What it is: Native Reranking enables developers to improve retrieval accuracy using Voyage AI reranking models directly from the MongoDB aggregation pipeline. A single new stage, $rerank, can improve retrieval accuracy by an average of 23.84% over full-text search and 10.82% over vector search when using Voyage rerank-2.5 compared to non-reranked results.Who it's for: Atlas developers building RAG applications, search and recommendation engines, or AI agents.Why it matters: Previously, adding a reranker model to your retrieval pipeline meant calling a separate API outside your query layer, handling reordering in application code, and managing additional credentials and billing. Native Reranking eliminates that complexity: $rerank is now a single stage in the same pipeline you're already using in MongoDB.How to get started: Native Reranking is now available in Atlas for clusters running Latest Version with Auto-Upgrades (MongoDB 8.3). Learn more in the documentation.

What it is: voyage-context-4 is the next-generation contextualized chunk embedding model and a drop-in replacement for voyage-context-3, producing one vector per chunk that captures full document context -- now with a new mixture-of-experts backbone, built-in auto-chunking, transparent handling of documents beyond the 32K-token window, and native overlapping-chunk support. It outperforms voyage-context-3 by 1.40% (document-level) and 2.08% (chunk-level) across 39 datasets and is priced at $0.12 per 1M tokens, down from $0.18.Who it's for: This feature is for developers and retrieval teams building semantic search, RAG, and agentic applications, especially with long documents and chunk-level retrieval. It's ideal for customers who want maximal retrieval accuracy without manually tuning their embedding pipeline.Why it matters: By capturing full document context in every chunk embedding, voyage-context-4 improves retrieval quality across nearly every domain while removing chunking as a design concern, with no extra LLM calls or preprocessing logic. As a drop-in replacement priced below voyage-context-3, it raises accuracy and lowers cost at the same time.How to get started: voyage-context-4 is available now through the Voyage AI API and the MongoDB Atlas Embedding and Reranking API -- simply swap in the model name voyage-context-4 or pass a full document with enable_auto_chunking=True. New users get 200 million free tokens.

What it is: LangGraph.js now supports MongoDB as the backend for long-term agent memory, adding to the short-term memory (Checkpointers) already available. The MongoDB Memory Store keeps and retrieves cross-session data, with support for semantic memory search powered by either an client-side embeddings provider or MongoDB Atlas Automated Embeddings, which generates and indexes vector embeddings server-side via Voyage AI models.Who it's for: JavaScript and TypeScript developers building LangGraph agents and want a unified database for conversation history, long-term memory, and semantic search.Why it matters: MongoDB is now a first-class option at every layer of LangGraph.js memory. Teams already running on MongoDB can now keep agent memory in the same database as their operational data: no additional infrastructure for conversation state, long-term storage, or vector search. Semantic memory search lets agents retrieve memories based on meaning, surfacing past context that matches the current conversation. Automated Embeddings removes the last piece of friction: instead of provisioning and calling a separate embedding service, MongoDB handles vectorization server-side, keeping application code focused on agent logic.How to get started: See LangGraph.js Memory documentation for step-by-step examples covering MongoDB checkpointers, stores, and semantic search or head directly to our official tutorial.