#ai (2023-06)

Can support LocalAI and ChatGPT

Would be great to hear what you learned from the experience

I dabbled only briefly with langchain, enough to ingest hundreds of sales calls exported from Gong, and ask questions of them. Unfortunately, my quick wins ended there and I realized I’d have to go way deeper to tune things, at which point it was over my head.

I learned the water is pretty deep. An important takeaway is that understanding how the thing came to its conclusion is a very important differentiating feature.

I’ve been drowning in papers and models of what essentially amounts to cognition, i.e., how do you train your LLMs to talk to itself in a productive manner?

And an interesting takeaway is that rather than having a singular sequence of thoughts, setting up multiple agents who compete and evaluate each other’s performance results in higher fidelity output.

the latency of the product is really high, that’s why i need to next build a stream of consciousness for the user to enjoy.

I’m twiddling my thumbs about creating a good admin / user interface for managing an internal corpus. I’m hoping you or @Matt Calhoun has made some progress on this, so I can borrow from your efforts.

Honestly, my work in this area has been very rough and “pet project” focused, so everything I’ve done has lacked the polish you are referring to here (like hard coded paths, and manually loading docs into certain dirs, etc).