I like the idea of suggestions a lot. As we get more and more nodes, we will basically need to make a search engine kind of thing in the future anyway. So these suggestions seems a like a good step towards that.

The only thing I don't like is that suggestions based on the specific inputs/outputs of specific nodes. This just doesn't scale. Imagine if we went all in with adding suggestions. We'd have to maintain multiple suggestions for each node.

I'd make suggestions based on data type + direction (input vs output). E.g. Load Image is a reasonable suggestion for pretty much all nodes that take an image, and Upscale Image is a good suggestion for pretty much all nodes that output a PyTorch model.

Edit: That said, we register suggested nodes for a data type based on inputs/outputs. E.g. the ImageOutput of Load Image could have a .suggest() function make this input as something to be suggested.

Future talk: To further improve suggestions, we need user data. Search engines give good results, because they use metrics (e.g. how many sites you visited from a search, how long you stayed, whether you changed the search query) to score their search results, and then use those score to adjust the rank of results.

With enough data, we could provide suggestions not just based on data type, or node, or node output/input, but even based on node + input data + connected nodes. That's the power of data.

That said, I'm not suggesting (hehe) that we should spy on our users. We basically just need a large collection of chains to use as data to get rankings. So we should probably encourage users to share more chains and make it easier to import shared chains. I imagine users would probably share a lot more chains/sub-chains if there was an easy way to copy'n'paste them directly from discord or elsewhere. We could then use that as a data source.