I've used simple TF/IDF with a custom stemmer as a similarity scoring tool against the XDK documentation. It worked pretty well, I'd be curious to know how that

I've used simple TF/IDF with a custom stemmer as a similarity scoring tool against the XDK documentation. It worked pretty well, I'd be curious to know how that'd fare vs. the neural model (i.e. I suspect it'd be better in identifying the rare, high-signal words that were excluded by the author's 400-most-common word limit). For my side project: As we received emails from developers asking for clarification/help with the APIs, the system would provide relevant documentation URLs so that anyone could pick up an inquiry and brush up on the API (and have a handy link if the the docs could be leveraged in the response).