Tabular Foundation Models (TFMs) logopeterindia.net
Knowledge Hub

Reranker Models

Reranker models for search and RAG: hosted APIs, open-weight models, restricted licences, libraries and late-interaction tools.

35 topics•6 categories•Updated October 2026

What Are Reranker Models?

5 entries

A reranker takes a query and a set of candidate documents from a first-stage search and reorders them by relevance, so only the best evidence reaches the LLM. Cross-encoders and LLM-based rerankers score each query and document pair together, which is slower but usually more precise than comparing stored embeddings. Details below come from each provider's documentation or model card; descriptions are not rankings, so test candidates on your own data.

01Embedding ModelsEmbedding APIs and open-weight models, on this site. 02Vector DatabasesVector databases for first-stage retrieval, on this site. 03RAG Architectures ElucidatedChunking, reranking and production RAG, on this site. 04Efficient RAG ArchitectureRecall, expand, rerank, ground, on this site. 05RAG vs. Jev + RAGRelevance gating compared with rerankers, on this site.

Hosted Reranking APIs

9 entries

Managed rerankers, described from each provider's own documentation.

06Cohere RerankCohere lists rerank-v5.0-pro and rerank-v5.0-fast with 32k-token context, 100+ languages, calibrated 0 to 1 relevance scores and an optional Context Focus feature; rerank-v4.0 and rerank-v3.5 are earlier models. Rerank 5 accepts text only, and semi-structured data is passed as YAML strings. 07Voyage AI rerankersVoyage lists rerank-3 as its highest-accuracy reranker and rerank-3-lite as the latency-optimised option, both with 32,000-token context; rerank-2.5 and earlier models remain accessible. 08Jina Reranker APIJina describes a reranker for maximizing search relevance and lists jina-reranker-v3.5, v3, v2-base-multilingual and m0 models. 09ZeroEntropyZeroEntropy describes training small, specialized models, including rerankers and embeddings, for production AI systems. 10Amazon Bedrock reranker modelsAWS documents a Rerank API and reranking inside Bedrock Knowledge Bases that reorders retrieved chunks by relevance; reranking supports text only. 11Pinecone hosted rerankingPinecone documents hosted reranking with cohere-rerank-4-fast, cohere-rerank-3.5, bge-reranker-v2-m3 and pinecone-rerank-v0. It says cohere-rerank-3.5 is deprecated as of 1 July 2026. 12Google ranking APIGoogle Cloud documents a ranking API in Agent Search that ranks and reranks documents to improve search and RAG quality. 13Azure AI Search semantic rankingMicrosoft documents semantic ranking, which uses deep learning language models to rerank search results. 14NVIDIA NeMo Retriever rerankingNVIDIA documents text reranking microservices as part of NeMo Retriever for RAG and agentic retrieval.

Open-Weight Rerankers

5 entries

Models whose weights can be downloaded. Licences are from each Hugging Face model card; read the licence before commercial use.

15Qwen3-Reranker-8BAlibaba Qwen reranker in 0.6B, 4B and 8B sizes, paired with the Qwen3-Embedding series (Apache-2.0). 16BGE Reranker v2-M3BAAI multilingual cross-encoder reranker with about 0.6B parameters (Apache-2.0); also hosted by Pinecone. 17mxbai-rerank-large-v2Mixedbread reranker with about 1.5B parameters (Apache-2.0). 18GTE Multilingual Reranker BaseAlibaba-NLP multilingual reranker with about 0.3B parameters (Apache-2.0). 19ms-marco-MiniLM-L6-v2Small cross-encoder from the Sentence Transformers team with about 23M parameters (Apache-2.0), a common low-latency baseline.

Rerankers With Restricted Licences

2 entries

Models whose licences limit commercial use or use a custom licence. Check the terms before building a product on them.

20Jina Reranker v3.5About 0.6B parameters under CC BY-NC 4.0; Jina describes it as the successor to jina-reranker-v3 and directs commercial users to contact Jina. 21NVIDIA Llama Nemotron Rerank 1B v2About 1.2B-parameter reranker listed on Hugging Face under a custom licence rather than a standard open-source one; read the licence file.

Libraries & Late-Interaction Tools

7 entries

Libraries for running rerankers and related retrieval methods. Release dates are from PyPI.

22Sentence Transformers CrossEncoderDocumentation for using and training cross-encoder rerankers; the sentence-transformers package on PyPI was at 6.2.0 (21 October 2026). 23rerankers (AnswerDotAI)Lightweight Python library with a unified API for common reranking and cross-encoder models; PyPI shows 0.10.0 from May 2025, so check its activity. 24FlashRankLightweight re-ranking library for search and retrieval pipelines; PyPI shows 0.2.10 from January 2025, so check its activity. 25ColBERTStanford late-interaction retrieval model that scores token-level matches, an alternative to cross-encoder reranking. 26RAGatouilleLibrary for using and training ColBERT late-interaction retrieval in RAG pipelines; PyPI shows 0.0.9.post2 from May 2025, so check its activity. 27Passage Re-ranking with BERTPaper that introduced BERT-based passage reranking, the basis of many cross-encoder rerankers. 28ColBERT paperPaper describing efficient passage search with contextualized late interaction over BERT.

Frequently Asked Questions

What is a reranker model?

A model that scores how relevant each candidate document is to a query and reorders the candidates. It usually runs after a fast first-stage search, such as vector or keyword search, and passes only the top results to the LLM.

How is a reranker different from an embedding model?

An embedding model encodes the query and each document separately, so documents can be indexed ahead of time. A cross-encoder reranker reads the query and a document together and scores them jointly, which is slower but is usually more precise, so it is applied to a short candidate list.

Do I need a reranker for RAG?

Not always. Rerankers help most when first-stage retrieval returns many near-matches and the LLM context is limited. They add latency and cost, so measure retrieval quality with and without one on your own questions.

How many documents should I rerank?

Rerank enough candidates to cover the relevant ones, then keep a small top set for the LLM. Provider limits differ; for example, Pinecone documents 100 to 250 documents per request depending on the model. Test the candidate count against your latency budget.

What are late-interaction models such as ColBERT?

They keep a vector per token and score a query against a document by matching tokens, which sits between single-vector embeddings and full cross-encoders. Tools such as ColBERT and RAGatouille support this approach.

Are open-weight rerankers free for commercial use?

Not always. Qwen3-Reranker, BGE Reranker v2-M3 and mxbai-rerank list Apache-2.0, while Jina Reranker v3.5 lists CC BY-NC 4.0 and NVIDIA's Nemotron reranker uses a custom licence. Read each model card.

How current is this page?

Links and model details were checked on 31 October 2026 against provider documentation and Hugging Face model cards. Provider descriptions follow their own wording and are not rankings, and benchmark scores are not reported because they change.