Faster memory and sharper context

Memory answers in 1.2 s where it took 3.2 s

Improvement

A B2B distributor with 6,000 documents waited 3.2 seconds, on average, for its context answer. On bad days, 5 seconds. We turned on it the retrieval path that already runs in our own workspace. After 24 hours, the median was 1.2 seconds and the worst case 1.8. Each answer brought 9.4 documents instead of 7.9, and no call failed. The comparison uses real calls: 18 after the change against 140 from the previous eight days.

The right context shows up more often, with fewer documents

Improvement

We measured search against the real corpus of two accounts, asking each document about its own subject. One in four did not show up even among the first ten. Two causes. The filter that avoids repeating a similar document treated two documents from the same client as if they were the same memory, and pushed to the end of the queue the one that answered; it now only demotes an earlier version of the same document or a near-identical copy. The other cause was the cap on documents per answer, too low for a large corpus: with 8 instead of 6, the chance that the right document shows up among the first ten rises 7 points in a 6,000-document account. The delivered context is about 30% smaller than it was with 12.