Yesterday was a strong start, but I suspect keeping up the momentum will be the real challenge.
Accomplished the following:
- learned the basics for LlamaIndex (formerly GPTIndex) and Langchain
- found and recorded some examples and resources for deploying chatbots built on said structures
- isolated some evidence-related resources to feed into a prototype, which even (*gasp*) required me to spend some upfront money!
- created a few working chatbots over said material, one of which ultimately seems promising
- picked a kitchy product name without overthinking it (which, for me, is huge) - “lawbrarian” –> short, .com was 13 bucks, easy to understand and spell, tells you what the product is
I pivoted immediately to prototyping with evidence-related material rather than impaireds at the suggestion of my wife, who tends to be right about most things. And I’m trying to take a different, more practical and targeted approach this time around - she is a user, after all, as civil litigation counsel, and evidence knowledge is a pervasive need across all litigators, meaning it’s perhaps a smarter place to start than something ultraspecialized like impaired material.
Incidentally - while llamaIndex was a bit quicker to get up and running, seems like going full langchain may be the move. A lot of what you’ll need to stand up a full product will require integrating llamaindex with langchain anyways, and I found the full-langchain solution to work a bit better and easier in any event. Plus the project seems to have more attention with better docs. Llamaindex appears, to my untrained eye, to be moving more slowly and focusing very narrowly on only one part of the overall architecture, not to mention a couple of annoying compatibility bumps with the namechange.
Keeping up with the theme of moving fast and developing in accordance with user feedback, I’m not going to spend time now optimizing/experimenting with what I prototyped yesterday, but instead I’m going to quickly find a way to stand up some form of working deployment of what I have and stick it in front of users. I may add some more material if the immediate feedback is “but I don’t have any evidence questions right now”, but we’ll cross that bridge if/when.
Today: we’re going to pick the simplest stack for getting this on the web so I can work on putting it in front of users to try. Even though I’m also working in court today.
Aaaaand…. go!