The retrieval stack
The cheapest retrieval system that could possibly work, and what it taught me before I reached for anything heavier.
In progress2 of 3 published · Feb 2025 – May 2025
Field
Machine learning in production - Test.
Everything between a question arriving and an answer leaving — what gets retrieved, how fast it serves, and the evaluation that tells you whether either one worked.
- 01The retrieval stack
- 02
- 03Serving at speed
Start here →Begins with Why my first vector store was a CSV — 1 minute, no prerequisites.
FirstRunning · 2 of 3
ThenComplete · 1 parts
Test Series
A deliberately over-built essay that puts every rendering feature the blog supports on one page — headings, math, a live diagram, an interactive 3D plot, code chrome, margin notes and figures — so nothing can hide.
Complete1 published · Jul 2026
ThenRunning · 1 of 5
Serving at speed
Once retrieval works, the next wall is latency — and most of the latency you can actually control turns out to be a question of scheduling, not hardware.
In progress1 of 5 published · Sep 2025