topicmodel lets you discover what topics are covered in a bunch of documents. You can also classify documents into topics and find the similarity of each document with each topic.

To categorize each line in docs.txt into topics, run:

export OPENAI_API_KEY=...
uvx topicmodel docs.txt --output topicmodel.txt

For example, if docs.txt has:

Mars has a thin atmosphere.
The moon orbits Earth.
Stars shine at night.
Bread needs yeast.
Basil smells fresh.

Run:

uvx topicmodel docs.txt --ntopics=2

It groups each line into 2 auto-discovered topics and print something like:

1: Space and Astronomy 2: Food and Ingredients

The best_match column is the closest topic to the text. The rest of the columns are the similarity between the text and each topic.

Create this topics.txt:

Astronomy
Cooking

Run:

uvx topicmodel docs.txt --topics topics.txt

This groups each line into the 2 topics in topics.txt along with the similarities:

You can visualize how documents cluster into topics using the --plot option. Extending the example from Discover Topics, run:

uvx topicmodel docs.txt --ntopics=2 --plot=plot.svg

This creates a plot.svg file showing a 2D visualization of your documents using UMAP dimensionality reduction. Each point represents a document, and regions are colored by their topic. Documents closer together are more similar, making it easy to see how well topics separate and which documents might be borderline cases.

• --docs: File containing documents. Required. Can be .txt, .csv or .json file or a JSON string
  • .txt: Each line is treated as a document.
  • .csv: Each row is treated as a document. Only the first column is used.
  • .json: This should have an array of objects. Only the first key is used. Example: [{"text": "Apples are great"}, {"text": "Bananas are yellow"}]
  • JSON string: You can pass the the JSON directly as input. Example: uvx topicmodel '[{"text": "Apple"}, {"text": "Banana"}]' --ntopics 2
• --topics: Optional file with existing topics you want to match with. Can be .txt, .csv or .json
• --output: Path to save results. Can be .csv, .json or .txt.
• --model: Default: text-embedding-3-small. OpenAI embedding model. Use text-embedding-3-large for higher quality.
• --name_model: Default: gpt-4.1-mini. Model to name clusters.
• --ntopics: Default: 20. Approx. number of topics to auto-discover. Increase for more granular clusters.
• --nsamples: Default: 5. Documents to show the naming model from each cluster. Higher values may improve topic names but increase cost.
• --truncate: Default: 200. Characters from each document to send to the naming model. Adjust based on document length; shorter saves tokens.
• --hierarchy: Optional. Generate hierarchical topic names (default "2 level depth").
• --plot: Optional. Save UMAP cluster visualization as a svg file (.svg).
• --prompt: Prompt sent to the naming model. Modify to control naming style.

The default --prompt is:

Here are clusters of documents. Suggest 2-4 word topic names for each cluster. Capture the spirit of each cluster. Differentiate from other clusters.

Environment variables:

# Use a different OpenAI compatible provider, e.g. openrouter:
export OPENAI_BASE_URL=https://openrouter.ai/api/v1

# Embeddings are cached in this path. You can change it. The default is:
export TOPICMODEL_CACHE=~/.cache/topicmodel/embeddings.db

git clone https://github.com/gramener/topicmodel.git
cd topicmodel
uvx ruff --line-length 100 .
uvx --with pytest-asyncio,httpx,pandas,numpy,scikit-learn,tiktoken,tqdm pytest

Modify the pyproject.toml file to change the version number.

uv build
uv publish

This is deployed to pypi as Anand.S

• 0.1.2: 07 Aug 2025. Include best_score in output
• 0.1.1: 07 Aug 2025. Help shows defaults. Informative errors. More tests
• 0.1.0: 25 Jul 2025. Initial release

MIT