Searching Diagrams: How Can Google's PaLM 2 AI Help?

This article explores some of my first uses of Google’s new PaLM 2 models for text completion and embeddings. Unlike OpenAI, these models are less tested but certainly less costly. And by that, I mean “free”. It’s refreshing to experiment with AI models where you don’t have to be concerned with billing. You do have to be concerned with Google Apps Script quotas as you would building automation processes. But they’re pretty hard to breach.

LucidChart Diagram Search (there isn’t one)

I pretty much had a meltdown the other day - I needed to find a diagram in LucidChart that I created about a year ago. I knew it was in there, but I failed to tag it, name it, or otherwise mark it in a way that it could be easily found. This is the case with almost all of my diagrams. They are one-offs that are created quickly to help another developer understand my goal or perspective concerning high-level architecture.

I should be more proactive about organizing these resources, but I’m not. Nor will I likely ever take the necessary time to do that extra work.

Two hours later, I had the diagram I needed. It was then I decided to figure out how to use AI or some other process to help me avoid this debilitating productivity thrashing in the future.

The Core Requirement

I needed to go from diagrams like this…

To this…

This diagram describes a process for generating a PDF document from data in an Airtable base. The process begins with the data being retrieved from the base and then being used to generate an HTML document. The HTML document is then used as a template for a PDF document, which is then generated and saved in the base.

Suppose I can successfully (and consistently) create a text summary that generally reflects a diagram's flow. In that case, I can probably create a better findability approach that will allow me to locate anything in seconds instead of minutes or even hours.