Topically is a work-in-progress suite of tools designed to help users make sense of text collections such as messages, articles, emails, and news headlines by leveraging large language models from Cohere. The repository, maintained by jalammar with a maintenance commitment through April 2023, addresses a specific challenge in natural language processing: automatically generating meaningful names for clusters of short texts based on their semantic content.
The primary feature of Topically is its ability to name clusters of short texts without requiring users to manually assign labels. The tool works particularly well when integrated with existing topic modeling libraries like BERTopic. In a typical workflow, BERTopic handles the clustering of texts into groups, and Topically then generates descriptive names for those clusters. This approach improves upon traditional keyword-based topic labels by using generative language models to produce more contextually appropriate and human-readable cluster names. The repository includes example code and a Google Colab notebook demonstrating integration with BERTopic, making it accessible for users to experiment with the workflow.
Topically's architecture is intentionally simple and modular. The codebase consists of two main classes: the Topically class, which maintains the client connection to Cohere's platform and exposes the primary interaction point through the name_topics method, and the ClusterNamer class, which handles prompt preparation and calls Cohere's Generate endpoint to produce suggested topic names. This straightforward design reflects the project's early-stage status while providing a clear foundation for future development.
The tool operates by sending requests to Cohere's managed language models through their API. Users need to obtain an API key from Cohere's dashboard, which can be used for free during the prototyping phase. Topically includes bundled prompts for generating titles, but the documentation emphasizes that users can achieve better results by customizing prompts to their specific use cases and providing examples of good cluster names from their own data. The tool is optimized for short texts due to context length limitations of GPT models, though users working with longer documents can experiment with excerpts or summaries.
Installation is straightforward through pip, with a basic installation command and an optional installation variant that includes BERTopic as a dependency. The project is licensed under MIT, making it freely available for both commercial and personal use. The repository maintains contributor guidelines and requires a Contributor License Agreement before accepting pull requests, reflecting Cohere's standard open-source practices. Support is available through GitHub issues or the Cohere Discord community, providing multiple channels for users to ask questions and report problems.
