v9.1.0

9.1.0

Features

• New Pharia Kernel connector (KernelTask) for calling Skills from a Task
• Add HybridQdrantInMemoryRetriever enabling hybrid search for in-memory Qdrant collections

Fixes

• Add warning to PromptBasedClassify and PromptBasedClassifyWithDefinitions to be cautious when using them with model families other than luminous

Full changelog: v9.0.2...v9.1.0

v9.0.2

9.0.2

Features

• Update dependencies

Full changelog: v9.0.1...v9.0.2

v9.0.1

9.0.1

Fixes

• Fixes an incompatibility where models with tokenizer with no whitespace prefix could not be used for qa examples. Now, no error will be thrown.

Full changelog: v9.0.0...v9.0.1

v9.0.0

9.0.0

Features

• Introduce Benchmark and StudioBenchmark
  • Benchmark allows you to evaluate and compare the performance of different Tasks with a fixed evaluation logic, aggregation logic and Dataset.
  • Add how_to_execute_a_benchmark.ipynb to how-tos
  • Add studio.ipynb to notebooks to show how one can debug a Task with Studio
• Introduce BenchmarkRepositoryand StudioBenchmarkRepository
• Add create_project bool to StudioClient.__init__() to enable users to automatically create their Studio projects
• Add progressbar to the Runner to be able to track the Run
• Add StudioClient.submit_benchmark_lineages function and include it in StudioClient.submit_benchmark_execution

DocumentIndexClient

• Add method DocumentIndexClient.chunks() for retrieving all text chunks of a document.
• Add metadata filter FilterOps.IS_NULL, that allows to filter fields based on whether their value is null.

Fixes

• The Document Index SearchQuery now correctly allows searches with a negative min_score.

Deprecations

...

Breaking Changes

• The env variable POSTGRES_HOST is split into POSTGRES_HOST and POSTGRES_PORT. This affects all classes interacting with Studio and the InstructionFinetuningDataRepository.
• The following env variables now need to be set (previously pointed to defaults)
  • CLIENT_URL - URL of your inference stack
  • DOCUMENT_INDEX_URL - URL of the document index

Full changelog: v8.0.0...v9.0.0

v8.0.0

v8.0.0

Features

• You can now customise the embedding model when creating an index using the DocumentIndexClient.
• You can now use the InstructableEmbed embedding strategy when creating an index using the DocumentIndexClient. See the document_index.ipynb notebook for more information and an example.

Breaking Changes

• The way you configure indexes in the DocumentIndexClient has changed. See the document_index.ipynb notebook for more information.
• The EmbeddingType alias has been renamed to Representation to better align with the underlying API.
• The embedding_type field has been removed from the IndexConfiguration class. You now configure embedding-related parameters via the embedding field.
• You now always need to specify an embedding model when creating an index. Previously, this was always luminous-base.

Full changelog: v7.3.1...v8.0.0

v7.3.1

7.3.1

Features

• Update dependencies

Full changelog: v0.7.3.0...v7.3.1

v7.3.0

7.3.0

(Note: The git tag of this version is incorrect (0.7.3.0). However, the version number in the pyproject.toml is set correctly to 7.3.0)

Features

• Add support for Llama3InstructModel in PromptBasedClassify
• Add TextControl to 'to_instruct_prompt' for instruct models
  • Add 'attention_manipulation_with_text_controls.ipynb' to tutorial notebooks
• Introduced InstructionFinetuningDataHandler to provide methods for storing, retrieving and updating finetuning data samples given an InstructionFinetuningDataRepository. Also has methods for filtered sample retrieval and for dataset formatting.
• Introduced InstructionFinetuningDataRepository for storing and retrieving finetuning samples. Comes in two implementations:
  • PostgresInstructionFinetuningDataRepository to work with data stored in a Postgres database.
  • FileInstructionFinetuningDataRepository to work with data stored in the local file-system.
• Compute precision, recall and f1-score by class in SingleLabelClassifyAggregationLogic
• Add submit_dataset function to StudioClient
  • Add how_to_upload_existing_datasets_to_studio.ipynb to how-tos

Fixes

• Improved some docstring inconsistencies across the codebase and switched the docstring checker to pydoclint.

Full changelog: v7.2.0...v7.3.0

v7.2.0

7.2.0

Features

• Add support for stages and files in Data client.

Fixes

• Update names of pharia-1 models to lowercase, aligning with fresh deployments of the api-scheduler.

Full Changelog: v7.1.0...v7.2.0

v7.1.0

7.1.0

Features

• Add Catalan and Polish support to DetectLanguage.
• Add utility function run_is_already_computed to Runner to check if a run with the given metadata has already been computed.
  • The parameter_optimization notebook describes how to use the run_is_already_computed function.

Fixes

• The default max_retry_time for the LimitedConcurrencyClient is now set to 3 minutes from a day. If you have long-running evaluations that need this, you can re-set a long retry time in the constructor.

Full Changelog: v7.0.0...v7.1.0

v7.0.0

7.0.0

Features

• You can now specify a hybrid_index when creating an index for the document index to use hybrid (semantic and keyword) search.
• min_score and max_results are now optional parameters in DocumentIndexClient.SearchQuery.
• k is now an optional parameter in DocumentIndexRetriever.
• List all indexes of a namespace with DocumentIndexClient.list_indexes.
• Remove an index from a namespace with DocumentIndexClient.delete_index.
• ChatModel now inherits from ControlModel. Although we recommend to use the new chat interface, you can use the Pharia1ChatModel with tasks that rely on ControlModel now.

Fixes

• DocumentIndexClient now properly sets chunk_overlap when creating an index configuration.

Breaking Changes

• The default model for Llama3InstructModel is now llama-3.1-8b-instruct instead of llama-3-8b-instruct. We also removed the llama3.0 models from the recommended models of the Llama3InstructModel.

• The default value of threshold in the DocumentIndexRetriever has changed from 0.5 to 0.0. This accommodates fusion scoring for searches over hybrid indexes.

  Full Changelog: v6.0.0...v7.0.0