Release notes - 3.5.0

Release date: 29/10/2025 (early morning)

Optimized training-flow

We’ve completely reworked the Oswald training-flow to make it faster, more efficient, and more reliable. Over time, training had become slower and more complex, so this overhaul focuses on streamlining the process and reducing unnecessary retraining cycles.

• How it works:

  • When you edit only scenarios or functions, Oswald now reloads only those database entries instead of retraining the entire model — significantly improving speed.

  • When updating NLP settings, entities, or intents, Oswald will no longer auto-train. It will now wait until a user manually starts a training cycle.

  • Oswald will no longer incorrectly display that it can be trained right after a publish action or a training cycle.

Note: Because of the scale of this update, bots on the acceptance environment might experience temporary downtime during the coming week.

Prompt-language

You can now define prompts per language in the LLM-hub. This allows multilingual bots to respond more accurately and naturally in each supported locale.

• How it works:

  • Each bot locale can now have its own prompt in the LLM-hub.

  • If a locale does not have a specific prompt, Oswald will automatically use the prompt of the default locale.

Improvements

• Smarter model handling: Smaller GPT-5 instances now use less reasoning power, resulting in faster response times.

• URL handling: Redirects (status 302) will no longer be flagged as bot issues, preventing unnecessary clutter in the issues page.

What this means for you

This update significantly improves Oswald’s performance and reliability, especially for teams that regularly edit scenarios and NLP configurations. Because it’s a large and fundamental change to the training flow, we strongly recommend that you thoroughly test your chatbot on the day of launch, October 29th.

Oswald will deploy this major update to the acceptance environment tomorrow, where it will remain until the Move to Production (MTP) on October 29th. If you encounter temporary downtime on the acceptance environment, this is part of the rollout process.

Of course, before and after the MTP, our engineers will also thoroughly test all production chatbots. Even so, we strongly recommend that you plan some testing capacity within your team, since each chatbot setup is unique and may behave differently after the update.

While we aim for a smooth rollout, minor bug fixes and adjustments may still occur in acceptance during this phase, meaning the environment might be temporarily unstable at times.

Your current production setup will remain unaffected until the MTP. If you have questions, feel free to contact our support team.