Skip to content

Multilingual Relevance Reranker

SageMaker models SageMaker model

Put retrieved English, Spanish and Portuguese passages in relevance order before you select context for a search result or a language model. You supply the passages and apply access and version filters. The SageMaker model package returns the original source IDs, indexes and uncalibrated relevance scores.

Delivery
SageMaker model package in your own AWS account

Where it fits

Use a second ranking stage after keyword or vector retrieval. Keep the exact submitted source pool so you can resolve each result and inspect its text. The package uses the original BAAI/bge-reranker-v2-m3 weights; it adds offline serving and SageMaker integration, with no Software Sushi fine-tune.

A ranking does not verify a complete or correct answer. An unrelated question still receives a ranking. You need source review and an application-level policy for deciding whether to answer.

Deployment and billing

The qualified region is us-east-1. Use ml.g5.xlarge for real-time inference and ml.g6e.xlarge for batch, with the explicit GPU AMIs in the getting-started guide. Enable network isolation. Weights and dependencies are bundled; inference does not download a model or call a seller service.

Software costs USD 0.50 per instance-hour, with a 14-day software trial. AWS infrastructure and other AWS services are billed separately, including during the trial. Idle endpoints continue to incur charges. Check the active offer before subscribing and delete test resources when finished.

Walkthrough

Watch the full video or the vertical overview. The illustrative interface replays recorded AWS outputs on fictional development cases, including a Portuguese failure and an unrelated-question control. It is not an included application or an accuracy sample. English narration is AI-generated.

Next steps

  • Getting started: subscription, regional model-package ARN, endpoint and batch examples.
  • Usage: bounded request shape, result mapping and errors.
  • Limits and evaluation: measured scope and operational boundaries.
  • Support: troubleshooting and safe support details.

Upstream project

BAAI/bge-reranker-v2-m3 provides the original publicly available weights under Apache-2.0. Read the sources and license notices for the integration, dependencies and private evaluation sources.

Last updated November 15, 2026 · Suggest an improvement