Videos / AI

Episode 10 of 22 · AI, end to end

Vector databases and RAG, end to end

Chapters

1. The press has run0:00
2. And yours were never in0:14
3. Fluent and wrong0:29
4. Train it, or tell it0:46
5. Put it in the prompt1:04
6. Two pipelines, not one1:23
7. Text becomes a point1:42
8. No shared words, same place1:58
9. The angle, not the length2:15
10. Where the numbers come from2:34
11. The encoder is a decision2:53
12. What nearness is not3:12
13. The cut is the ceiling3:30
14. Too small loses its subject3:47
15. Too big dilutes the match4:07
16. Measure it in tokens4:26
17. Overlap buys the seam4:44
18. Split where the document splits5:01
19. The metadata rides along5:20
20. Small to find, large to read5:39
21. A library with no catalogue5:55
22. Approximate on purpose6:12
23. The layered graph6:30
24. Cells, and compressed vectors6:48
25. The dial nobody turns7:08
26. What the database adds7:25
27. Two witnesses7:45
28. Fuse the two lists8:04
29. Retrieve wide, rerank narrow8:24
30. The question is not the query8:41
31. Recall at k, measured8:59
32. The briefing folder9:17
33. Make it cite, let it refuse9:36
34. Failure one: it never arrived9:54
35. Failure two: it arrived and was ignored10:13
36. Score the halves apart10:30
37. The whole thing10:48

Vector databases and RAG, end to end. A model ships the way a printed encyclopaedia ships. Every fact it has was set on the day it went to press, and the press has already run. The weights are frozen numbers. Nothing t

Mark as watchedSaved in this browser only.