Hands-down my most productive interface to LLMs for [years since GPT3.5] years running.
Do you have any examples of what decent RAG functionality might look like? And where the current plugins fall short?
Why is a plugin necessary?
Sometimes you'll have to push a big ugly commit.
But other times the manual diff review can save you from a headache, like if you have some obscure syncing going on, like syncing READMEs and other markdown files to external repos to manage all markdown with the same Obsidian interface.
Also if you need to maintain a high-standard for the contents of your notes while still utilizing AI tools, the manual diff review can prove invaluable in ensuring trusted resources don't turn into slop
One caveat is that the obsidian android app DOES NOT seem to save files to storage until the note unloads, which can break things if you pull in the middle of making changes.
[1] Though I have had to fix my termux clone of the vault enough times that I now just nuke it and re-clone instead of bothering with git - but that's more of a "termux likes to break git" issue than anything
But as an Obsidian power-user, I regularly paste screenshots into notes
There is a plugin that allows templating the screenshot file name, so naming the pasted screenshot, using the same as the note where it's being pasted, and a timestamp, for example, is easy.
I'd hope that any inference performed in a given note would take as context not only the full text of the note, but its nearest neighbors—although, frankly, what I really want are the implicit definitions of proper nouns, ie links.
I also frequently generate articles based on the subject's backlinks, ie mentions in other notes.
Both of those functionalities are provided by Obsidian's core featureset, so all that's missing—and what might provide outsized value—is indexing notes based not only on the embeddings of their literal text, but on meta-text such as: the questions they might answer; the definitions of their principle subjects; counterfactual statements they falsify.
All of which would allow the editor to bring up a relevant note when writing/reviewing another–ie the "unknown knowns".
I think this technique is akin to "knowledge enrichment", or "meta-embeddings," but I think actually those terms have lower-level meanings.
Of course, generating these embeddings would make the indexing process slower and more failure prone, and so I'd be even more likely to fail to get it working and call it a dud.
That would be perfect!
Smart Connections does seem to have improved a lot recently.