Over the weekend, I tested LLMs and Embeddings to solve the hallucination problem.
As an extension to last week's experiment. The combination of both is mind blowing!
I have been trying to create an echo chamber for twitter, in a way that is not just curates your "For you" tab with all your interactions, but helps you tune it to your needs.
I used LLM as a feature extractors (extracting topics, sentiment, etc.,)
Embedded all the topics, sentiment, tweets, etc., and ran through a complex logic that will create a weighted average score for each of the tweets from my feed.
Voilà! My feed is exactly as what I need. I can find the top 20 tweets that I can engage with, without the need for reading 200 tweets.
It's currently costing me $0.1/100 tweets crawled, but there's a lot of room to reduce this cost.