat models currently use literal_eval to parse messages.

The official Python documentation states:

Warning It is possible to crash the Python interpreter due to stack depth limitations in Python’s AST compiler.

Also official CPython contributor vstinner states:

It was never safe.

Since messages clearly contain user input, we currently consider this an issue which prevents us from using this implementation in production. Are there any plans to rework this part of the code and eliminate the literal_eval?

[kevinkupski]

Might also relate to #694

[lkuligin]

yes, adding a flag would help with that.