Then I decided to try and switch to the local mode, and after ~ an hour figuring out how to build a markdown version of the docs I needed, hit the "object has no attribute 'generate_from_text'" error, as someone else also reported [1].

So I cloned the source and started to look around, and the method really doesn't exist, even though it's called from main.py. A comment above it says "Assuming LLMMinGenerator has a method to process raw text" and I immediately feel the waft of vibe coding... this is all a mirage. I saw a long README and assumed it was real, but that was probably written by an LLM as well. Would have been obvious by the 'IntegratedKnowledgeManifest_SKF' and 'GenerationTimestamp' keys in the 'SKF format' definition - the former makes no sense, and neither has any reason to be this verbose when the goal is compression.