Last few days have been a bit of a whirlwind.
After creating and validating the proof of concept, it was time to sit down and think about an MVP.
Assuming an MVP is a simple version of the product that (a) works, and (b) you’re allowed to charge for, this involves creating my own content database instead of stealing it (obviously).
Running with the theme of building things as simply as possible, this in my current opinion means using my professional experience to whip up a Notion database (really, a collection of pages) of useful material, and then using Langchain’s Notion QA recipe to deploy something simple.
This will take a minute, but is something I’m capable of. It also has the added benefit that I am not stuck with the quality or organization of the published materials, but can use them all to distill and create a practitioner-friendly database that I know will work well in a semantic search/OpenAI completion pipeline.
(That said, with the rapidly increasing context windows coming like GPT-4’s 32K option, the semantic search aspect may become less and less important)
In the meantime, part of the big 90 day kick-in-the-ass has also led me to explore monetizing my existing application for a professional group that could really use it. We’ll see if it works out, but it would be useful for them and nice for me, so here’s hoping. If it doesn’t work out, then I’ll probably kill my original project so I can focus on something working. Also, the same group would be interested in what I’m building with Lawbrarian, so it may all dovetail nicely.
When you start doing productive things, other productive things tend to happen.
I know this already, but it’s always nice to be reminded and experience it firsthand.
Something else I’ve experienced over the last few days is the pains of trying to deploy something memory/compute intensive on Heroku. It may well be time to find another solution, and perhaps even something run primarily with simple client-side vanilla JS or browserified node that uses baked-in data, or something like Supabase (which I have long thought looks very cool and easy, and I would be happy to pay them frankly). Bit of a shame because I like Python, and it will remain what I reach for when playing/experimenting, but the world of deploying modern apps really does seem geared towards JS first nowadays. I really, really like the whole jamstack thing and throwing things up on Netlify or Vercel that just work. And with back-end solutions like Supabase (or even Heroku’s battle-tested managed postgres), or for my newer purposes things like Weaviate or Pinecone, it’s harder to justify paying for the good PAAS services for python, or taking the time to learn and manage my own servers. Maybe PyScript will catch up, who knows.
Now I’m off on a 4 day trip where the very same wife who gave me 90 days wants me to ’not think about this stuff’. Classic. Like that’s gonna happen.