Rerank 3.5

Rerank 3.5 sits after your first-stage retriever and makes finer-grained relevance judgments by jointly encoding the query with each candidate chunk. That cross-attention lets it catch subtle semantics—synonyms, paraphrase, discourse cues—and demote near-duplicates or off-topic snippets, so your top-k actually contains what the user needs. It’s optimized for fast batch inference and stable scoring, works well with cosine or raw logits, and handles long queries and multi-sentence passages without falling back to keyword bias. Teams typically pair it with an embedding retriever for recall, then let Rerank 3.5 sort the shortlist before generation. The result is higher click-through for search, better grounding and fewer hallucinations in RAG, and cleaner inputs for agents that depend on precise evidence.