Reranking

Every reranker sits at the same seam in the retrieval-augmented generation stack: downstream of the search that decided what a model could possibly see, upstream of the answer it gives. Get that seam wrong and no amount of prompt tuning recovers a document that never made the shortlist; get it right and one second-stage pass often buys more precision than a bigger first-stage retriever ever will. That trade only pays off once you know which candidates are worth the expensive rescore and which reranker actually earns its latency budget — which is what this topic’s six articles walk through in order.

The reranking reading path: from mechanism to market

Start with what reranking is and how cross-encoders rescore retrieved documents — the shortest path to the mechanism every later article assumes: a second, heavier model reading each query-document pair together instead of separately. Then read the prerequisites behind cross-encoders, bi-encoders, and listwise scoring in the same sitting — it maps the three reasoning shapes a reranker can take, which decides how much latency you are actually buying. Before you wire one into a pipeline, the hard technical limits of cross-encoder reranking is worth the ten minutes: latency scales quadratically with token length, and relevance learned on general benchmarks drifts on specialized domains.

When you are ready to add one, the integration guide for Cohere Rerank-4-Pro, Voyage Rerank-2.5, and Zerank-2 compares the managed APIs against self-hosting and flags the licence trap hiding in one of them. For where the market moved since, the open-vs-closed reranker race on the 2026 Agentset leaderboard tracks how fast open-weight models closed the quality gap. Close with the ethical cost of outsourcing search ranking to a closed API — if a vendor’s scoring decides what your users see, read it before the contract renews, not after.

How reranking differs from hybrid search and contextual retrieval

Two neighbours get folded into “reranking” in planning docs, and the confusion changes which fix a team reaches for.

• Reranking is not hybrid search. Hybrid search runs before reranking and decides what gets recalled — fusing dense and sparse retrieval so relevant documents make the candidate list at all. Reranking never widens that list; it only reorders what hybrid search already found. A team that adds a reranker to fix a recall problem is solving the wrong stage.
• Reranking is not contextual retrieval. Contextual retrieval acts before indexing, prepending context to a chunk so it is findable in the first place. Reranking acts after retrieval, on chunks that were already found. Fixing missing context with a heavier reranker treats the symptom one stage too late.

Common questions about reranking

Q: Should I use a hosted reranking API or self-host an open-weight model? A: Self-hosting BGE Reranker v2-m3 or mxbai-rerank-large-v2 only pays off once you have measured retrieval quality and have the serving budget to run it; otherwise a managed API is the faster path. The integration guide compares both.

Q: Are open-weight rerankers now as good as Cohere or Voyage? A: On the Agentset leaderboard, yes — a 4B open-weight reranker now leads the ELO ranking ahead of Cohere’s flagship, and the absolute quality gap on nDCG@10 has closed too. The 2026 leaderboard race tracks the shift.

Q: Does a reranker’s licence matter if I only call it through a hosted API? A: Yes — hosted access does not erase the underlying licence. Zerank-2 ships under a CC-BY-NC licence that blocks commercial use even through a paid endpoint, exactly the kind of fine print outsourced ranking decisions hide from procurement.

Q: Why does a reranker that scored well in benchmarks misrank my legal or financial documents? A: Most cross-encoders are trained on MS-MARCO-style web queries, and that relevance judgment drifts on legal, financial, and argument-heavy text with a different notion of “relevant.” The hard limits of cross-encoder reranking covers this wall alongside latency.

Part of the retrieval-augmented generation theme · closest neighbour: hybrid search. Coming to RAG pipelines from classical software engineering? Start with the story: Debugging RAG Failures: Why Developers Need a New Diagnostic Model.