MangoAgent

from mango.agent import MangoAgent

Constructor

MangoAgent(
    llm_service: LLMService,
    tool_registry: ToolRegistry,
    db: NoSQLRunner | None = None,
    agent_memory: MemoryService | None = None,
    schema: dict[str, SchemaInfo] | None = None,
    introspect: bool = False,
    max_iterations: int = 8,
    memory_top_k: int = 3,
    max_turns: int = 5,
)

Methods

setup() → None

Initializes the agent. Runs schema introspection if introspect=True and db is set. Builds the system prompt. Must be called before the first ask().

Called automatically on the first ask() if not called manually.

ask(question, on_tool_call=None) → AgentResponse

async

Asks a question and returns the complete answer.

Parameters:

• question: str — natural language question
• on_tool_call: Callable[[str, dict, str], None] | None — optional callback called after each tool execution with (tool_name, tool_args, result_text)

Returns: AgentResponse

response = await agent.ask("How many users signed up last week?")

ask_stream(question) → AsyncGenerator[dict, None]

async generator

Same as ask() but yields events as they happen. Use for real-time UIs.

Parameters:

• question: str — natural language question

Yields dicts with type field:

async for event in agent.ask_stream("..."):
    match event["type"]:
        case "tool_call":
            # event["tool_name"], event["tool_args"]
        case "tool_result":
            # event["tool_name"], event["success"], event["preview"]
        case "answer":
            # event["text"]
        case "done":
            # event["iterations"], event["input_tokens"],
            # event["output_tokens"], event["memory_hits"],
            # event["tool_calls_made"]

new_session() → MangoAgent

Returns a new agent with the same configuration but a fresh conversation history. Schema and system prompt are reused — no re-introspection.

session = agent.new_session()

reset_conversation() → None

Clears the conversation history in place.

agent.reset_conversation()

Properties

AgentResponse

@dataclass
class AgentResponse:
    answer: str                  # natural language answer
    tool_calls_made: list[str]   # names of tools called
    input_tokens: int            # total input tokens across all LLM calls
    output_tokens: int           # total output tokens across all LLM calls
    iterations: int              # number of LLM calls in this turn
    memory_hits: int             # number of memory examples injected