Ellf helps coding assistants build text-processing systems

Explosion’s October 14 introduction positions Ellf around the data work that accompanies code. In its support-ticket example, extracting fields and choosing a routing category become separate components. Agents can draft labels, while people review and correct them. The example reserves separate, human-labelled data for evaluation, rather than treating the agent’s own suggestions as the standard against which it is judged.

The product page describes a wider development cycle: prepare data, organise annotation, train models and evaluate the resulting system. Templates provide starting workflows that a team can adapt to its data and task. Annotation and training use a processing cluster operated locally or in the team’s cloud account. Ellf also supports external model APIs; the advertised control over that cluster should not be read as a promise that every possible configuration keeps every request local.

The setup guide makes the integration concrete. Its command-line client requires Python 3.10 or later. Logging in with the agent option installs skills and adds a knowledge-base connection to the chosen coding assistant. The same login command refreshes skills and should be repeated for a new virtual environment. A coding assistant is recommended, not mandatory: teams can operate through the application or command line. Other assistant and model configurations are not guaranteed to work as intended.