To avoid a re-index, can we embed only new documents wi...

The objectives of embedding models depend on inner products, and the resulting space is only defined up to an orthogonal transform. Consequently, cross-model cosine similarity is essentially random. The old vectors will effectively disappear rather than degrade. If a full re-index is impossible, use two indexes queried separately and merged via rank fusion, or attempt to learn an alignment between the spaces, though this should be validated by checking the distribution of cosines on a sample set.