VoyageAI by MongoDB: voyage-4-lite

voyage-4-lite is an embedding model from Voyage AI, the company behind the Voyage family of retrieval models, now part of MongoDB. It sits in the embeddings category rather than the chat category, which means it does not generate text. Instead, it turns text into vectors that you store and compare. It is positioned as a lightweight, general-purpose model, built for low latency and cost rather than maximum accuracy on the hardest retrieval problems.

The model accepts text as input and returns embeddings as output. Its context length is 32,000 tokens, and the maximum output is 28,800, so long documents can be embedded without splitting them into many small chunks. Voyage AI trained it with Matryoshka learning and quantization-aware training, which is what allows a single model to produce embeddings at 2048, 1024, 512, and 256 dimensions. You choose the dimension that fits your storage and latency budget, and shorter vectors remain usable rather than collapsing.

Those properties point to a fairly specific set of uses. Semantic search over a document collection, where you embed chunks once and query them with an embedded question, is the core case. The same vectors support retrieval augmented generation, deduplication and near-duplicate detection, clustering and topic grouping, and classification or routing when you compare an input against labelled examples. Because the model handles 32,000 tokens of context, you can embed whole pages or long passages directly, which suits code search, support ticket triage, and internal knowledge bases. The lower dimension options matter when you are storing millions of vectors and want to keep the index small.

You call it through api14's OpenAI-compatible API using the model id "voyageai/voyage-4-lite". Because the endpoint follows the OpenAI embeddings shape, existing clients and libraries work with only the base URL and model id changed. The model exposes no additional supported parameters, so requests rely on the standard fields the embeddings endpoint accepts, such as the input text and the model id. There is no sampling, temperature, or tool configuration to set, since none of that applies to an embedding model.

If your task is text generation, summarisation, or reasoning, this is the wrong model and a chat model will fit better. voyage-4-lite is also described as lightweight and general purpose, so if you need the strongest retrieval quality on specialised or highly technical corpora, a larger Voyage embedding model is the more likely choice.