evofabric.core.agent

evofabric.core.agent#

AgentNode#

class evofabric.core.agent.AgentNode(AsyncStreamNode)[source]#

Agent node integrating memory management, tool invocation, LLM invocation, and input/output formatting. Inherits from AsyncStreamNode

Workflow:

  1. Sequentially call the memory module’s retrieval_update() to update the model context.

  2. Format the input according to the input_msg_format template.

  3. Call LLM Client to get the response.

  4. Execute the invoked tools on demand.

  5. Format the reply according to the output_msg_format template.

  6. Call the memory module’s add_messages() method sequentially to update memory.

Parameters:

  • client (ChatClientBase) – LLM backend client.

  • system_prompt (str, optional) – System prompt.

  • inference_kwargs (Dict, optional) – Inference parameters when calling LLM.

  • tool_manager (Union[ToolManagerBase, List[ToolManagerBase]], optional) – Tool Manager or Manager List.

  • memory (Union[MemBase, List[MemBase]], optional) – Memory component or component list.

  • output_schema (Optional[Type[BaseModel]], optional) – The Pydantic model expected to be returned by the LLM.

  • output_msg_format (Optional[str], optional) – jinja2 template string used for rendering LLM output; must also provide output_schema

  • input_msg_format (Optional[str], optional) – jinja2 template string used to format input state into a single user message.

__call__(self, state: State, stream_writer: StreamWriter) → StateDelta[source]#

Execute agent reasoning.

Parameters:

  • state (State) – Input state must contain the messages field.

  • stream_writer (StreamWriter) – Streaming output writer.

Returns:

Incremental state, containing the updated message.

Return type:

StateDelta

UserNode#

class evofabric.core.agent.UserNode(AsyncNode)[source]#

A node that obtains user input from the terminal and provides asynchronous input to the graph.

Inherits from AsyncNode.

UserNode automatically handles the following exceptions:

  • EOFError: User input stream ended (e.g., end-of-file marker)

  • KeyboardInterrupt: User interrupts input (Ctrl+C)

  • Other Exceptions: Catch all unexpected errors

An empty state increment will be returned in all abnormal situations: {"messages": []}

Parameters:

  • prompt_message (str) – User input prompt. Default: ``Please enter your input: ``

  • input_key (str) – The key name for storing user input in the state. Default is user_input

__call__(state: State) → StateDelta[source]#

Obtain user input and return status increment.

Parameters:

state (Dict) – Current Status

Returns:

State increment including user input

Return type:

StateDelta

Contents