Should I use embeddings to cluster customer feedback?

Yes—often embeddings are a good choice for clustering customer feedback, especially when the feedback is short, messy, or varies in wording.

When embeddings help

Use them if you want to group feedback by semantic meaning, not just shared words. For example:

• “The app keeps crashing”
• “It freezes when I open it”

These are phrased differently but mean the same thing. Embeddings can place them close together in vector space, making clustering much better than keyword-only methods.

Good use cases

• Finding recurring complaint themes
• Grouping feature requests
• Reducing a large pile of survey comments into topics
• Detecting emerging issues over time

When embeddings may not be enough

They’re less ideal if:

• You need very precise, pre-defined categories
• Feedback is extremely domain-specific and embeddings don’t capture the jargon well
• You need easily explainable clusters from the start
• The dataset is tiny and simple keyword rules would suffice

Practical approach

A common pipeline is:

1. Clean text lightly
   Remove obvious noise, but don’t over-process.
2. Generate embeddings Use a strong sentence embedding model.
3. Cluster Try algorithms like:
   • HDBSCAN if you don’t know the number of clusters
   • K-means if you do know roughly how many topics there are
4. Inspect and label clusters Read representative comments from each cluster.
5. Iterate Adjust preprocessing, clustering method, or embedding model.

Tips

• Combine embeddings with metadata like product area, customer segment, or date if useful.
• Use dimensionality reduction only for visualization, not necessarily for clustering.
• Expect some clusters to be mixed; human review is usually needed.
• If you want topics that are also easy to interpret, consider BERTopic or similar topic-modeling methods built on embeddings.

Bottom line

Yes, embeddings are usually a strong default for clustering customer feedback when the goal is semantic grouping. They work especially well for open-ended text where exact wording differs but meaning is similar.

If you want, I can also recommend:

• a specific embedding model,
• the best clustering algorithm for your case, or
• a complete workflow in Python.