The LLM approach is simpler and more flexible since it works with every library and language out of the box.

Looking up official documentation would require shipping sophisticated parsers for each language, plus a way to map tokens to their corresponding docs. I'd also need to maintain those mappings as libraries evolve and documentation moves. Some ecosystems make this easier (Rust with docs.rs), but others would be much harder (AWS documentation, for example).

I also want explanations visible directly in hover hints rather than linking out to external docs. So even with official documentation, I'd need to extract and present the relevant content anyway.

Beyond that, the LLM approach adapts to context in ways static docs can't. It generates explanations for code within the documentation you're reading, and it works on code that doesn't compile, like snippets with stubs or incomplete examples.

It could be interesting in the future to look into doing some type of hybrid approach where an LLM goes out and searches up the documentation, that way it's a little bit more flexible. But that would also be a bit slower and more costly.