Voyage AI by MongoDB

Building Gen AI prototypes is straightforward. Scaling them to handle millions of production queries? That's where things get challenging. In production, you need to optimize for three things simultaneously: accuracy, speed, and operating costs. Small inefficiencies in retrieval compound quickly, since retrieval happens on every query, often multiple times in agentic workflows. Our latest guide covers three powerful techniques for optimizing embedding-based retrieval: ⚖️ Asymmetric Retrieval with Voyage 4: Use different embedding models for queries versus documents. Voyage 4's shared embedding space lets you embed queries with voyage-4-lite (fast and economical) while searching documents embedded with voyage-4-large (high quality). This approach is particularly effective for high-volume applications where query embedding costs become significant. 🗜️ Vector Quantization: Compress embeddings from high-precision formats (float32) to low-precision formats (int8 or binary). This massively reduces memory footprint and accelerates search with minimal accuracy loss. 🪆 Matryoshka Representation Learning: Train embeddings such that early dimensions capture the most important information. This allows you to truncate embeddings to reduce storage costs and speed up search while preserving semantic meaning. Voyage 4 supports 2048, 1024, 512, and 256 dimensions from a single model. In our latest guide, we show you how to evaluate each of these techniques, so you can make data-driven decisions for your specific use case. Read the full guide: https://lnkd.in/euVPX8yf

📢 Big news for Google Cloud developers! Voyage AI models have officially joined the Model Garden on Vertex AI. You can now access our frontier-level embeddings and capabilities in multiple GCP regions—including the industry-first shared embedding space from the Voyage 4 series—alongside Google’s native tools. SOTA Retrieval: Voyage models lead standard benchmarks with models optimized for accuracy and cost. Multimodal Excellence: Use voyage-multimodal-3.5 for unified search across text, images, and video. Easy Setup: Deploy directly from Vertex AI with just a few clicks. Find us in the Vertex AI Model Garden: https://bit.ly/3MmK649

🚀 Voyage AI models are on the Amazon Web Services (AWS) Marketplace! You can deploy Voyage 4 and voyage-multimodal-3.5 directly within your AWS VPC. Leverage your existing AWS credits and infrastructure to build high-performance RAG and agentic search systems. Full Data Privacy: Models run in your own account and VPC. Seamless Integration: Deploy as real-time inference endpoints via Amazon SageMaker. Unified Billing: One consolidated bill through your AWS account. Get started on AWS: https://lnkd.in/eNeSw9MF

Stop choosing between accuracy and cost. The Voyage 4 model series offers unmatched price-performance flexibility for developers. Join Apoorva Joshi on Feb 19 at 9 AM PST to learn about: - Shared Embedding Spaces: Flagship accuracy with lite-model latency. - MoE Architecture: SOTA accuracy at 40% lower cost. - Matryoshka Learning and Quantization: Maintaining accuracy while slashing query latency Register here: https://lnkd.in/gU2FR7HQ

🚀 Meet voyage-4-nano: Our first open-weight embedding model voyage-4-nano is ideal for local development and prototyping while providing an easy path to production. ✨ Shared Embedding Space: voyage-4-nano shares an embedding space as its larger siblings (voyage-4-large, voyage-4, and voyage-4-lite). This means you can use voyage-4-nano locally alongside our flagship models without re-indexing a single vector. 🔓 Open Weights: Freely available on Hugging Face under the Apache 2.0 license. Download, run, and start building. 📏 Flexible & Efficient: Supports Matryoshka Representation Learning (MRL) and multiple quantization options (int8, binary). Get started today: https://lnkd.in/egayDGHE

Video search has long relied on a workaround—indexing only transcripts and metadata instead of the actual visual content. Voyage AI's latest multimodal embedding model, voyage-multimodal-3.5, changes that. It natively embeds text, images, AND video frames in a unified vector space. This means you can now combine visual cues with semantics and keywords from transcripts to build robust video search systems. Our new tutorial shows you how to build an agentic video search system that dynamically routes between vector and hybrid search, prioritizing visual cues, semantics, and keywords as needed for each query. Learn how to process and embed videos, implement multiple search strategies, and build production-grade multimodal retrieval systems: https://lnkd.in/e9PeeF2E

📢 Announcing voyage-multimodal-3.5: a new multimodal retrieval frontier with video support voyage-multimodal-3.5 adds explicit support for video frame sequences to enable semantic search over video content with natural language queries. voyage-multimodal-3.5 is also the first production-grade video embedding model to support Matryoshka learning, offering flexible dimensionality to optimize cost and latency with minimal quality loss. Key capabilities: • Processes text, images, screenshots, PDFs, tables, figures, and videos through a single encoder, preserving contextual relationships and eliminating the modality gap • 4.65% more accurate than Google Multimodal Embedding 001 across video retrieval datasets • 4.56% more accurate than Cohere Embed v4 across visual retrieval datasets • Within 0.29% of voyage-3-large for text-only retrieval Read the full blog to learn more: https://lnkd.in/e_zxFuxN

Stop ignoring your video data—start searching it. 📽️🔍 Video content is exploding, but finding specific content in long-form videos remains a significant challenge. In this webinar, Apoorva Joshi, Senior AI Developer Advocate, will share: - How Voyage AI’s video embeddings work - How to implement hybrid search in MongoDB - How to build an end-to-end video search application - How to optimize your video search application Don't let your video archives sit idle. Unlock them with high-performance AI. Join us on January 29 at 9 AM PST. Register here: https://lnkd.in/edmJWq-Z

Building production AI means balancing accuracy, cost, and deployment flexibility. Today, we're making that easier: ✨ Voyage 4 series: Shared embedding space lets you mix models across pipeline stages. Use voyage-4-large for indexing, voyage-4-lite for queries, and voyage-4-nano locally. 🎥 voyage-multimodal-3.5: Native video retrieval and the first production-grade video embedding model to support Matryoshka embeddings for flexible dimensionality. 🌍 New availability: Deploy on MongoDB Atlas, Google Cloud, Amazon Web Services (AWS), or Microsoft Azure. The most accurate models. Available everywhere you build. Read the full announcement: https://lnkd.in/eHRhjD_P