Retrieve once
Jina v5 Small turns the corpus and claims into vectors. Cosine search finds 20 candidates. These candidates are frozen for every reranker.
A RETRIEVAL EXPERIMENT / OCTOBER 2026
Good answers
start with better evidence.
Can a general decision model do the job of a dedicated reranker? Five approaches. The same scientific claims. Every result open to inspection.
300test claims
20 → 5same candidates, new order
—claims with evidence in top 20
1 GPUBGE reranking
01 / THE SCOREBOARD
What changed after reranking?
OBSERVED TIME / 20 PASSAGES
BGE runs locally; API times include cloud processing and network time. Shared embedding work is excluded. This is a deployment comparison, not a hardware speed contest.
DECISION MODELS / REPORTED API COST
—02 / FOLLOW THE EVIDENCE
Open the result. Check the claim.
Green marks a document in SciFact's relevance labels. Unmarked documents are not labeled relevant; this is not a medical fact-check. Raw scores use different scales across models; compare the document order.
03 / OPEN NOTEBOOK
A small experiment, with its limits visible.
Rerank independently
BGE ranks claim–passage pairs. Jev, Span-02 and D1 Ultra receive the same question about useful evidence, including findings that support or contradict the claim. Documents are ordered by relevance.
Check against labels
We evaluate the top 5 against SciFact labels. Error bars use 10,000 bootstrap samples. Paired comparisons use the same claims.
What these results can—and cannot—tell us
This measures evidence retrieval, not generated-answer accuracy. One English scientific dataset cannot establish a universal winner. Models may have seen benchmark material during training. Passages are shortened using both tokenizers; the same text reaches every method. Labels may miss useful unjudged documents. BF16 can create ties; ties keep the original search order. Local processing time and remote API time include different costs. Later models were added after earlier results were available. The same relevance question was reused without model-specific tuning. Probability calibration was not measured.