Dynamic Content Insertion (Python)

[ojhurst]

So I have just been experimenting with a certain mental model and I kind of wanted to run it by you guys and see if it matches what you guys are doing or if it's different. But basically, you know how like iOS has a text expander. You type in a little phrase MBP and it expands it into MacBook Pro, right? It's purely deterministic, you know, potentially Python-esque, right? There's no LLM. And anyways, I also was starting to think of like this, kind of like this text expander or text shortcut thing. And I was also kind of making a parallel of that idea to the idea of essentially a meaning, a meaning expander. Right. So it's very similar, but also revolved around entities. So first, I'm trying to identify the entity, which is the MacBook Pro. and then I trying to actually find all of the aliases which are maybe my laptop or um mbp um that kind of thing and so there a bunch of different aliases that then that then point to um my my entity my proper entity the macbook pro and then from there there is this um there is essentially the definition which is like hey this is james's macbook pro the ip address is 192.168 blah blah blah and he and he uh he uses tail scale with it okay and so the idea being that um If I ever type MBP or my laptop, I get a direct substitution, essentially. I basically get a dynamic content insertion. I basically get that translation injected into my next prompt to basically help it understand what I'm talking about.

When we're ingesting a session, there's a whole heuristic or methodology for trying to identify what these proper nouns are, what those entities are, and then it can take its best stab at defining them.

And finally, at least in my case, here I am. I just installed Signet and supposedly it's supposed to make things better. What I found myself doing was almost as an auditing mechanism. I'm saying, hey, if Signet is working for one, and then when it does have a match, I'm actually asking it to surface that to me in my prompt. So I don't know if we could make an argument for that being a configuration. But yeah, it's almost like auditability. If I, hey, I just installed this new thing and I kind of just want a little bit of transparency. Is this working? How is it working? And to basically for onboarding to say, wow, that is working. That's pretty darn cool.

[NicholaiVogel]

@ojhurst So funnily enough this maps really close to how Signet already works!

Entities get extracted from session transcripts and memory artifacts, dreaming deduplicates the aliases (so 'mbp' and 'my laptop' would eventually resolve to the same entity), and the loop completes during recall when entities are identified in queries.

I've got an autolink-entities script for my Obsidian vault that does exactly the wikilink expansion you're describing (scan vault for entity names and replace with live links), and Signet has all the plumbing for that same behavior, it just never surfaces it to the user.

I'm curious did you have any ideas of what you'd want that user visible layer to look like? I'm interested in exploring this further.

[ojhurst]

I have it inject its findings inline in claude code personally ...

[NicholaiVogel]

@ojhurst lets see how this goes, trying something in #795