app.py

import os
import json
import requests
import gradio as gr

BACKEND_PUBLIC_URL = os.getenv("BACKEND_PUBLIC_URL", "http://134.199.203.136:8080").rstrip("/")
AMD_LLM_MODEL = os.getenv("AMD_LLM_MODEL", "Qwen/Qwen2.5-7B-Instruct")
DEMO_MODE = os.getenv("DEMO_MODE", "true")


def backend_get(path: str):
    url = f"{BACKEND_PUBLIC_URL}{path}"
    try:
        response = requests.get(url, timeout=25)
        response.raise_for_status()
        try:
            return response.json()
        except Exception:
            return {"raw": response.text[:3000]}
    except Exception as exc:
        return {
            "error": str(exc),
            "url": url,
            "hint": "Check that the AMD backend is running and reachable."
        }


def backend_post(path: str, payload: dict):
    url = f"{BACKEND_PUBLIC_URL}{path}"
    try:
        response = requests.post(url, json=payload, timeout=90)
        response.raise_for_status()
        try:
            return response.json()
        except Exception:
            return {"raw": response.text[:5000]}
    except Exception as exc:
        return {
            "error": str(exc),
            "url": url,
            "hint": "Check the backend route in /docs. If this route does not exist, use the fallback demo."
        }


def check_backend():
    health = backend_get("/health")
    amd = backend_get("/api/amd/performance")

    return json.dumps(
        {
            "backend_url": BACKEND_PUBLIC_URL,
            "health": health,
            "amd_performance": amd,
            "model": AMD_LLM_MODEL,
            "demo_mode": DEMO_MODE,
        },
        indent=2,
        ensure_ascii=False,
    )


def run_santa_ana_demo():
    payload = {
        "scenario": "santa_ana",
        "demo_mode": True,
        "model": AMD_LLM_MODEL,
    }

    result = backend_post("/api/demo/run", payload)
    return json.dumps(result, indent=2, ensure_ascii=False)


def run_fallback_demo():
    result = {
        "status": "fallback_demo",
        "message": "Synthetic fallback demo. AMD backend may be offline or demo route may be unavailable.",
        "scenario": "Santa Ana wildfire synthetic demo",
        "signals": [
            {
                "type": "text",
                "content": "Smoke visible near hillside. Two families requesting evacuation assistance.",
                "priority_hint": "high"
            },
            {
                "type": "image_report",
                "content": "Simulated image evidence: smoke plume near residential area.",
                "priority_hint": "medium"
            },
            {
                "type": "csv_location",
                "content": "Lat/lon cluster indicates repeated alerts from the same area.",
                "priority_hint": "high"
            }
        ],
        "consolidated_incident": {
            "incident_type": "wildfire",
            "priority": "P1",
            "location": "Santa Ana synthetic demo area",
            "evidence_count": 3
        },
        "recommended_resources": [
            "evacuation support",
            "medical triage unit",
            "fire response team",
            "public alert message"
        ],
        "dispatch_message": "P1 wildfire risk near residential area. Dispatch evacuation support and fire response. Monitor wind direction and confirm affected households.",
        "amd_note": "Full backend is expected to run on AMD Developer Cloud. This fallback does not claim live GPU inference."
    }

    return json.dumps(result, indent=2, ensure_ascii=False)


with gr.Blocks(title="ReliefLens AI") as demo:
    gr.Markdown(
        """
        # 🚨 ReliefLens AI

        **Agentic disaster-response triage powered by AMD Developer Cloud.**

        ReliefLens AI turns chaotic multimodal field reports into structured incidents,
        priority labels, evidence links, resource recommendations, and dispatch-ready messages.

        This public Space calls a FastAPI backend deployed on AMD Developer Cloud.
        """
    )

    with gr.Row():
        gr.Textbox(value=BACKEND_PUBLIC_URL, label="AMD Backend URL", interactive=False)
        gr.Textbox(value=AMD_LLM_MODEL, label="Target model", interactive=False)

    with gr.Tab("1. AMD Backend Status"):
        gr.Markdown("Check whether the AMD Cloud backend is reachable.")
        status_btn = gr.Button("Check Backend")
        status_output = gr.Code(label="Backend status", language="json")
        status_btn.click(fn=check_backend, outputs=status_output)

    with gr.Tab("2. Santa Ana Demo"):
        gr.Markdown(
            """
            Runs the synthetic Santa Ana crisis-response scenario through the AMD backend.

            If the backend endpoint is not available yet, use the fallback tab.
            """
        )
        run_btn = gr.Button("Run Santa Ana Demo via AMD Backend")
        run_output = gr.Code(label="Demo output", language="json")
        run_btn.click(fn=run_santa_ana_demo, outputs=run_output)

    with gr.Tab("3. Offline Fallback Demo"):
        gr.Markdown("Safe local fallback for judges if the AMD VM is offline.")
        fallback_btn = gr.Button("Run Fallback Demo")
        fallback_output = gr.Code(label="Fallback output", language="json")
        fallback_btn.click(fn=run_fallback_demo, outputs=fallback_output)

    with gr.Tab("4. Safety / Limitations"):
        gr.Markdown(
            """
            ## Safety and limitations

            - This demo uses synthetic disaster-response data.
            - It is not connected to emergency services.
            - It should not be used for real-world emergency dispatch.
            - AMD Cloud validation should be shown with `rocm-smi`, backend status, and performance telemetry.
            - If the AMD VM is destroyed to save credits, the fallback demo remains available.
            """
        )


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