Re-embedding millions of documents is expensive. Can I map my old 768-dimensional vectors to the new model's 1024 dimensions?
Why there's no shortcut
Each model learns its own embedding space. Dimension 17 in one model has no relationship to dimension 17 in another, and the spaces may be rotated, stretched, and organized differently. Padding with zeros or truncating makes the error go away, but the resulting similarity scores are meaningless.
Answer 1: Re-embed all documents with the new model (the correct approach)
This is the only approach that gives reliable results. Plan for it like a data migration:
new_model = SentenceTransformer("new-model-name")
new_vecs = new_model.encode(documents, batch_size=64,
show_progress_bar=True,
normalize_embeddings=True)
np.save("embeddings_new_model.npy", new_vecs)
Saving the embeddings to disk means you won't have to recompute them if you need to rebuild the index.
