Lesson 1: Inside an embedding: vectors, distances and odd scores

Module 4 · Lesson 1 · About 45 minutes · Needs: Module 1 (top-k retrieval by cosine); the basics are also in the AI Engineering course, Module 4 Lesson 2 and Module 5 Lessons 2 and 3

The situation

Kwame's prototype found the right page for "the smoothie maker stopped and felt hot": the X200 error codes, where E7 means the motor has overheated. That is Ellie Shaw's fault too. The question and the page share one word, "the". Grace asks how.

Kwame has the opposite worry. To drop weak matches he tried a cut-off of 0.5: almost every document passed. With a smaller model, nothing passed. Both questions need the same answer: what an embedding is, how it is made, and what its scores mean. The Hollin data is simulated.