Home▸AI & Machine Learning▸Retrieval Evaluation and Reranking Lab (2D)

Retrieval Evaluation and Reranking Lab (2D)

A 2D companion to the retrieval-pipeline lab: candidates sit on a radial map around the query, first-stage retrieval and reranking pull them inward, and the final top-10 lands closest to the center while Recall@10, nDCG@10 and latency update live.

AI & Machine Learning2DModerate60 FPS📱 Mobile-adapted⇄ 3D version

This 2D companion runs the same two-stage retrieval pipeline as the 3D version on a flat radial map: documents sit around a central query marker, first-stage retrieval and the reranker pull surviving candidates inward stage by stage, and the four sliders — pool size, rerank budget, hard-negative density and reranker strength — let you watch Recall@10 and nDCG@10 rise or fall against the simulated latency cost in real time.

⚙ Under the hood

2D retrieval pipeline lab with a radial candidate map, live Recall@10/nDCG@10/latency readouts, and the same first-stage-retrieval-then-rerank mechanic as the 3D version.

information retrievalRAGrerankingevaluation metricsrecallnDCG

2D · HTML5 Canvas 2D · 60 FPS target · runs fully client-side, no install

What does this simulation show?

A two-stage retrieval pipeline: a first-stage retriever pulls a broad candidate pool by embedding similarity, then a reranker re-scores a limited budget of those candidates before the final top-10 is chosen.

What do Recall@10 and nDCG@10 measure?

Recall@10 is the fraction of all relevant documents that made it into the final top-10. nDCG@10 additionally rewards ranking relevant documents higher within that top-10, normalized against a perfect ranking.

Why does a bigger rerank budget increase latency?

The reranker is a more expensive model than the first-stage retriever, so scoring more candidates (a larger N) or using a stronger reranker costs more compute time, which the simulated latency readout reflects.