CourseAI System Design: The Complete Course · Module 14: Scale and Evolution · part 95 of 107
Part 95 · Module 14: Scale and Evolution

Lesson 2: Changing the architecture without stopping it

32 min read·
Module 14 · Lesson 2 · About 50 minutes · Needs: Module 3 lesson 3 (re-embedding cost), Module 6 lessons 1 and 3 (embedding choice, the score floor, the alias swap), Module 7 lesson 4 (model change management). No key and no model calls; two embedding models download from Hugging Face

The situation

Week 14, Monday 1 February 2027. Gareth asks: "Which of today's choices will you regret in a year, and what does undoing each one cost?" Helen wants one of them undone now, as routine work: the embedding model. MiniLM reads 256 word pieces, so Module 6 capped chunks at 180 words, while real documents average 6,000 tokens.

The candidate is BAAI/bge-base-en-v1.5 (MIT licence, read on the Hugging Face Hub on 19 October 2026): 768 dimensions, 512 word pieces. The running example is G028: t-tallis-u035 asks "How many days of annual leave do I get?"; the answer is 30 days, in t-tallis-d002.

The rest of this course is yours to keep

This course is bought on its own, once, and stays readable afterwards, including the parts added to it later.

Or open everything with a plan