We continue to improve the Graft platform and deal with bugs, the highlights of which are listed below. As always if you spot an issue please contact Graft Support
New Features
Updated Enrichment Page
- New entry control flow:
- Cannot enter without ingest and embedding entity first
- Symmetric flows for pre-trained and user-defined enrichments
- New initial experience for user-defined enrichments:
- Graft will train a user-defined enrichment as soon as it’s created, stats for that model will be available on first visit to enrichment details page
- New control flow for pre-trained and user-defined enrichment details:
- Overview card shows training statistics and links to sql tab for viewing predictions
- Flywheel card guides user through (1) making predictions, (2) adding labels/correcting predictions, (3) retraining enrichment and checking changes to accuracy, recall, precision.
- User can now choose to save or discard the model which is obtained via retraining.
- For pre-trained enrichments, Flywheel card prevents the user from doing anything other than generating predictions
- Flywheel card will prevent user from retraining or generating predictions when doing so would only spend money (and not produce any observable changes)
- Summary tab now shows events for both pre-trained and user-defined enrichments
- New UI for adding labels (either via Enrichment creation or Flywheel card)
- User can choose whether labels are part of entity or whether the entity needs to be joined with another data source
- Both control paths provide a preview to help user gain confidence that they’re making the right selections
- Active Learning
- Active learning is now found under Correct Predictions tab on enrichment detail page
- Various UI tweaks remove artifacting on pop-ups
- Enrichment Detail control flow removes most of the lock-outs that used to appear over the Correct Predictions UI
- Update default embedding strategy not to truncate long text
Improvements
- Add SSL support for MongoDB connector
- Retrain enrichments from scratch using cumulative data
- Improved error handling Support all embedding strategies when working with extracted PDF text
Bug fixes
- Reinstate Reddit connector after API updates