Latency vs. corpus

growing the corpus...
Per-call latency as the corpus growsFrom 20 to 300 repositories, the original illustrative model increases ripgrep latency from 1.55 to 17.05 seconds and turbopuffer search from 0.29 to 0.348 seconds. Model completion remains 4.2 seconds. Both axes are logarithmic.
20 repos

Bash-only agent

15.3 ripgrep calls
27.9 s
23.7 s tools4.2 s model

Bash + turbopuffer

5.0 search calls
5.7 s
1.5 s tools4.2 s model
RIPGREP, PER CALL
1.55s
1.0× the 20-repo call
TURBOPUFFER, PER CALL
0.29s
1.00× the 20-repo call
ROLLOUT WALL-CLOCK
28s vs 6s
4.9× wall-clock apart
ripgrepturbopuffer searchmodel completion