I think it would be quite a bit more interesting if you could converse with the model. The back and forth "is this paper about foo related to this other paper a

I think it would be quite a bit more interesting if you could converse with the model. The back and forth "is this paper about foo related to this other paper about bar?" would probably be a better way of getting at the interconnections. This should be doable now. The thing that might hold it back is the latency in the experience. You could mask it with the AI equivalent of "ummm ..." to get to maybe 5-10s.