OK. My first try was to search for "plato dialogue concerning friendship". Google gave me the result I expected (a reference to the Lysis dialogue) through wiki

OK. My first try was to search for "plato dialogue concerning friendship". Google gave me the result I expected (a reference to the Lysis dialogue) through wikipedia and a bunch of articles about it (the most helpful being a link to the Stanford Encyclopedia of Philosophy, ranked third). It didn't link the text though in the first page (it only appears in the third page of results, with a copy at the MIT Classics archive). Samuru gives me a bunch of general articles on Plato first (oddly enough, the first results are articles from the SEP, but not the article on "Plato on Friendship and Eros".), some noise and then information on the Lysis. The text itself appeared at 25th place. Something I find interesting is that one of the snippets samuru gave me (on the 5th result) has a pretty good description of the lysis as the item most likely to be the "plato dialogue concerning friendship": "the dramatically later Lysis presents Plato's more developed understanding of love and friendship than the dramatically earlier Symposium and Phaedrus". From this description of the Lysis one could gather that the text of the Lysis itself should be a very relevant result to the query; at the very least, that information about it should be weighted as more relevant to the query than info on the Symposium or the Phaedrus, and then info on those over all else. From this, I think, one could build a better representation of a good answer to the query than in google or samuru. I think natural language analysis is very promising here. I hope work on this area yields good results, but it seems like a hard problem.