WhiteBox — following an answer back to its sources

By WhiteBox, a few things I'd been playing with started turning into the same problem. Memento had retrieval. I'd also gotten more interested in giving messy information some structure instead of treating everything as plain text.

When a model generates an answer, a lot of the research gets squashed into it. If the answer is wrong, or the part you cared about got dropped along the way, you can't get back to it.

For WhiteBox we took a medical textbook and built it into a graph before anyone asked a question. Concepts linked back to the passages they came from. Different names for the same thing could be merged into one. Relationships the textbook stated outright stayed as relationships in the graph, instead of getting lost in a chunk of text.

Then the chat answer wasn't the end. You could go into the graph and keep working through the material yourself.

I don't want provenance to be a tiny citation badge under a finished answer. I want enough of the work left behind that people can check it, go off in another direction, reuse it somewhere else, or decide the generated answer was wrong in the first place.

The team was me, Kaaustaaub Shankar and Raihan Rafeek, the same crew from Memento and later Gyrus.