bge-reranker-v2-m3

BGE Reranker v2 M3 is a multilingual relevance model built on BGE-M3. It receives a query and a candidate passage together and returns a score for their relationship. BAAI positions this edition as an efficient multilingual reranker, including for Chinese and English retrieval, where a search application needs to refine an existing list of candidates.

Relevance from a query–passage pair

A cross-encoder processes both texts jointly, allowing their content to interact before the relevance score is produced. This differs from an embedding model, which encodes texts independently for reusable vector comparison. The reranker’s output is a scalar score for a particular pair, rather than a document vector or a written answer.

Use the model after an initial search has selected plausible passages. Pair the question with each candidate and order the candidates by the resulting scores. This focuses the more detailed comparison on a manageable shortlist instead of asking a cross-encoder to inspect every passage in a large collection.

The v2 M3 edition

The maker identifies BGE-M3 as the base model and lists multilingual support as a principal feature. Within its v2 reranker family, M3 is the efficient multilingual option; the Gemma-based and layer-selection variants are distinct models with different implementation choices. The name M3 here does not mean the reranker returns the dense, sparse and multi-vector embeddings of the original embedding model.

This distinction matters when choosing the result your application requires. Select an embedding model to represent and search a document collection. Select this reranker when you already have candidate text and need to decide which passage is most relevant to the current query.

Preparing candidates that can be ranked usefully

Provide the actual question and the complete candidate text needed to interpret it. Preserve headings or nearby definitions when they determine the meaning of a passage. In the published Transformers example, the query and passage are tokenised together with truncation and max_length=512. That is an example setting rather than a model-wide limit. Keep the question and required evidence within the selected pair budget, and check what truncation removes. A result mentioning the same product as a query can still be irrelevant if it addresses the wrong operation or condition; prepare examples that distinguish those cases when evaluating retrieval quality.

For a document assistant, keep candidate identifiers alongside their text, score the question–passage pairs and pass the selected original passages to the answering step. The answering model then receives the evidence rather than a relevance score without its source. This is an illustrative retrieval workflow; the reranker itself neither searches an external collection nor writes the answer.

Reading and normalising scores

The manufacturer returns raw relevance scores, with a higher value indicating a closer match. Its documented optional normalisation applies a sigmoid function to place the score between 0 and 1. A normalised score has a convenient range, but that transformation alone does not establish a calibrated probability that the candidate is correct.

Choose ranking or filtering thresholds using representative query–passage pairs from your workload. Include passages that share keywords but answer a different question, as well as relevant passages with different wording. This makes the selected threshold meaningful for the collection being searched.

For an Ailoot API project, define what the final context should contain and how many candidates can be examined efficiently. The useful result is a better ordered evidence set for the next application step. Keep this edition’s relevance output separate from document embeddings, generated prose and the layer-selection controls documented for other rerankers.