@adrienne
Started working with Claude to come up with a plan to redesign the embeddings architecture for the podcast transcripts that power our AI bot @warpee.eth and episode search on our website.
It occurred to me that if I were to design it today there are probably more established best practices and patterns to leverage
Claude made the following high level recommendations:
- update the model to newest embedding model
- create 2 indexes, one for episode summaries and one for full transcripts
- increase chunk size
- add chunk overlap
- remove filler words ( the ums and ahs that add noise)
I am having Claude build all this following a blue green deployment approach so I can test the new embeddings without impacting the existing ones, and can cutover only when I confirm it’s materially improved
Better embeddings mean more accurate search which means better quality answers.
Then I want to upgrade Warpee to follow a skills based agent model so I can plug in more capabilities in the future and then I want to make it all available for agents behind x402 or MPP
2 replies
1 recast
8 reactions