You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Add support for Vertex AI models, on top of the Gemini, Gemma, and OpenAI models that were supported in the initial v0.1.0 elease.
Add an HTMLRenderer library for rendering an ExecutionResult, ExperimentResult, or ExperimentSummary as a block of HTML suitable for interactive display in colab.
Various bug fixes and incremental improvements to ReActAgent, PythonPlanningAgent, tracing, and handling of multimodal content chunks.
v0.1.0
Initial release including support for three kinds of models:
GeminiAPI: remote connection to any of the models supported by the Gemini API.
Gemma: possibility to load and use open-weights Gemma models.
OpenAIAPI: remote connection to any of the models supported by the OpenAI API.
OneTwo core includes support for asynchronous execution, batching, caching, tracing, sequential and parallel flows, prompting templating via Jinja2 or composables, and multimodal support.
OneTwo standard library includes off-the-shelf implementations of two popular tool use strategies in the form of a ReActAgent and a PythonPlanningAgent, along with a PythonSandbox API, simple autorater critics, and a generic BeamSearch implementation that can be composed with arbitrary underlying agents for producing multi-trajectory strategies such as Tree-of-Thoughts.