v1.1.3

Bump to version 1.1.3

v1.1.2

Bumb version to 1.1.2

v1.1.1

Merge pull request #22 from jataware/feat/flexible-json-handling

Add unwrapping of json "blocks" returned by gpt-4-turbo

v1.0.6

Added convenience synchronous versions of async methods on the Agent base class

• all_messages_sync
• query_sync
• observe_sync
• error_sync
• execute_sync
• oneshot_sync

v1.0.5

Added new auto-context functionality that allows providing an auto-context function to the agent that can be run automatically to update the agent's context before queries, etc for use in dynamic situations when state may change outside of the normal conversational flow.

v1.0.4

Merge pull request #15 from jataware/tool-method-update-fix

Slightly more complicated solution than hoped for, but all working

v1.0.2

Bump version for release

v1.0.1

Changes

• Made rich printing optional via the rich_print flag or by setting the DISABLE_RICH_PRINT environment variable

v1.0.0

Breaking Changes:

• reworked how context messages are created/handled. Now all messages are created with add_context, which returns an id to that message
  • message and context messages are now Message and ContextMessage classes, rather than plain dictionaries
  • timed messages can be created by setting the lifetime argument when creating a context message
  • messages can be deleted with the clear_context function
• renamed loop controller values to be more clear:
  • STOP -> STOP_SUCCESS
  • ERROR -> STOP_FATAL

Non-breaking changes

• created dedicated functions for determining which type of tool a given object is (i.e. one of func tool, method tool, class tool, class tool instance)

v0.3.0

• added dependency injection
  • reference to the underlying agent
  • reference to the underlying tool/tool-name
  • managing react loop flow
• refactored how decorators modified functions/classes. Now objects retain original functionality and properties for the LLMs use are attached
• Added the ability to create custom spinners/functionality while LLM is "thinking", by passing in a custom python context manager
• made LLM accept API key as either a function argument or environment variable. Check is done at agent instantiation, rather than import time
• Added context management to chat agent message history
  • permanent context: messages that permanently stay in the chat
  • timed context: messages that stay in the chat for some number of timesteps
  • managed context: messages that stay in the chat until a deleter callback is called
• Added a oneshot method on agent for accessing the LLM without including any chat history or context
• fixed issue with prompt not changing to reflect when ask_user was set to False
• added a python tool for letting the LLM write and run code
• Made ReActAgent a subclass of Agent
• fixed bug around prompting for tools that take a single argument of dict/list to be forwarded correctly from LLM to function/method
• added a demo video of using archytas with the python tool