In practice, this means a RAG pipeline built on Voyage 4 will retrieve the correct piece of information from a 10,000-page legal contract or a month-long customer support chat with over 94% accuracy, compared to ~70% for generic embedding models.
Use voyage-4-lite for low-latency user queries without needing to rebuild the index. voyage 4
For a mobile game, the scale is staggering. The map replicates the M7 Highway, stretching from the capital, Moscow, all the way to the Volga region and beyond. The distance isn't scaled down for convenience; driving 1,000 kilometers in the game feels like a genuine commitment. This commitment forces the player to respect the road. You cannot simply hold the accelerate button; you must scan the horizon for oncoming trucks, watch for potholes that could destroy your suspension, and plan your stops for gas. In practice, this means a RAG pipeline built
Most embedding models truncate at 8k or 16k tokens. Voyage 4 embraces long-form retrieval. You can embed an entire book chapter as a single vector. This is revolutionary for applications like: The map replicates the M7 Highway, stretching from
The standout feature of is the introduction of shared embedding spaces . Historically, switching between a "large" model for accuracy and a "lite" model for speed required re-embedding your entire dataset—a process that is both time-consuming and expensive.
: Supports flexible embedding dimensions (256, 512, 1024, or 2048) with minimal loss in retrieval quality, allowing users to optimize for storage and speed.