Embeddings and all the Black Magic
The first AI concept that truly fascinates me is embeddings.
There is so much literature about it that I won't try to explain it again. For that, open up Medium and search for an introduction to this subject. I find a good starting point in this article.
Here comes the black magic:
Embeddings catch the meaning of a text-based source. And that meaning is surprisingly language-unaware.
You can embed an English document
and search it with an Italian query.
In my first experiment, I indexed a markdown-based documentation project that I wrote in English. Then, I managed to find useful information out of a query that I wrote in my local Italian dialect.
I was stunned.
Embeddings For Dummies
Embeddings are not something you do; you would instead buy them in a marketplace such as OpenAI, which is possibly the most famous. There are also many open-source alternatives, but hosting and running them requires a heck of a piece of hardware.
I am an eager consumer of ChatGPT and so I've decided to start with OpenAI.
Setup your API Key
Initially, I got confused because I couldn't find any "API Tokens" section in the ChatGPT interface, and I was puzzled about how to get to use it programmatically 🧐.
It took some googling to figure out that ChatGPT and OpenAI are two different things. I hope you appreciate me sharing this with you, even if it casts silly spotlights on my weekend-induced naivety.
So here is the receipt:
- Go to OpenAI and Login
- Navigate to Settings / Billing
- Add some funds: you need a minimum deposit of 5€ to be able to use the API!
- Move to "API Keys" and generate your key
👉 I tried to generate the key before making the money transfer, but I got "insufficient funds" even though I could see some trial credits. This is not something you quite expect.
Once you have your key, create a .env file in your JupiterLAB project (checkout the repo here) and store the key:
# .env
OPENAI_API_KEY=xxx
And restart your project:
make restart
Generate Embeddings
As we said, generating embeddings is just a matter of turning text into an array of numbers.
Checkout the step-by-step source code here: