Voyage 4 puts every model size in one embedding space
- Source
- x.com
- Date
📢 ICYMI: We recently released the Voyage 4 model family. Voyage 4 features an industry-first shared embedding space, eliminating the need to re-index your data when switching between models in the series. This breakthrough is designed to solve the trade-off between frontier-level accuracy and production efficiency.
The shared embedding space enables asymmetric retrieval, for example, embedding documents with voyage-4-large for high accuracy while embedding queries with a smaller model to keep per-query latency and cost low. Asymmetric retrieval is denoted in the figure with a *
voyage-4-large sets a new accuracy standard, outperforming voyage-4, voyage-4-lite, Gemini Embedding 001, Cohere Embed v4, and OpenAI v3 Large by an average of 1.87%, 4.80%, 3.87%, 8.20%, and 14.05%, respectively.
The first 200 million tokens are free. Learn more in our release blog: https://t.co/6HeupDGb6t
Context
Voyage AI designed Voyage 4 so different-sized models produce compatible , the numerical representations used to match queries with documents. That allows a search system to process its document collection with voyage-4-large while using a smaller model for incoming questions. The family's release article calls this asymmetric retrieval.
The useful flexibility is on the query side: a team can change which Voyage 4 model handles questions without regenerating all the stored document embeddings merely to make them compatible. It separates the model choice for indexing documents from the cost and latency of serving repeated searches.
- embedding — A list of numbers representing a piece of text's meaning, so that similar meanings end up numerically close and can be searched.
Checking sign-in…
Loading comments…