The best way to learn is by doing, and taking the least amount of baggage with you. So I have approached this by being completely naive and not trying to learn everything all at once. I have tried to learn the bare minimum to get started and then build on that knowledge as I go. This helps me understand much bigger projects and how they work, by helping me understand the problems they are trying to solve. More learning by fire than book learning.
I have two very different side projects that I am working on, where intelligent retreival is required:
- A front end to the SEC Edgar database where you can ask questions. It originally started as a way to investigate SPACs during the 2021 SPAC boom. The current state of the app is at embarc.com.
- An app that pulls all the data about me from all over the internet and my devices and puts it in one app that I can then ask questions of. Like, "How did I meet so and so" and it checks when we connected on LinkedIn, cross references with my calendar and also my location history for that day and the days before.
This project is an attempt to build a framework that can be used to build intelligent retrieval systems. The main requirements are:
- I need the ability to test and iterate quickly on ideas, models, and data using both local and cloud resources. And compare the results of different approaches.
- Be able to put into production and decide to change approaches later without having to rewrite everything. So switching embedding models part way through a project should be easy.
- When there is a lot of processing, particularly locally, it needs to be resumable -- the user may close the app while it is processing and then come back to it later.
- I want to be able to code things using a node editor for no-code and low-code users, and then be able to see the code that was generated and edit it.