Zero-shot topic modeling with nr_topics parameter results in empty topic_sizes_ and get_topic_info()

Have you searched existing issues? 🔎

• I have searched and found no existing issues

Desribe the bug

When using zero-shot topic modeling with the nr_topics parameter set, and all (or most) documents are assigned to zero-shot topics, the topic_sizes_ attribute remains empty and get_topic_info() returns an empty DataFrame.

Root cause:
The issue occurs in the fit_transform() method when all documents are assigned to zero-shot topics. The code has this conditional:

if len(documents) > 0:
    # Cluster reduced embeddings
    documents, probabilities = self._cluster_embeddings(umap_embeddings, documents, y=y)
    if self._is_zeroshot() and len(assigned_documents) > 0:
        documents, embeddings = self._combine_zeroshot_topics(
            documents, embeddings, assigned_documents, assigned_embeddings
        )
else:
    # All documents matches zero-shot topics
    documents = assigned_documents
    embeddings = assigned_embeddings

When all documents are assigned to zero-shot topics, len(documents) == 0, so the else branch is taken. However, this branch never calls _update_topic_size(), which is normally called within _cluster_embeddings() or _combine_zeroshot_topics().

Additionally, when nr_topics is specified, the fallback call to _sort_mappings_by_frequency() (which would call _update_topic_size()) is skipped because if not self.nr_topics: evaluates to False.

Expected behavior:

• topic_sizes_ should contain the count of documents per topic
• get_topic_info() should return a populated DataFrame with topic information

Actual behavior:

• topic_sizes_ is an empty dictionary {}
• get_topic_info() returns an empty DataFrame
• topic_representations_ works correctly (populated as expected)

Note: The issue does NOT occur when nr_topics is not specified, because _sort_mappings_by_frequency() gets called, which internally calls _update_topic_size().

Reproduction

from bertopic import BERTopic

# Sample documents and zero-shot topics
docs = ["I need help with my voucher", "Gift card not working", "Customer service was poor"] * 50
zeroshot_topics = ["Voucher inquiries", "Gift card issues", "Customer service feedback"]

# BUG: Setting nr_topics causes the issue
model = BERTopic(
    zeroshot_topic_list=zeroshot_topics,
    zeroshot_min_similarity=-1,  # Force all documents to zero-shot assignment
    nr_topics=4,  # This triggers the bug
)

topics, _ = model.fit_transform(docs)

# Demonstrates the bug
print("topic_sizes_:", model.topic_sizes_)  # Expected: Counter with topic counts, Actual: None
print("topic_representations_ works:", bool(model.topic_representations_))  # This works: True

BERTopic Version

0.17.0

Start in fit_transform() and trace the all-zero-shot else branch alongside _cluster_embeddings(), combine_zeroshot_topics(), and sort_mappings_by_frequency(). Use the supplied zero-shot reproduction with nr_topics=4, then verify that topic_sizes contains document counts and get_topic_info() returns a populated DataFrame while topic_representations remains populated.