BM25 retrieves. JEV reranks.

Explore the TREC-1 Wall Street Journal subset: 173,252 indexed articles across its 1987–1992 volumes, about 0.5 GB of source text. Each saved BM25 search returns 1,000 articles; JEV scores the first 100 as complete documents.

NIST collection statistics

Sample of 20 search queries

BM25 order

Ranking quality · top 100

MetricBM25JEV

0.0000

—

0.0000

—

0.0000

—

0.0000

—

0.0000

—

0.0000

—

Percentage change is relative to BM25. Recall@100 stays fixed because JEV reorders the same 100 candidates.

Results

Top 10 shown · first 100 reranked

Human judgments

01
WSJ880517-0142Not relevant
Diagnostic/Retrieval To Acquire Photronics
—
02
WSJ880816-0130Not relevant
Diagnostic/Retrieval Systems
—
03
WSJ870303-0004Not relevant
Diagnostic/Retrieval Bid for Rospatch Corp. Is Rejected by Firm
—
04
WSJ861205-0165Not relevant
Who's News: Diagnostic/Retrieval Systems Inc.
—
05
WSJ880525-0156Not relevant
Who's News: Diagnostic/Retrieval Systems Inc.
—
06
WSJ870304-0091 Relevant
Kodak Unveils Data Retrieval, Storage Systems
—
07
WSJ920224-0020Not relevant
Fidelity Follows Other Discount Brokers In Raising Commission Rates on Trades
—
08
WSJ870429-0078 Relevant
Kodak Introduces 12 New Data Systems
—
09
WSJ911212-0022 Relevant
SEC Postpones Opening Of Electronic Filing System
—
10
WSJ870305-0052Not relevant
Corrections & Amplifications: Dow Jones News/Retrieval
—

Ranks 101–1,000 receive no JEV scores.Saved experiment replay

Across 50 TREC-1 topics

Top-ten ranking quality on fixed BM25 candidates.

nDCG@100.4640 0.6634BM25 → JEV

Reranking time211.5 sUncached run

Estimated API cost$0.328Successful responses

5,000 scored articles · jev-1.13.0 · recorded replayFull-run report