Post Aay0wtFqnRwNR4GvOi by tomhazledine@mastodon.social
(DIR) More posts by tomhazledine@mastodon.social
(DIR) Post #Aay0wtFqnRwNR4GvOi by tomhazledine@mastodon.social
2023-10-20T16:12:04Z
0 likes, 0 repeats
Starting to feel pretty pleased with the #LLM generated “related posts" section at the bottom of all my blog posts.Similarity is calculated by using embeddings and cosine-similarity, and the description is generated using #openai #GPT4Tweaked the prompt to make the recommendations a little terser and more fact-based, which has made a big difference.(Massively inspired by listening to @simon talk about the power of embeddings - and you know what? I think he was right!)
(DIR) Post #Aay0wuGF3TqcYZSke8 by simon@fedi.simonwillison.net
2023-10-20T16:29:55Z
0 likes, 0 repeats
@tomhazledine ooh, generating text saying why it's similar is a neat twist!
(DIR) Post #Aay1pE0lq6g9Ru7x7w by tomhazledine@mastodon.social
2023-10-20T16:39:42Z
0 likes, 0 repeats
@simon Started trying to do this by hand a few years ago, but quickly gave up because it was way too much effort for even my tiny blog.LLMs FTW! 💪Ended up being a two-step process to get around token limits when dealing with long posts:1. Get the LLM to generate a summary2. Use those summaries to generate the “why these are similar" text