Where it started (the one-model, one-table rework):
https://t.co/kh8r0jxL8P
Full write-up with the SQL, the eval setup, and the migration details:
https://t.co/DtsVFnRItV
5 months ago I collapsed a production RAG pipeline down to one embedding model and one table.
Since then I found out my "hybrid search" wasn't actually hybrid. The BM25 arm had been dead the whole time and nothing warned me.
Here is what production RAG taught me 🧵
Same conclusion as 5 months ago, just stronger now:
Simple architecture beats clever architecture. But simple architecture with evals beats everything, because simplicity you can't measure is just hope.
Original story and full write-up below 👇