// the find
RasaHQ/rasa
💬 Open source machine learning framework to automate text- and voice-based conversations: NLU, dialogue management, connect to Slack, Facebook, and more - Create chatbots and voice assistants
Rasa is a Python framework for building task-oriented chatbots with NLU (intent classification, entity extraction) and dialogue management. It's explicitly in maintenance mode — the company has moved on to a cloud product called Hello Rasa/CALM. You're looking at a framework that won't see new features, only security patches at best.
The intent/entity/story architecture is well-thought-out and handles multi-turn conversations without needing an LLM at inference time — useful if you need predictable, auditable dialogue flows. The component pipeline (featurizers, classifiers, policies) is genuinely modular; you can swap in spaCy, BERT, or MITIE embeddings without rewriting logic. CI tooling is thorough — model regression tests on a schedule, not just unit tests, which is rare in ML projects. The graph-based execution model added in later versions (visible in data/graph_schemas/) makes the training pipeline inspectable and cacheable.
It is in maintenance mode, full stop. The README says so explicitly and the company is actively pushing users to a paid cloud product. Adopting this for a new project means betting on community forks or accepting you'll eventually need to migrate. The intent-based paradigm it's built on is now the wrong abstraction — LLMs handle 'what did the user mean' better than training data, which is exactly why Rasa abandoned this framework internally. Dependency footprint is heavy: Poetry-managed, needs specific Python versions, and the NLU pipeline pulls in TensorFlow or PyTorch depending on configuration — painful to containerize and keep updated. No path to migrate a Rasa Open Source bot to the new CALM system without a near-complete rewrite.