Trigram candidates, then regex verify
Candidates first, proof after: /appl./ over 11 docs
trigram app proposes, the regex disposes (teaching data)
Lexical pipeline: filter, then verify
regex /appl./
fixed: app
trigram app
8 candidates
verify each
doc 10 drops
7 matches
exact hits
->
->
->
red AND apples: intersect postings red and apples to doc 0
no fixed trigram: fall back to full scan or selective filter first
IVF probe vs recall@3
nprobe 1: 0.900
nprobe 2: 0.967
nprobe 4: 1.000 exact
ground truth is exact top-k
Ranking, latency and tenants (teaching rules)
hybrid score: lexical signal plus vector signal, one ranker
report p95 latency beside recall, never recall alone
tenant filter before vector search, or one index per tenant
fresh writes visible only after the index refresh window