Voyage 4 documents compatible embedding spaces across large, standard, lite, and nano models. This permits testing larger-model document vectors with cheaper query encoding while retaining Qdrant; it does not establish equal accuracy or cross-family compatibility.

Asymmetric embeddings as a component-level experiment

The Voyage 4 announcement documents shared-space compatibility among the four general-purpose models. The current model catalog retains that capability.

My proposed experiment is to encode document chunks with voyage-4-large and compare query encoding with voyage-4-lite versus the larger model, holding vector dimensions and index configuration fixed. The tradeoff concerns recurring query cost and latency versus retrieval accuracy; vendor benchmarks do not establish the outcome on this corpus.

This is an inference API substitution, not migration of storage, publishing, or retrieval to a provider platform. Do not infer compatibility with Qwen, old Voyage generations, or the separate contextualized embedding model. Other model migrations should be treated as new representations requiring backfilling and evaluation.