src/data/location.py

"""Normalize EM-DAT location field into a single short search token.

The EM-DAT `location` field is inconsistent — sometimes a single city
("Valencia Province"), sometimes a 10-item list of tiny administrative units
("Gorenjska, Goriska, Jugovzodna Slovenija, Koroska, ..."), sometimes null.
Neither raw form is directly usable in a news search query.

This module wraps a single LLM call that returns ONE token suitable for
dropping into a GDELT query. When the location is too diffuse or absent,
it falls back to the country name.
"""

from __future__ import annotations

import logging
import re

from src.llm.client import LLMClient
from src.models.schemas import FloodEvent

logger = logging.getLogger(__name__)


def extract_location_token(
    event: FloodEvent, llm_client: LLMClient, config: dict
) -> str:
    """Return a single place token for use in news queries.

    Falls back to the country name on any LLM failure or empty output.
    """
    variables = {
        "country": event.country,
        "location": event.location or "null",
    }
    try:
        response = llm_client.call_with_config(
            prompt_key="location_extract",
            knowledge_key="search",  # reuse expert_search as knowledge placeholder
            variables=variables,
            config=config,
        )
    except Exception as e:
        logger.warning(f"Location extraction failed for {event.event_id}: {e}")
        return event.country

    # Take first non-empty line, strip quotes/punctuation
    token = ""
    for line in response.splitlines():
        s = line.strip().strip('"').strip("'").strip()
        if s:
            token = s
            break

    # Sanity checks: must be short, alphabetic-ish, not empty
    if not token or len(token) > 50 or not re.search(r"[A-Za-z]", token):
        logger.warning(
            f"Location extraction produced invalid token '{token}' for {event.event_id}; "
            f"falling back to country"
        )
        return event.country
    return token