Al Linkedin Loop lead searcher - JSON structure ¿?

Hi everyone!!

So I've been working lately on a personal Al system for LinkedIn's lead search and I kinda got stuck with the memory and registry system...

Basically, what this system does is:

Uses context from system prompt
Uses Brave for a first search and gives a first list of contacts to analyze.
Tavily makes a deeper search on the list that Brave first extracted and filters.
A FileSystem registers the prior searches for future triggers and uploads the memory.
Turns the lists from Tavily into a mini-report and sends it to inbox through Github

I've added an excalidraw diagram for it to be better understood in any case :)

So the problem:

The thing is that I still don't know how to aim the JSON formatting or how to approach an archive for it to register the prior searches and contacts that are no longer interesting, and how to make this scalable and persistent. I thought of creating an excluded.json which compiles both accepted and rejected contacts and a reported.json which registers the ones that were taken into consideration with a brief explaining for the system prompt to be richer.

Is that split worth it, or would you just keep one record for everything and filter?
Who writes the file, the agent or deterministic code?
Anything you'd have designed differently from day one?

Thanks a lot in advance to everyone and I'd love to hear thoughts on this 😃

4 comments

Al Linkedin Loop lead searcher - JSON structure ¿?