Building Agents
########################

This system was designed to be extensible allowing you to
design new agents to be utilized by orchestrator or other tool agents.
This is achieved using ``recap.agents.base.BaseAgent`` class. 

Here is an example:

.. code-block:: python

    from recap.agents.base import BaseAgent
    from pydantic import BaseModel, Field

    class ExampleInputs(BaseModel):
        query: str = Field(
            description="Natural language query that will be parsed to filter and search the Meta-Review Dataset"
        )
        another_param: int = Field(description="This is another required example parameter")
    
    class NewAgent
        description = "Description of agent to be used by orchestrator. Describe inputs, expected outputs, when to use, etc.
        args_schema = ExampleInputs
        name = "new_agent"

        def __init__(self, model, tools: list=None) -> None:
            super().__init__(model=model, tools=tools)

        ...

        # Invoke method must match args_schema
        def invoke(self, query: str, another_param) -> str:
            ...
            # Do something
            ...
            return json_string

To implement base image class there are 4 requirements:

- name: This is used to identify the tool on frontend and load assocatied prompt.md
- description: A description of the agent used as tool description for orchestrator agent
- args_schema: Pydantic BaseModel with the inputs to the tool call
- invoke: Python function used as tool for .run_as_tool(). The inputs must match schema defintion and it is expected to return a json string.

Once you have your agent created you can run your agent directly:

.. code-block:: python

    # Create model for agent to use if necessary
    from langchain_ollama import ChatOllama
    model = ChatOllama(
        model="gemma4:e4b",
        temperature=0,
        num_ctx=(2048 * 4),
        base_url=os.getenv("OLLAMA_BASE_URL"),
    )

    new_agent = NewAgent(model=model)
    new_agent.invoke("This is my user query", 5)


Or use at as tool for another agent:

.. code-block:: python

    # Create model for agent to use if necessary
    from langchain_ollama import ChatOllama
    from orchestrator.a
    from recap.agents import Orchestrator

    model = ChatOllama(
        model="gemma4:e4b",
        temperature=0,
        num_ctx=(2048 * 4),
        base_url=os.getenv("OLLAMA_BASE_URL"),
    )

    new_agent = NewAgent(model=model)
    orchestrator = Orchestrator(model=model, tools=[new_agent.as_tool()])