MongoDB's new Voyage AI boosts retrieval accuracy up to 30% with native reranking

MongoDB announced at its Bengaluru event new Voyage AI capabilities designed to solve two major enterprise AI hurdles: retrieval accuracy and compliance. The centerpiece is Native Reranking (public preview), powered by Voyage AI models that outperform Google and Cohere on public retrieval embedding benchmarks. It delivers up to a 30% boost in retrieval quality directly inside the database, with no external APIs, keys, or round-trips. Alongside this, MongoDB introduced voyage-context-4 (GA), a new embedding model built for long documents that processes full context rather than isolated chunks, preserving meaning across complex enterprise content. Hybrid Search (GA) combines full-text and vector search in a single query, keeping embeddings automatically updated from live data.

These capabilities are now available for MongoDB Enterprise Advanced and Community Edition, bringing Atlas-grade retrieval to on-premises, private cloud, and local environments. This is critical for regulated industries with data residency mandates. Early adopter Emergent Labs, an AI-native app platform running 2 million applications, reports that MongoDB's unified retrieval stack prevents stale or wrong data from compounding errors in agent-based workflows. MongoDB's Chief Product Officer Ben Cefalo emphasized that bolting on separate systems creates more vendors, latency, and failure points. By embedding accurate retrieval directly into the operational database, MongoDB gives enterprises a production-ready stack that works wherever their data lives.