gradio_agent_inspector

Agent Inspector for ADK

Installation

pip install gradio_agent_inspector

Usage

import json
from pathlib import Path
import gradio as gr
from gradio_agent_inspector import AgentInspector
import os


def simulate_conversation():
    dir_path = Path(os.path.dirname(os.path.realpath(__file__)))

    initial_state = {
        "state": {},
        "events": [],
    }

    states = []
    for i in range(2):
        session_value_p = dir_path / "session-sample" / f"value-{i}.json"
        with session_value_p.open("r", encoding="utf-8") as f:
            session_value = json.load(f)

            # attach event trace and graph info to the event
            for e in session_value["events"]:
                event_trace_p = dir_path / "event-trace" / f"trace-{e['id']}.json"
                if event_trace_p.exists():
                    with event_trace_p.open("r", encoding="utf-8") as trace_f:
                        event_trace = json.load(trace_f)
                        if "gcp.vertex.agent.llm_request" in event_trace:
                            event_trace["gcp.vertex.agent.llm_request"] = json.loads(
                                event_trace["gcp.vertex.agent.llm_request"]
                            )
                        if "gcp.vertex.agent.llm_response" in event_trace:
                            event_trace["gcp.vertex.agent.llm_response"] = json.loads(
                                event_trace["gcp.vertex.agent.llm_response"]
                            )
                        e["trace"] = event_trace
                event_graph_p = dir_path / "event-trace" / f"graph-{e['id']}.json"
                if event_graph_p.exists():
                    with event_graph_p.open("r", encoding="utf-8") as graph_f:
                        event_graph = json.load(graph_f)
                        e["graph"] = event_graph
            states.append(session_value)

    return initial_state, states


def update_conversation_state(state_index, states):
    if (state_index + 1) >= len(states):
        return states[state_index], state_index
    else:
        new_index = state_index + 1
        return states[new_index], new_index


initial_state, conversation_states = simulate_conversation()

with gr.Blocks() as demo:
    gr.Markdown("# Agent Inspector")

    state_counter = gr.State(-1)

    agent_inspector = AgentInspector(json.dumps(initial_state))

    with gr.Row():
        next_btn = gr.Button(
            f"▶️ Next ({0} / {len(conversation_states)})", variant="primary"
        )
        reset_btn = gr.Button("🔄 Reset", variant="secondary")

    def next_state(current_counter):
        new_state, new_counter = update_conversation_state(
            current_counter, conversation_states
        )

        json_state = json.dumps(new_state)
        next_button_label = f"▶️ Next ({new_counter+1} / {len(conversation_states)})"

        return json_state, new_counter, next_button_label

    def reset_conversation():
        json_state = json.dumps(initial_state)
        next_button_label = f"▶️ Next ({0} / {len(conversation_states)})"

        return json_state, -1, next_button_label

    next_btn.click(
        next_state,
        inputs=[state_counter],
        outputs=[agent_inspector, state_counter, next_btn],
    )

    reset_btn.click(
        reset_conversation, outputs=[agent_inspector, state_counter, next_btn]
    )

    # examples = gr.Examples(
    #     examples=[
    #         s for s in conversation_states
    #     ],
    #     inputs=[initial_state],
    # )

if __name__ == "__main__":
    demo.launch()

AgentInspector

Initialization

str | Callable | None

str | None

str | I18nData | None

Timer | float | None

Component | Sequence[Component] | set[Component] | None

bool | None

int | None

int

bool | None

bool

bool

str | None

list[str] | str | None

bool

int | str | tuple[int | str, ...] | None

list[str] | str | None

Events

User function

The impact on the users predict function varies depending on whether the component is used as an input or output for an event (or both).

• When used as an Input, the component only impacts the input signature of the user function.
• When used as an output, the component only impacts the return signature of the user function.

The code snippet below is accurate in cases where the component is used as both an input and an output.

• As output: Is passed, passes text value as a {str} into the function.
• As input: Should return, expects a {str} returned from function and sets textarea value to it.

def predict(
    value: str | None
) -> str | None:
    return value