Make TREC CAsT evaluation tractable and publish Gemini baseline - #8

Draft

bigboateng wants to merge 15 commits into

mainfrom

cursor/cast-gemini-eval-2551

Commits on Jul 13, 2026

Fix CAsT document store dependencies

cursoragent
and
bigboateng
committed
e7333fd
Scale kNN graph construction with NumPy

cursoragent
and
bigboateng
committed
dcd7309
Bound graph distance caching during evaluation

cursoragent
and
bigboateng
committed
669cab1
Accelerate kNN shortest paths with SciPy

cursoragent
and
bigboateng
committed
797b7cd
Vectorize full-corpus evaluation scoring

cursoragent
and
bigboateng
committed
1243a6b
Cache sparse graph distances compactly

cursoragent
and
bigboateng
committed
9a71f59
Type sparse distance results explicitly

cursoragent
and
bigboateng
committed
7dc5fd7

Commits on Jul 15, 2026

Checkpoint and retry Gemini rewrites

cursoragent
and
bigboateng
committed
bf3a642
Order Gemini baseline imports

cursoragent
and
bigboateng
committed
8b0b04b
Prefetch and pace Gemini rewrites

cursoragent
and
bigboateng
committed
1ffadb5
Use available Gemini Flash model for CAsT

cursoragent
and
bigboateng
committed
81d2509
Use stable Flash-Lite rewrite model

cursoragent
and
bigboateng
committed
34eb806
Bound Gemini request stalls

cursoragent
and
bigboateng
committed
dc1ec4f
Require real resolved queries for oracle baseline

cursoragent
and
bigboateng
committed
1d8ea0f
Publish TREC CAsT Gemini evaluation result

cursoragent
and
bigboateng
committed
d541982