Lesson 3: Embedding variants: smaller vectors, token vectors, sparse terms
28 min read·13 Oct 2026
Module 4 · Lesson 3 · About 50 minutes · Needs: Lessons 1 and 2 of this module (unit-length vectors, cosine scores, prefixes, the domain-fit test)
The situation
Kwame has read that binary vectors are 32 times smaller and wants to switch them on before the partner launch. Owen asks what each stored vector costs. Hannah has one condition: whatever Hollin stores, a search for "batch B-2291" must still put the X200 jar recall notice at the top. Ellie's blender is an X200, so its jar could be from that batch.
Grace adds a question: Kwame keeps hearing about token-level and sparse embeddings that match codes like that batch number better. Are they worth it? The running question is q13, "batch B-2291". The Hollin data is simulated.