backend/app/services/inference_service.py

from collections.abc import AsyncGenerator

from backend.app.guardrails.safety import is_unsafe_prompt, needs_human_review
from backend.app.models.schemas import ChatMessage, ChatResponse, IncidentClassificationResponse


def _last_user_message(messages: list[ChatMessage]) -> str:
    for message in reversed(messages):
        if message.role == "user":
            return message.content
    return ""


NEPALI_GUIDANCE = (
    "आपतकालीन प्रतिक्रियाको सुझावकात विधिहरू: "
    "तत्काल खतरा मूल्याङ्कन गर्नुहोस्, "
    "सुरक्षित ठाउँमा जानुहोस्, "
    "आपतकालीन सेवा (१०१ / १०२) मा फोन गर्नुहोस्, "
    "कमजोर वर्गको सुरक्षा गर्नुहोस्, "
    "र हर १५ मिनेटमा आधिकारिक सूचना अनुसरण गर्नुहोस्।"
)


def generate_response(messages: list[ChatMessage], language: str, region: str) -> ChatResponse:
    user_text = _last_user_message(messages)
    if is_unsafe_prompt(user_text):
        return ChatResponse(
            answer=(
                "I cannot provide guidance for unsafe actions. Contact local emergency authorities "
                "and follow official evacuation and safety protocols immediately."
            ),
            confidence=0.99,
            needs_human_review=True,
            citations=["Local emergency management authority", "Official public safety bulletins"],
        )

    if language.strip().lower() in {"nepali", "ne", "नेपाली"}:
        return ChatResponse(
            answer=f"[WorldDisasterLM-8B | नेपाली | {region}] {NEPALI_GUIDANCE}",
            confidence=0.74,
            needs_human_review=False,
            citations=[
                "NDRRMA नेपाल विपद् व्यवस्थापन प्राधिकरण",
                "WHO आपतकालीन प्रतिक्रिया मार्गदर्शन",
                "UNDRR Sendai Framework 2015-2030",
            ],
        )

    answer = (
        f"[WorldDisasterLM-8B | {language} | {region}] Recommended next steps: assess immediate hazards, move to a safe "
        "location, call emergency services, protect vulnerable groups, and verify updates from "
        "official alerts every 15 minutes."
    )
    confidence = 0.74
    return ChatResponse(
        answer=answer,
        confidence=confidence,
        needs_human_review=needs_human_review(confidence, answer),
        citations=["UNDRR preparedness guidelines", "WHO emergency response guidance"],
    )


async def stream_response(messages: list[ChatMessage], language: str, region: str) -> AsyncGenerator[str, None]:
    response = generate_response(messages, language=language, region=region)
    for token in response.answer.split():
        yield token + " "


def classify_incident(text: str) -> IncidentClassificationResponse:
    lowered = text.lower()

    mapping = {
        "earthquake": "earthquake",
        "tsunami": "tsunami",
        "flood": "flood",
        "wildfire": "wildfire",
        "pandemic": "public_health",
        "epidemic": "public_health",
        "chemical": "industrial",
        "nuclear": "industrial",
        "refugee": "humanitarian",
        "drought": "climate",
        "heatwave": "climate",
    }

    incident_type = "unknown"
    for keyword, event_type in mapping.items():
        if keyword in lowered:
            incident_type = event_type
            break

    if any(token in lowered for token in ["mass", "collapse", "critical", "urgent", "dead"]):
        severity = "critical"
    elif any(token in lowered for token in ["severe", "major", "injured", "evacuate"]):
        severity = "high"
    elif any(token in lowered for token in ["moderate", "contained", "localized"]):
        severity = "medium"
    else:
        severity = "low"

    return IncidentClassificationResponse(
        incident_type=incident_type,
        severity=severity,
        rationale="Keyword and severity heuristic classifier; replace with fine-tuned classifier model.",
    )