"""채널 URL 발견 — 겹들을 엮어 결과를 만드는 자리."""
from dataclasses import dataclass, field

import httpx

from common.logger import LOG
from services.grounding.channels import (
    DiscoveredLink,
    classify_url,
    collect_links,
    filter_links,
    search_count,
)
from services.llm.perplexity import (
    DEFAULT_MAX_TOKENS,
    DEFAULT_MODEL,
    DEFAULT_TIMEOUT,
    PerplexityError,
    PerplexityNotConfigured,
    call,
    is_configured,
    read_usage,
)
from services.prompts.channel_discovery import RESPONSE_SCHEMA, SYSTEM_PROMPT, build_prompt

# 후기 유입을 막기 위해 공식 채널 도메인만 검색한다.
SEARCH_DOMAIN_FILTER = [
    "yanolja.com",
    "goodchoice.kr",
    "place.naver.com",
    "map.naver.com",
    "naver.me",
]

# 이 횟수를 넘으면 프롬프트·도메인 필터를 의심한다 — 검색 요금은 토큰 요금과 별도다.
SEARCH_COUNT_WARN_THRESHOLD = 8



@dataclass
class ChannelDiscovery:
    """채널 발견 결과."""

    links: list[DiscoveredLink] = field(default_factory=list)
    raw: dict = field(default_factory=dict)
    search_count: int = 0
    filtered_out: list[tuple[str, str]] = field(default_factory=list)

    def reason_counts(self) -> dict[str, int]:
        """탈락 사유별 건수."""
        counts: dict[str, int] = {}
        for _url, reason in self.filtered_out:
            counts[reason] = counts.get(reason, 0) + 1
        return counts


async def discover_channels(
    name: str,
    address: str | None = None,
    category_hint: str | None = None,
    *,
    model: str = DEFAULT_MODEL,
    include_blogs: bool = False,
    client: httpx.AsyncClient | None = None,
) -> ChannelDiscovery:
    """상호명으로 채널 URL 후보를 찾는다."""
    if not (name or "").strip():
        raise PerplexityError("상호명이 비어 있다")

    body = {
        "model": model,
        "messages": [
            {
                "role": "system",
                "content": SYSTEM_PROMPT,
            },
            {"role": "user", "content": build_prompt(name, address, category_hint)},
        ],
        "max_tokens": DEFAULT_MAX_TOKENS,
        "temperature": 0,                              # URL 수집이라 창의성이 해롭다
        "response_format": RESPONSE_SCHEMA,
        "search_domain_filter": SEARCH_DOMAIN_FILTER,  # 야놀자·여기어때·네이버로 한정
    }
    payload = await call(body, client=client)

    found, dropped_raw = collect_links(payload)
    links, dropped_quality = filter_links(found, include_blogs)
    filtered_out = dropped_raw + dropped_quality
    searches = search_count(payload)

    result = ChannelDiscovery(
        links=links, raw=payload, search_count=searches, filtered_out=filtered_out
    )

    # 내부 검색 횟수는 품질·지연 관측값이다.
    usage = read_usage(payload)
    reasons = result.reason_counts()
    reason_text = " ".join(f"{k}{v}" for k, v in sorted(reasons.items())) or "없음"
    LOG.i(
        f"[perplexity] '{name}' 검색={searches}회 발견={len(found)} 통과={len(links)} "
        f"탈락={len(filtered_out)}({reason_text}) "
        f"tokens in={usage.input_tokens} out={usage.output_tokens} · 약 ${usage.cost}"
    )
    if searches > SEARCH_COUNT_WARN_THRESHOLD:
        LOG.w(
            f"[perplexity] 검색 {searches}회 — 기준({SEARCH_COUNT_WARN_THRESHOLD}회) 초과. "
            f"지연·오탐 후보가 늘 수 있으니 도메인 필터·프롬프트를 확인하라"
        )

    # raw 는 응답 전체를 그대로 둔다 — 나중에 환각 추적에 쓴다(사실 근거로는 쓰지 않는다).
    return result