Problem
The tool output is in JSON and maybe to verbose, which translates to a waste of tokens ingested
Proposed solution
Use CSV or a markdown table, and perhaps figure out a way to return less text in the initial search of insights, only expose full text when pulling a specific item.
Alternatives considered
I learned about Mnemon from this gist by Singapore's foreign minister
One commenter suggested a way to make mnemon return more relevant results:
I have asked an LLM to make this suggestion into a design for a wrapper around mnemon to implement this, and this is what it suggested
I have not tried this, I hope it's legit but I don't understand enough on the subject to judge.
more info
TOON as an alternative to JSON for LLMs
Problem
The tool output is in JSON and maybe to verbose, which translates to a waste of tokens ingested
Proposed solution
Use CSV or a markdown table, and perhaps figure out a way to return less text in the initial search of insights, only expose full text when pulling a specific item.
Alternatives considered
I learned about Mnemon from this gist by Singapore's foreign minister
One commenter suggested a way to make mnemon return more relevant results:
I have asked an LLM to make this suggestion into a design for a wrapper around mnemon to implement this, and this is what it suggested
I have not tried this, I hope it's legit but I don't understand enough on the subject to judge.
more info
TOON as an alternative to JSON for LLMs