Disclaimer: niche use case and not covered by the current tutorial. Perhaps a short footnote would be enough to make users aware of this?
I'm not absolutely certain, but it seems that using precomputed embeddings without providing an embedding_model can limit access to some BERTopic features that need to embed new text.
Examples:
- transform(new_docs) -> needs to embed the new docs
- find_topics("query") -> needs to embed the query ?
- zero-shot topic modeling (zeroshot_topic_list) -> needs to embed the given labels ?
- some representation models ? KeyBERTInspired / MaximalMarginalRelevance ?
Fairly niche and probably more advanced use case, but perhaps a brief footnote mentioning these limitations could help beginners avoid stumbling into these edge cases?
Disclaimer: niche use case and not covered by the current tutorial. Perhaps a short footnote would be enough to make users aware of this?
I'm not absolutely certain, but it seems that using precomputed embeddings without providing an embedding_model can limit access to some BERTopic features that need to embed new text.
Examples:
Fairly niche and probably more advanced use case, but perhaps a brief footnote mentioning these limitations could help beginners avoid stumbling into these edge cases?