@Curtis Hays the point of jev is not to tradeoff speed for output quality, but to answer queries lightning fast by reading and hopping through the markdown files until we have sufficient info to complete an agentic task

after this post was made, we now have tools like jevgrep, which searches within files and return the right context

i have a way of making my local pc run scheduled jobs, replacing need for vps and ur local pc can be sleeping until it wakes up to run the job. but first, that local pc needs to be macbook M series, it has this command called pmset which lets you wake up the mac at the time you set daily. e.g. pmset wakeup daily 2am once the mac is awake, your scripts running on cron or some scheduling tool, calls the program, and when successfully done its job, calls pmset sleepnow to continue sleeping what if the job fails due to no internet? well the failed cli will schedule another wake up in the next hour then goes to sleep, to retry the job later, continuously retrying until it succeeds or reaching max num of retries that you set. you see where this goes, make the mac wake up to do its job, if it fails, then try again by repeating wakeups periodically. This method lets me get best of both worlds, my data stays on my machine without separate vps to manage and pay for, and it runs jobs periodically like a vps.

I'm using oh-my-pi, pi agent with more builtin features and has nice tui

@David Vogel after playing around with it, yes you are right the system prompt is more bloated than native pi's but i was able to change the system prompt to pi's and it works fast now. the read tool and write tool feels better with hashline editing, and it has built in web search. i was able to pass the --tools cli flag to limit num of tools to further reduce system prompt bloat. loving the power and flexibilityso far with oh my pi

I would like to use hermes but I read that it doesnt work as a coding agent? Ideally I want to have a coding agent which is backed by second brain, so Im using pi agent at the moment

It's the same question that I have, when pulling entries, there might be irrelevant entries being read in which pollutes the context. We should have the LLM score the entries before pulling them in as context, then score them later after reading to score how they useful they really are to answer the question. This way we can always fine tune our entries to make it be more relevant for future queries. But how to implement this? Perhaps writing these scores to a log?