Day 38 results: recall versus speed, the one dial

Measured on: (machine, OS, Python version)

Output

Paste part 1, part 2, part 3 and the scoreboard here.

One sentence per prediction

  • P1 (brute-force recall@k):
  • P2 (IVF nprobe=1 recall@k):
  • P3 (IVF nprobe=1 speedup over brute force):
  • P4 (IVF nprobe=8 recall@k):

The wall, in my own words

Brute force was exact, 100 percent recall. Why is it exact, and why does its cost grow with the number of vectors:

The IVF bargain

Probing one cell, what recall and what speedup did I get, and what exactly caused the recall it missed:

The dial

As I turned nprobe up, what happened to recall and what happened to latency, and at what point did probing more stop buying any recall:

The number that matters

At one cell, IVF hit __ percent recall at __ times the speed of brute force. Turn the dial to 8 cells and recall came back to ___ percent. In one line, why “approximate” is a setting you choose, not a flaw:

Can’t explain yet