Release Notes - 2026-03-06

Date: March 6, 2026

Improvements

Configure AI provider API keys: You can now set your own API keys for OpenAI, Gemini, and Anthropic directly from Settings > Secrets.
Evaluator preprocessing script: Define a custom script in an evaluator to control how the AI response is extracted before evaluation — useful when the response is embedded in complex output structures.
Promotion run history: A list of the last 10 promotion runs is now available per tenant, including status, start date, and end date.
Manual reporting improvements: The chat ID is now optional when creating manual reports. Reports can also be submitted directly from the Troubleshooting section.
Missing API key warning: The portal now shows a visible warning when no AI provider API key is configured, making it clear why the chat may not work.
Feature elements API: The API now supports updating and deleting individual elements within a feature (PATCH and DELETE on /features/{id}/elements).
Metadata config types are now immutable: Once a metadata config is created, its type cannot be changed, ensuring data consistency.

Safe node deletion: Deleting a node used in an ingestion config now automatically updates the config and shows a warning explaining the change.
Safe metadata deletion: Deleting a metadata config linked to ingestion configs now updates those configs and notifies you of the impact.

Fixed lessons with very large content being saved incompletely.
Fixed an error when switching a metadata config to free values.
Fixed a crash in the evaluator UI when using incompatible metric and chart type combinations.
Fixed prediction errors when using a preprocessing script in the evaluator.
Fixed a silent failure when deleting an analyzer in Running state — the portal now prevents the action with a clear message.
Fixed past evaluations showing as "Unknown" when their associated metric was deleted.
Fixed the role selector showing invalid tenants in user admin.

Gemini and Anthropic genconfig update required: Existing Anthropic and Gemini generator configurations must be updated to use the new OpenAI-compatible SDK format. All models previous to Claude 4.5 and Gemini 3 Flash will no longer be available.