Match each 2026 reranker offering to its distinguishing trait

Match pairs·Medium·4.0 · 0·~2 min·Asked atDatadog IBM Jpmorgan

Attempt it

Drag each answer to line up with its matching prompt

Cohere Rerank 3.5

Open source, self-hostable, several sizes from base to large with permissive license

Voyage Rerank-2

Domain tuned variants (code, legal, finance) on top of a general reranker

BGE Reranker v2 (BAAI)

Hosted multilingual cross-encoder with strong latency for top-100 in single digit ms

Jina Reranker v2

Late interaction model, token level scoring rather than sentence level

ColBERT v2 / PLAID

Open source with long context support up to 8K and multilingual training

Five rerankers, five distinguishing traits. **Cohere Rerank 3.5** is the dominant hosted cross-encoder in 2026. It is closed source and called via a paid API, but it is fast, Cohere reranks a top-100 candidate set in single-digit milliseconds on their managed infrastructure. It is multilingual out of the box. **Voyage Rerank-2** is also hosted, but its differentiator is domain-tuned variants. Voyage ships specialized rerankers for code, legal, and finance on top of a general reranker. If your corpus is one of those domains, the specialized variant typically outperforms the general one. **BGE Reranker v2** from BAAI is open-source with a permissive license. You can self-host it. It comes in several sizes (base, large, m3) so you can pick a tradeoff between latency and quality. It is the default choice when you cannot or do not want to send your corpus to a hosted API. **Jina Reranker v2** is another open-source self-hostable reranker. Its differentiator is long-context support up to 8K tokens and multilingual training. If your chunks are long or your corpus is multilingual, it is a strong fit. **ColBERT v2 / PLAID** is architecturally different from the others. The others are sentence-level cross-encoders that produce one score per (query, chunk) pair. ColBERT is a late-interaction model, it scores at the token level, comparing each query token against each chunk token, then aggregates. The benefit is finer-grained matching; the cost is more memory and storage for token-level embeddings. The four sentence-level cross-encoders (Cohere, Voyage, BGE, Jina) differ in hosting model and specialization. ColBERT is the architectural outlier.

The five rerankers split along two axes: **architecture** (sentence-level cross-encoder vs token-level late-interaction) and **hosting** (closed hosted API vs open self-hostable weights). **Cohere Rerank 3.5**, closed, hosted, multilingual, low-latency. The dominant production default for teams willing to pay per call. Optimized batching means top-100 reranks in single-digit milliseconds. The closed nature is the limitation: vendor lock-in and per-call cost. **Voyage Rerank-2**, closed, hosted, with **domain-specialized variants**. Voyage ships dedicated rerankers for code, legal, and finance corpora on top of a general reranker. The specialized variants beat the general reranker on their domains by meaningful margins. Pair naturally with Voyage embedding models. **BGE Reranker v2**, open-source from BAAI, permissive license, self-hostable. Available in base, large, and m3 sizes for latency vs quality tuning. The 2026 default when you need to self-host (data residency, on-prem, cost control). Pairs naturally with BGE-M3 embeddings. **Jina Reranker v2**, also open-source and self-hostable. Differentiator is **long-context support up to 8K tokens** and multilingual training. Best fit when chunks are long (e.g. legal sections, full pages) or the corpus spans many languages. **ColBERT v2 / PLAID**, architecturally distinct. The other four are sentence-level cross-encoders: they take a (query, chunk) pair and produce one relevance score from a single forward pass. ColBERT is a **late-interaction** model: it stores token-level embeddings for chunks, and at query time it scores each query token against each chunk token (the maxsim operation) and aggregates. The benefit is more granular matching, especially on long chunks. The cost is storage (token embeddings, not sentence embeddings) and somewhat more complex serving infrastructure. The practical default in 2026 is Cohere Rerank 3.5 for hosted production, BGE Reranker v2 for self-hosted, Voyage Rerank-2

TL;DR

Cohere Rerank 3.5 is the hosted multilingual default; Voyage Rerank-2 ships domain-tuned variants; BGE and Jina are open-source self-hostable; ColBERT is architecturally different with late-interaction token-level

Memory aid

Easy to grasp

Imagine five careful interviewers, each with a specialty. The first one is the fast-talking generalist you hire from an agency. The second works at a firm with specialist hires for legal, finance, or coding interviews. The third is freelance and you can hire them onto your own team. The fourth is also freelance and good with long resumes in any language. The fifth has a completely different style, they read the resume word by word instead of skimming the whole thing. All five interviewers re-rank a small shortlist, but each has a personality and a quirk worth knowing.

Key concepts

Context engineering Retrieval-augmented generation Embeddings

Concept explanation~2 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

The 2026 reranker landscape has five offerings worth knowing. Four of them are sentence-level cross-encoders that differ along hosting and specialization axes; the fifth is architecturally different and trades storage for token-level matching granularity.

This card walks each one through its distinguishing trait, and frames the choice as two orthogonal decisions: architecture and hosting.

Cross-encoder vs late-interaction architecture

Sentence-level cross-encoders (Cohere Rerank 3.5, Voyage Rerank-2, BGE Reranker v2, Jina Reranker v2) take a [CLS] query [SEP] chunk [SEP] input and produce one relevance score per pair via attention across the joint sequence. One forward pass per (query, chunk) pair. Conceptually simple, fits standard transformer serving, well-supported across vector databases.

ColBERT v2 / PLAID is a late-interaction model. Each chunk is encoded into a sequence of per-token embeddings, stored persistently in an index alongside the chunk text. At query time the query is also encoded per-token. The relevance score is the MaxSim aggregation: for each query token, find its maximum similarity to any chunk token, then sum those maxima.

What this buys

Granularity. A sentence-level cross-encoder collapses the whole chunk into one score, which means a perfect match on one sentence and weak matches everywhere else can produce the same score as moderate matches everywhere. Late-interaction sees the per-token alignment directly and can credit the perfect-match sentence appropriately.

What it costs

Storage. Chunks need per-token embeddings (tens of vectors per chunk) rather than one chunk-embedding. The index grows by 1-2 orders of magnitude. PLAID is the optimized serving system that uses quantization and centroid pruning to make late-interaction serving practical at scale, but the operational complexity is higher than a sentence-level reranker.

Closed-API rerankers, Cohere and Voyage

Open-weight rerankers, BGE and Jina

Picking a reranker in 2026

Situations where this technique stops working.

2–4 min · Everything important, quickly.

Reranker	Hosting	Architecture	Distinguisher
Cohere Rerank 3.5	Closed hosted API	Sentence-level cross-encoder	Low-latency multilingual default
Voyage Rerank-2	Closed hosted API	Sentence-level cross-encoder	Domain-tuned variants (code, legal, finance)
BGE Reranker v2	Open / self-host	Sentence-level cross-encoder	Permissive license, multiple sizes
Jina Reranker v2	Open / self-host	Sentence-level cross-encoder	Long-context up to 8K, multilingual
ColBERT v2 / PLAID	Open / self-host	Late-interaction (token-level)	MaxSim token scoring, finer granularity at storage cost

Real products, models, and research that use this idea.

Cohere Rerank 3.5 is the dominant hosted reranker for production RAG stacks in 2026, integrated natively into Pinecone, Weaviate, Qdrant, and the major orchestration frameworks.
Voyage AI's Rerank-2-code variant is widely used in coding-agent stacks (Cursor, Windsurf, Cline) where corpus is source code.

What an interviewer would ask next. Try answering before peeking at the approach.

QWhen is ColBERT's late-interaction architecture worth the storage cost?

Long chunks where sentence-level scoring loses information, or queries with rare-token alignments. Run a head to head eval against a cross-encoder before committing to the storage overhead.

2 more follow-ups an interviewer would ask next. Sign in to reveal them.

Red flags & common mistakes

The phrases that signal junior thinking. Click to expand.

Most common mistake

Treating all five rerankers as interchangeable. Architectural differences (cross-encoder vs late-interaction) and hosting models (closed API vs open weights) and domain tuning each change which one is right for a given workload.

60 second bullets to scan on the way to the call.

State which reranker is the architectural outlier and why
Distinguish hosted (Cohere, Voyage) from self-hostable (BGE, Jina) options

Primary sources. Browse if you want the original framing.