VoyageAI by MongoDB: rerank-2.5

rerank-2.5 is Voyage AI's instruction-following reranker, designed to refine a set of search candidates. It jointly examines the query and each document as a cross-encoder, then returns a relevance ordering. The maker's release emphasizes natural-language control over what counts as relevant, multilingual retrieval and a longer query–document context than the preceding rerank-2 edition.

Give the model the candidates to compare

The input is a query and a list of candidate text documents. These can be passages selected by embedding search or a lexical first-stage search. The reranker evaluates the supplied text rather than creating corpus embeddings, so its practical job is to refine the candidates the initial search already found.

Keep the original candidate ordering and source identifiers when preparing that list. The returned indices refer to those input positions. For a RAG task, supply the passages the answer stage might actually use, including the title or qualifications needed to interpret them. A relevant document that is absent from the submitted candidate set cannot appear as a newly found result of reranking.

Specify the intended notion of relevance

Instructions can be appended or prepended to the query. The release describes three useful roles: emphasize a particular part of a document, prefer a type of source and disambiguate the query with application context. These alter the relevance task presented to the model.

For a technical search, an illustrative instruction is “Prioritize passages that explain the failure and recovery behavior; introductory overviews are secondary.” For a research collection, specify whether the title, method or result matters most. For an ambiguous term, state the subject area the search is about. Keep the question and the instruction compatible, so one does not ask for a broad topic while the other silently changes the task.

These preferences guide relevance scoring. If the application has strict eligibility requirements—such as source version, date or access scope—prepare the eligible candidate set separately. The model's relevance instruction and a deterministic collection restriction serve different purposes.

Ranked output and later result use

Results are ordered by descending relevance and include a floating-point relevance score and the candidate's original input index. The top-k option selects how many leading results to return; without it, the manufacturer specification returns results for all submitted candidates. An optional return-documents setting includes the corresponding source text; otherwise the response contains indices and scores.

Use the ordering to select passages for display or a later answer stage, keeping their source identity intact. Inspect the actual text when evaluating the ranking: the score is a measure of relevance to the submitted query, not a generated answer or a verification that the document's claims are correct. This makes the model useful for refining technical, legal, financial, medical, conversation and multilingual search material without converting ranking into a domain conclusion.

Context, total-token budget and truncation

The documented query limit is 8,000 tokens, including instructions placed in the query. The query plus any one candidate document must fit within 32,000 tokens, and the document list can contain at most 1,000 items. The total allowance is 600,000 tokens, counted as query tokens multiplied by the number of documents, plus all document tokens. A long instruction is therefore counted repeatedly across the candidate pairs.

The output top-k count does not change the number of candidates in that input-token calculation. Default truncation shortens over-length queries or documents to meet the context limit; disabling it makes an over-length query or pair return an error. Prepare answer-bearing passages deliberately and keep the relevant qualification within the text the model will see. Evaluate the refined order on representative questions and instructions, then pass the selected original passages to the next stage.