0 Fork 0

Files

T

ZacharyZhang-NY 6de7439db6 Score the product path against corpus ids, not storage ids

The frozen baseline claimed the product's top-10 shared 0.0% with the fast
BM25 path. That was a harness bug: gold names a corpus source
(mmlongbench:PH_2016.06.08_Economy-Final.pdf) while the store names a content
hash (PH_2016_06_08_Economy_Final_10e57152f44d1a2b), so the two id spaces never
met and the right document, ranked first, scored as a miss.

- bench/systems.py: `_corpus_ids` joins store sources through the
  `source_map()` the counterpart already uses, and refuses a store whose
  sources are not corpus files instead of silently scoring zero.
- bench/systems.py: `live_product_rankings(..., as_holders=...)` asks the
  product as the query's own holder, which is how the counterpart is scored.
- bench/run.py: the live section now reports recall@10 / MRR / hit@10 with a
  95% lower bound, and probes 300 queries instead of 40 — the SQL path answers
  in well under a second, so the old cap was a cost estimate that expired.
- bench/run.py: `--name CURRENT` writes a post-P0 report beside the frozen
  BASELINE without touching it, stamped with the commit and the store schema
  version the pass started on (the live pass migrates the store mid-run, so a
  plain stamp would be a lie).
- bench/stamp_holders.py: the pre-P0 store holds 165,973 memories nobody owns.
  The corpus index states the owner, so holders, scenarios and event times are
  stamped in place — 165,973 atoms across 50,278 sources, 0 left unowned.

Numbers, same machine, load 1.0-1.7:
- CURRENT (post-P0, schema v5): no_rerank recall@10 0.565 p50 0.643s;
  fusion+rerank 0.453 p50 7.000s; graph 0.001.
- Product as its holder: recall@10 0.870, MRR 0.740, n=173.
- Counterpart: mem0_flat holder scope 0.866, whole corpus 0.412; the holder
  boundary is worth +0.454 with an identical ranker and identical items.

Verification: pytest backend/tests bench/tests -q -> 173 passed; ruff and mypy
(fresh cache) clean on every touched file.

2026-10-07 19:32:32 -04:00

3.3 KiB

Raw Blame History

BASELINE_EXTERNAL: memory counterparts on the same corpus

generated 2026-10-07T19:16:08+00:00 at commit 3d792417cf
5308 queries (2928 answerable) against 50789 sources
machine: 32 cores, load average 2.9 when the report was written
extraction is lexical for both stores; Mem0's real add() and Zep's graph builder call a generator model, which this round does not have
chance: random guessing scores recall@10 0.00020 over the whole corpus and 0.3686 inside one holder's memory (median 43 sources per holder); read each row against its own floor

store	items	closed	recall@10	hit@10	hit@10 Wilson 95% lower	MRR	abstain (unanswerable)	p50 s	p95 s
mem0_flat [holder scope]	403859	0	0.866	0.867	0.855	0.629	0.000	0.083	0.105
mem0_flat [whole corpus]	403859	0	0.412	0.416	0.398	0.283	0.000	0.086	0.112
zep_temporal [holder scope]	403859	27072	0.866	0.867	0.854	0.627	0.058	0.084	0.113
zep_temporal [whole corpus]	403859	27072	0.429	0.433	0.415	0.297	0.000	0.094	0.121

n per modality: {"code": 2474, "conversation": 300, "image": 480, "pdf": 984, "ui_screenshot": 640, "video": 430}

[holder scope] answers only from the asking holder's memory; [whole corpus] searches every holder's memory, which is what the product's arms are shown. The gap between the two rows is the value of the holder boundary, not of a better ranker.

mem0_flat [holder scope] by modality

modality	n answerable	recall@10	hit@10	Wilson lower	MRR
code	1237	0.952	0.952	0.939	0.695
conversation	150	0.993	1.000	0.975	0.927
image	240	0.646	0.654	0.592	0.505
pdf	766	0.916	0.916	0.895	0.692
ui_screenshot	320	0.622	0.622	0.568	0.330
video	215	0.716	0.716	0.653	0.396

mem0_flat [whole corpus] by modality

modality	n answerable	recall@10	hit@10	Wilson lower	MRR
code	1237	0.349	0.349	0.323	0.216
conversation	150	0.500	0.573	0.493	0.373
image	240	0.525	0.533	0.470	0.362
pdf	766	0.683	0.683	0.649	0.515
ui_screenshot	320	0.075	0.075	0.051	0.033
video	215	0.116	0.116	0.080	0.069

zep_temporal [holder scope] by modality

modality	n answerable	recall@10	hit@10	Wilson lower	MRR
code	1237	0.952	0.952	0.939	0.695
conversation	150	0.978	0.987	0.953	0.905
image	240	0.646	0.654	0.592	0.505
pdf	766	0.916	0.916	0.895	0.692
ui_screenshot	320	0.622	0.622	0.568	0.330
video	215	0.716	0.716	0.653	0.396

zep_temporal [whole corpus] by modality

modality	n answerable	recall@10	hit@10	Wilson lower	MRR
code	1237	0.350	0.350	0.324	0.216
conversation	150	0.591	0.667	0.588	0.429
image	240	0.562	0.571	0.508	0.426
pdf	766	0.685	0.685	0.652	0.519
ui_screenshot	320	0.119	0.119	0.088	0.057
video	215	0.163	0.163	0.119	0.087

3.3 KiB Raw Blame History