v0.21.0

🚨 Breaking Changes

• Renamed hugging_face_hub_prompt_driver.py to huggingface_hub_prompt_driver.py.
• Renamed hugging_face_pipeline_prompt_driver.py to huggingface_pipeline_prompt_driver.py.
• Renamed hugging_face_tokenizer.py to huggingface_tokenizer.py.
• Moved ProxycurlClient Tool to dedicated repository.

🆕 New Features

• Added HuggingFaceEmbeddingDriver.
• Added SimpleTokenizer for use with LLM providers that don't provide tokenization APIs.
• Added streaming support to Chat util. Thanks @mattma1970!
• Added Image Generation Drivers (AzureOpenAiDalleImageGenerationDriver, LeonardoImageGenerationDriver, and AmazonBedrockImageGenerationDriver).
• Added streaming support to HuggingFaceHubPromptDriver.
• Added prompt_stack and prompt fields to StartPromptEvent to allow for inspection of the prompt before sending to the LLM.
• Created Tool Template repository.

🔧 Improvements

• Added python 3.12 support.
• Updated HuggingFaceHubPromptDriver to use new InferenceClient.
• Fixed OpenSearchVectorStoreDriver.query argument order. Thanks @igor-2lemetry!
• Fixed FileManager working directory not updating when using os.cwd().
• Fixed WebLoader failing when extracting empty pages.
• Fixed SummaryConversationMemory failing to deserialize.

📖 Docs

• Updated contribution guidelines for new Tools.
• Added docs for HuggingFaceEmbeddingDriver.
• Updated docs to reflect hugging_face -> huggingface rename.
• Added dedicated page for Tokenizers.
• Added note to Prompt Drivers page advising users that if they choose to override the Prompt Driver, they should also consider overriding the Embedding Driver.
• Fixed broken import on Task Memory page.

v0.20.0

🚨 Breaking Changes

• Removed BaseTask.add_child and BaseTask.add_parent. Use Workflow.add_task(s) and Workflow.insert_task(s) instead.
• Changed Workflows to always have a single input Task and single output Task; enabling Workflow Conversation Memory.
• Split dependencies into extras to reduce base installation size. Read more about this change here.
• Renamed ToolOutputProcessor to TaskMemoryClient.
• Removed Memory Actions in favor of TaskMemoryClient. If your Tools have data that needs to go to the LLM, you must set either set off_prompt=False on the Tool, or add TaskMemoryClient with off_prompt=False.
• Changed all Structure.run to return a Structure. This resolves inconsistencies in the run return values across Agents, Pipelines and Workflows. You can access the output Task value via Structure.output_task.output.value.
• Renamed AzureOpenAiChatPromptDriver's deployment_id and api_base to azure_deployment and azure_endpoint.
• Renamed Structure.tool_memory to Structure.task_memory.
• Renamed ToolMemory to TaskMemory.
• Renamed Structure.memory to Structure.conversation_memory.
• Removed TextQueryTask.load method.

🆕 New Features

• Added seed parameter to OpenAiChatPromptDriver. Read more about this parameter here.
• Added response_format parameter to OpenAiChatPromptDriver and AzureOpenAiChatPromptDriver. Read more about this parameter here.
• Added support for gpt-4-1106-preview(gpt-4-turbo) in OpenAiChatPromptDriver.
• Added new Workflow method insert_tasks. See example usage here.
• Added Conversation Memory to Workflows.
• Added off_prompt parameter to all Tools. With the exception of TaskMemoryClient (no default provided), off_prompt defaults to True meaning that the Tool results will go into Task Memory. If you'd like the results to go directly back to the LLM, set off_prompt to False.
• Added support for Rulesets and Rules to all Tasks.
• Added namespace parameter to TextQueryTask.
• Added encoding parameter to TextLoader.
• Added download_objects parameter to AwsS3Client.
• Added encoding parameter to TextLoader and TextArtifact.
• Added to_bytes method to TextArtifact.

🔧 Improvements

• Updated openai's sdk to ^1.1.0.
• Renamed Tool Memory to Task Memory. Soon, all Tasks will be able to output Artifacts into Task Memory, giving the ability to share Artifacts across Tasks.
• Added MetaMemory to store ActionSubtask results to improve the LLM's reasoning when using ToolkitTask. Other applications of MetaMemory will be added in future releases.
• Flattened Memory Actions and Tool Actions into Actions to improve the LLM's reasoning when using ToolkitTask.
• Improved system prompt for ToolTask and ToolkitTask to reduce action hallucinations.

📖 Docs

• Updated examples to reflect Tool Memory to Task Memory rename.
• Update examples to reflect flattening of Memory Actions and Tool Actions into Actions.
• Updated Overview and Prompt Driver pages to reflect pip extras changes.
• Updated Workflow example to use new Workflow.insert_tasks method.
• Added note to OpenAI Prompt Driver examples regarding seed and response_format parameters.
• Updated Tool examples to use TaskMemoryClient and off_prompt parameters.

v0.19.4

🔧 Improvements

• Lock openai at ^0.28 to prevent installation issues until #420 is implemented.

v0.19.3

🚨 Breaking Changes

• Updated Structure.add_event_listener to take an EventListener directly.

🔧 Improvements

• Fixed freeze when using the Stream utility multiple times.
• Fixed boto3 session not being passed from AmazonBedrockEmbeddingDriver to BedrockTitanTokenizer.
• Added Structure.remove_event_listener.

v0.19.2

🔧 Improvements

• Fixed requests scope in pyproject.toml.
• Updated Makefile.

v0.19.1

🔧 Improvements

• Moved requests from Tool requirements.txt to hard dependency in pyproject.toml.

v0.19.0

🚨 Breaking Changes

• Removed BufferConversationMemory.
• Simplified Event Listener format.
• Restrict Python support to >=3.9,<3.12. 3.12 support is coming, but is currently blocked by several downstream libraries.
• Removed Tokenizer's encode and decode methods.
• Removed TextToolMemory and BlobToolMemory.

🆕 New Features

• Combined TextToolMemory and BlobToolMemory under ToolMemory.
• Added Conversation Memory buffering via max_runs parameter.
• Added Conversation Memory pruning to prevent long running conversations from exceeding prompt token limit.
• Added prompt streaming to Prompt Drivers that can support it.
• Added EmailLoader.
• Added support for more file extensions in FileManager Tool.
• Added share_file activity to GoogleDriveClient Tool.
• Added serialization and deserialization for all Events.

🔧 Improvements

• Added BaseVectorStoreDriver.QueryResult.id.
• Fixed SummaryConversationMemory bugs.
• Refactored EmailClient to use EmailLoader.
• Refactored Embedding Drivers to use Chunkers for improvements on embedding large text.
• Added parsing of OpenAi rate limit headers for future use.
• Added __len__ magic methods for Artifacts (thanks @Bubble-Interface!).
• Fixed boto3 session not being passed from BedrockTitanEmbeddingDriver into BedrockTitanTokenizer.

📖 Docs

• Updated docs for new Event Listener format.
• Added page for Extraction Engines.
• Added sections for additional Task types.
• Added page for OpenWeatherClient Tool.
• Added example for PgVectorStoreDriver
• Updated page for Conversation Memory buffering changes.

v0.18.2

🔧 Improvements

• Updated snowflake-connector-python dependency that was causing issues in Python 3.11.

v0.18.1

🔧 Improvements

• Fixed prompt driver propagation to tool memory defaults in Structures.
• Removed engine defaults in TextToolMemory.
• Improved test coverage.

v0.18.0

🆕 New Features

• Added new embedding driver for Amazon Bedrock: BedrockTitanEmbeddingDriver.
• Added new prompt driver for Amazon Bedrock: AmazonBedrockPromptDriver.
• Added new Amazon Bedrock prompt models: BedrockClaudePromptModelDriver, and BedrockJurassicPromptModelDriver, BedrockTitanPromptModelDriver.
• Added new Amazon Bedrock tokenizers: BedrockClaudeTokenizer, BedrockJurassicTokenizer, and BedrockTitanTokenizer.
• Added ignored_exception_types to ExponentialBackoffMixin to ignore certain failures and not retry.
• Added extraction engines for extracting data into JSON objects and CSV rows: JsonExtractionEngine and CsvExtractionEngine.
• Added ExtractionTask to integrate extraction engines into workflows and pipelines.
• Added the GoogleDriveClient tool to access and search files in Google Drive.
• Added the GoogleDocsClient tool to access and edit Google Docs documents.
• Added PgVectorVectorStoreDriver to store and retrieve embeddings in Postgres.

🔧 Improvements

• Added InvalidRequestError to ignored_exception_types in OpenAI drivers.
• Updated all relevant tools to use Griptape loaders.
• Abstracted TextToolMemory and ToolOutputProcessor activities into TextMemoryActivitiesMixin.
• Updated the Calculator tool to use numexpr to improve security.
• Improved base prompts.
• Updated GoogleCalendarClient and GoogleGmailClient to support the owner_email property.
• Fixed OpenAI GPT-4 throttling issues in prompt drivers.