Embeddings & vector search, from scratch

Search lab: forty-two records become unit vectors of 384 float32 coordinates, 1,536 bytes each, and similarity becomes the angle between two arrows. Brute force scores all 42 (16,128 multiply-adds) and is correct by construction; on 1,024 records the layered graph reaches recall@10 0.9833 at ef=40 while touching 223.2 records, 21.79% of a scan, against 0.9875 for 372.6 at ef=100. Then the failures a vector cannot fix: "what does E4021 mean" returns its answer at rank 8 behind six incident records at 0.2273, because E4021 and E4022 sit 0.997252 apart, and the unanswerable question still returns ten rows whose best cosine is 0.0394.

This Memory Engineering lesson is part of a paid plan. Sign in with a plan that includes this course to read it.