jina-reranker-v2-base-multilingual

Jina Reranker v2 Base Multilingual is a text-ranking model for the second stage of retrieval. Give it a query and candidate documents or passages; it evaluates each query–document pair and returns a relevance score. The useful result is a better-ordered candidate set for search or a retrieval-augmented workflow. It does not write the answer to the query or create document embeddings.

Prepare the query and the candidate passages

Write the query so that it states the information need. Then provide the texts that an earlier retrieval step has found. Keep a stable identifier beside each candidate so that its relevance score can be connected to the original document, section or record.

For a support search, use the actual question and candidate excerpts that contain the relevant procedure. A query about resetting a particular device should be paired with passages that retain the device name, version and reset conditions. A passage that merely mentions the device can be less useful than one that explains the requested operation. Supply the content the model needs to compare, rather than only a document title.

Use coherent passages. A table heading, definition or exception can change how a paragraph relates to the query, so do not remove it merely to make every candidate equally short. The reranker evaluates supplied candidates; it cannot recover a relevant document that the earlier retrieval stage omitted.

Query–document scoring

This is a cross-encoder: the query and a candidate text are processed together. The direct scoring interface returns a floating-point relevance score for each corresponding pair. Higher scores indicate greater relevance in that comparison. The score is a ranking signal, rather than a generated explanation or proof that the document is true.

The maker also provides a reranking helper that returns ordered documents with their scores. A top-result control selects how many ranked candidates to retain. Preserve their identifiers and text when passing them into the next stage. For a question-answering workflow, the following model still needs the query and the selected evidence; the ranked list itself is not an answered question.

When comparing retrieval changes, keep the query and evaluation criteria fixed. Inspect whether the passages promoted to the top actually contain the requested information. Set any filtering threshold using examples from your own collection, instead of treating a relevance score as a universal measure of answer correctness.

Model window and longer documents

The model supports a context length of up to 1,024 tokens. Its documented rerank() helper handles longer documents with a sliding-window approach, scores the pieces separately and combines them into document-level results. Chunking is behavior of that helper; it should not be confused with a single model input that has an unlimited context window.

The helper exposes query-length and chunk-length controls and an overlap setting. Its documented default overlap is 80. The published usage example specifies a maximum query length of 512 and a chunk length of 1,024; these are example helper settings that should be kept distinct from a promise that every caller or route uses them.

For a long manual, retain enough nearby text for an instruction and its conditions to appear together. When assessing the result, inspect the relevant portion of a highly ranked document rather than assuming that every section of it answers the query equally well.

Multilingual and specialized text retrieval

The v2 model is multilingual. The maker's examples rank English queries against English, German, Spanish, Chinese and Japanese passages. This supports a collection where relevant material is written in several languages without requiring every candidate to be translated into the query's language first.

The documented task scope also includes code retrieval and ranking text representations related to functions or SQL. These remain relevance-ranking tasks: the model evaluates the supplied text. It does not execute a function, run SQL or produce a tool result merely because those kinds of candidates appear in the retrieval set.

Edition and practical limits

Keep this v2 Base Multilingual profile separate from Jina's earlier English-only reranker, Jina Reranker v3 and multimodal reranking profiles. They are different models with different task and input characteristics. This profile's input is query–text pairs; image or video content needs its own suitable representation or model.

The maker's reference implementation supports direct scores and a document-ranking helper, with optional accelerated attention. Those implementation choices affect the way inputs are prepared and processed; they do not establish that every integration exposes identical controls.

Use the ranked output for multilingual search, support knowledge bases, document retrieval and evidence selection before generation. Review the top passages, keep the original candidate mapping and distinguish relevance from factual validity. A useful retrieval result brings the right supplied evidence forward while leaving the next task—answering, summarizing or deciding—to the appropriate workflow.