Skip to content

Latest commit

 

History

History
90 lines (69 loc) · 3.45 KB

File metadata and controls

90 lines (69 loc) · 3.45 KB

LongMemEval-S Semantic Cache Path Report

Date: 2026-06-12

Issue: #1305

Sanitized JSON: longmemeval-semantic-cache-path-2026-06-12.json

Decision

Status: semantic_warm_query_cache_path_replay_report.

This report completes the warm query/candidate cache slice for #1305. It does not reuse the #1193 structural failure report as completion evidence. The structural 500Q report remains a failure boundary; this follow-up measures a separate semantic/cache path over a fresh 25Q opt-in semantic pilot with candidate-pack hashes in the sanitized case rows.

Cache Path

The measured path is B: warm query/candidate cache.

Metric Value
Semantic pilot cases 25
Available cold-fill calls 24
Complete query/candidate cache keys 24 / 24
Cold-fill cache misses 24
Warm replay hits 24
Warm replay hit rate 1.0000
Two-pass hit rate 0.5000
Warm lookup latency, avg / max 0.000079ms / 0.000300ms
Cold-fill latency, avg / max 6308.67ms / 21006.74ms
Cold-fill tokens 223947
Provider prefix-cache hit rate 0.0000

The cache key includes query hash, candidate window/span hash, reranker prompt version, model/provider id, source or dataset fingerprint, and policy version. The committed report does not emit cache key samples, cache values, raw candidate text, source text, questions, answers, provider responses, credentials, or local paths.

Runtime now emits line_reranker_candidate_pack_sha1 for reranker rows. That hash is derived from candidate line/routing metadata plus source-text hashes, not raw text. It gives a real candidate-window invalidation input without turning the report into a source-text leak.

Exact-Line Metrics

The semantic pilot remains narrow: all 25 cases are longmemeval_single-session-user.

Metric Value
First-stage evidence-line R@10 23 / 25 = 0.9200
Semantic-only evidence-line R@10 24 / 25 = 0.9600
Fused reranked evidence-line R@1 23 / 25 = 0.9200
Fused reranked evidence-line R@3/R@5/R@10 25 / 25 = 1.0000
Fused reranked evidence-line MRR 0.9600
Rerank regression count @10 0

Miss-family focus:

Miss family Cases Recovered by fused rerank
context_visible_exact_line_miss 2 2
exact_line_found_top_k 22 22
multi_evidence_partial_hit 1 1

The two context-visible exact-line misses were also same-session wrong-line cases in this pilot, and both were recovered by the semantic reranker before the warm-cache replay.

Boundaries

Cold online semantic rerank remains explicit opt-in and must not become a default hook path. The full 500Q semantic projection from this pilot is about 4478940 tokens and 52.57 single-worker minutes, with provider dollar cost not reported by the chat-completions response.

The warm replay proves query/candidate cache-key completeness and hot lookup mechanics for repeated identical queries on the sanitized pilot. It does not claim 500Q semantic quality, source-side semantic cache quality, live hook latency, official LongMemEval QA score, answer-generation quality, or SOTA.

The source-side semantic cache path remains unmeasured: it still needs a source-span warming job, offline build cost, hot-path latency, stale/invalidated cache counts, and comparison against lexical, structural, and cold semantic arms before it can be promoted.