bge-reranker-v2-m3 is live on the standard /v1/rerank endpoint
BAAI’s open-source multilingual reranker is now available in the default and svip groups at $0.01 per 1M tokens, billed on input only. It produces no embeddings and cannot be indexed — its job is to sit behind vector recall and re-sort what came back. In testing on the same 10 candidates, changing nothing but the ranking method lifted nDCG@3 from 0.53 to 1.00, pushing out the distractors that share a topic but never answer the question. Ordering was correct across Chinese, English, Japanese, Korean, Russian, French and Arabic, plus cross-lingual queries.
Recommended shape: recall 50–100 candidates with vectors or BM25 → rerank down to the top 3–5 → feed the LLM. Three pitfalls to avoid: relevance_score is not a confidence value comparable across queries (don’t set a fixed threshold), negation is a clear weakness, and long documents need chunking to 200–500 characters first.
An upstream TPM quota increase is being applied for; contact APIYI support to review higher volumes. Full measurements and tuning guidance are in the rerank overview and RAG tuning in practice.
← Back to Live Updates · 📚 Monthly archive