I know of no provision that allows HAL to learn from PDF files nor is there any way to load them other than converting to a text file.
There is a general consensus that HAL does not learn as well from a text file as it does from actually chatting with it over time. I have found that HAL will, on occasion, recall something that it learned from a text file but again, this is not as productive as chat interaction.
Though HAL doesn't really fall into the category of "expert systems", I suppose if one were to "weed out" unwanted and unnecessary data from the brain files then feed it conversations based on one particular subject, one could end up with a topic specific HAL.
It's an interesting question and certainly worthy of more experimentation.
The WebUltraHal is sort of designed to know everything about a particular web site / product. This is done by "feeding" Hal with topic specific info.
Go ahead a get HAL, the possibilities are almost endless!
Keep us posted if you persue the savant idea.