Skip to content

OrcaSDK Release Notes#

This document tracks notable changes to the OrcaSDK.

v0.1.13#

  • Added methods to retrieve the currently used org ID, scopes, API key name, and config.
  • Added progress bars for batch operations in predict, insert, update, and delete methods.
  • Added a "replace" option to the if_exists parameter in model creation methods to allow replacing existing models.
  • Enhanced explanation output to include input and prediction information.
  • Refactored memory suggestion API to return a list of LabeledMemorySuggestion objects that can be directly inserted into a memoryset.
  • Replaced head_type with balance_classes parameter in classification model create method.
  • Fixed prediction explanation timeout issue.
  • Aligned and fixed __repr__ methods across the SDK.
  • Fixed predictions not storing the correct memoryset when memoryset override is used.
  • Fixed API health check in OrcaSDK to properly validate responses and fail faster.
  • Removed default file type from datasource download.
  • Added ability for root users to upload locally finetuned embedding models.

v0.1.12#

  • Added a partitioned property on memorysets to check whether a memoryset uses partitioning.
  • Added a partitioned parameter to memoryset create to create empty partitioned memorysets.
  • Added ability to change whether a memoryset is partitioned during clone via the partitioned parameter.
  • Added a static compute method to classification and regression metrics to calculate metrics from a list of predictions.
  • Added shared logger to be used throughout the SDK that can be customized.
  • Added a consistency parameter to get, query, search, and predict methods.
  • Added support for model evaluate with pandas data frames and generic iterables of dictionaries.
  • Fixed a bug where expected labels/scores were not being saved on predictions when telemetry was disabled.
  • Removed scikit-learn and numpy dependencies (metrics calculation now happens on the API server).
  • Removed datasets dependency (was only needed for types and torch data parsing which was refactored to not need it anymore).

v0.1.11#

  • Added a cascade parameter to drop method on memorysets and finetuned embedding models to allow deleting related resources in one call (avoids foreign-key errors).
  • Added classification_models and regression_models properties on memorysets to list the models associated with a memoryset.
  • Added support for updating all memories that match a filter via the memoryset update method.
  • Changed the memoryset update method to return the number of updated memories (instead of the updated objects) to reduce network usage.
  • Added support for deleting all memories that match a filter via the memoryset delete method.
  • Added a truncate method on memorysets to delete all memories, or only those in a specific partition (partition_id defaults to UNSET; passing None truncates the global partition).
  • Removed partition parameters from the memoryset delete and query methods; use filter to target one or more partitions, or use truncate to clear a partition.
  • Fixed a bug where updating non-metadata fields on a memory could clear its metadata.
  • Removed torch, pandas, and pyarrow dependencies (they were only needed for typing).
  • Made gradio optional; install the notebook UI extras via orca_sdk[ui].

v0.1.10#

  • Removed deprecation warning from OrcaCredentials.set_api_key.

v0.1.9#

  • Added support for Python 3.14 including updating datasets to 4.4.2, pyarrow to 22.0.0, gradio to 6.3.0, and fixing several incompatibility issues.
  • Changed predictions to return all predictions by default when limit is None.
  • Changed predict and apredict to automatically batch requests to reduce network overhead.
  • Fixed evaluate to also include the confusion matrix, when running evaluate with a local dataset.

v0.1.7#

  • Added confusion matrix to classification metrics
  • Added ability to create empty memorysets
  • Added stricter checks for if_exists="open" during memoryset creation
  • Added support for running distribution, duplicate, cluster, and projection analyses on ScoredMemoryset
  • Tweaked representation of predictive models and embedding models
  • Fixed classification metrics calculation when test set classes don’t match model’s predicted classes.

v0.1.6#

  • Fixed bug that could lead to division by zero during metrics calculation

v0.1.5#

  • Added support for partitioned memorysets and models

v0.1.4#

  • Added use_gpu parameter to prediction methods to allow CPU-based predictions
  • Added support for using string columns as label columns
  • Added sample parameter to memoryset creation methods and model evaluate methods to allow sampling of rows
  • Added ignore_unlabeled parameter to prediction and evaluate methods
  • Added method to query datasource rows
  • Added support to finetune embedding models for regression tasks
  • Added support for querying prediction telemetry on memories
  • Updated SDK to use new job endpoints
  • Improved prediction caching
  • Fixed dependency vulnerability

v0.1.3#

  • Added async ClassificationModel.apredict and Memoryset.ainsert methods.
  • Added batching to Memoryset.insert, Memoryset.update, and Memoryset.delete methods to reduce network issues.
  • Renamed "neighbor" analysis to "distribution" analysis.
  • Allowed injecting custom httpx clients via context to cleanly override api keys and controll client lifecycle.
  • Fixed creation of orphaned datasources when using the if_exists="open" option during memoryset creation.
  • Removed deprecated Memoryset.run_embedding_evaluation method, use EmbeddingModel.evaluate instead.

v0.1.2#

  • Added support for None labels and scores to memorysets and models.
  • Added automatic retrying of requests to improve mitigate transient network and service issues.
  • Fixed bug when receiving additional field in API responses for metrics
  • Updated dependencies to resolve vulnerabilities