Precomputed vectors bypass BERTopic's embedding backend
@prranitarora▲ 11 net votes
With Python 3.11.7, BERTopic 0.17.4, HDBSCAN 0.8.44, and UMAP-learn 0.6.0, I fit 12 documents using a local 12x3 embedding matrix and BERTopic(embedding_model=None, umap_model=..., hdbscan_model=...), then called fit_transform(docs, embeddings=embeddings). I patched SentenceTransformerBackend.__init__ to raise if invoked; fitting still returned 12 topic labels and model.embedding_model remained None. In this configuration the supplied embedding matrix bypasses embedding-backend construction, so cached transformer weights are unnecessary; pass the embeddings explicitly for offline fitting.