ZacharyZhang-NY
c4a8c36416
Every number in CHECKLIST p2-1 and p0-3.7 came out of a scratch script: the
kind of artifact a reader cannot re-run, and the kind that quietly drifts
away from the code it describes. These reports are generated by committed
harnesses and each one stamps the commit that carries the code it measures.
DENSE.md 165,973 atoms / 153,857 vectors: build 1421.0s, search p50 75.6ms
p95 93.0ms, 0.998 agreement with a brute-force scan, one write into
a full index p50 518ms, 1,729 bytes per stored vector.
QUANT.md twelve encodings measured on the same vectors: int8 keeps 0.977 of
the ranking for a quarter of the bytes and no speed, bit is 14x
faster and keeps 0.728, so bit is a coarse filter and nothing more.
TERMS.md the stop-word bound swept, not guessed: 0.05 is the measured peak
(recall@10 0.803 at p50 69.5ms) and also the fastest of the rows
that keep recall.
SCALE.md the product's own search at 165,973 atoms: p50 71.7ms against the
frozen baseline's 20.16s.
LEGS.md what each retrieval leg is worth, per modality, with the page size
the product actually asks for in the table.
The acceptance bar the operator set is recall@10 >= 0.90 at n>=300. It is
not met. The best measured row is 0.858 (dense leg on, page widened to 30,
p50 414ms); the shipped configuration gives 0.848 at p50 72.5ms; the
in-repo counterpart at holder scope gives 0.866. The shortfall is
concentrated in picture golds (video 0.583-0.708, ui_screenshot 0.500-0.583,
image 0.750) while the word-shaped ones already pass (code 0.953, pdf 0.919).
Latency passes by three hundred fold; recall does not, and the gap is four
points of cross-modal retrieval, not index scale.
1.5 KiB
1.5 KiB
DENSE: the vector leg at corpus size (commit dff5977ab)
- database:
bench/runs/dense_scale/data/ely_eye.sqlite - atoms: 165973, memories with words: 153857, vectors: 153857, pending: 0
- model:
sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2, width 384 - build: 1421.0s for 153857 memories (108 per second)
- writing one memory into a full index: p50 518ms · p95 543ms over 50 writes
- the repair scan (
ely-eye build-dense-index) over a full index: 1.3s, and it is not on the search path - search p50 75.6ms · p95 93.0ms · max 123.6ms over 300 queries
- of which encoding the question: p50 12.4ms · p95 25.8ms
- against a brute-force scan: 0.998 of the top 10 over 40 probes (a returned atom counts when its true cosine is at least the k-th best, so ties cannot fail the check)
- bytes per stored vector: 1729 (266059776 bytes for 153857 vectors, dbstat page sizes over the vec0 shadow tables)
- machine: Linux 6.18.40.1-microsoft-standard-WSL2, 32 cores, 4 threads per query, load [3.18, 4.58, 5.42]
One million memories on this machine
- p50 455ms · p95 561ms
- vector table: 1.6 GiB
- first build: 153.9 minutes
- writing one memory: 3123ms — a vec0 delete by id scans the table, so this grows with the count too
- assumption: sqlite-vec 0.1.9 scans every vector, so latency and bytes grow linearly with the count; the model, the machine and the thread budget stay the same
Latency is wall time per search on the machine named above with the thread budget the benchmark pins; a run on a loaded machine is not comparable.