tools

tools are the model’s action space.

define a tool

from benchmax.envs.base_env import ToolDefinition

search_tool = ToolDefinition(
    name="search_corpus",
    description="Search the corpus using BM25 with optional filters.",
    input_schema={
        "type": "object",
        "properties": {
            "query": {"type": "string"},
            "limit": {"type": "integer"},
        },
        "required": ["query"],
    },
)

implement and register

self._tools = {"search_corpus": (search_tool, self._search_impl)}

async def list_tools(self):
    return [defn for defn, _ in self._tools.values()]

async def run_tool(self, rollout_id: str, tool_name: str, **tool_args):
    _, impl = self._tools[tool_name]
    return await impl(**tool_args)

notes

• input_schema should be explicit and stable.
• output should be parseable and consistent.
• keep tool names aligned between schema and implementation.