Use a LLM to infer the segments of a scene

[jbilcke-hf]

Context

Currently we use the Broadway library to convert a screenplay to a .clap file

But this uses heuristics (if/elses, pattern matching using regexes) which can scale (importing only takes a few seconds) but are "dumb" and can't get the bigger picture

Solution

Call the LLM when we import a story, for more accurate conversion to the timeline.

Note that this is a non-trivial ticket, there are technical constraints (context window, speed, memory, JSON validation etc)