import os
import logging
import numpy as np
from repositories.client_trips_repository import client_trips_repository

logger = logging.getLogger(__name__)

OPENROUTER_API_KEY = os.getenv('OPENROUTER_API_KEY')
EMBEDDING_MODEL = 'openai/text-embedding-3-small'
SIMILARITY_THRESHOLD = 0.75


def save_client_trip(tenant_id, destination, duration, travel_style, product_name=None,
                      url=None, extracted=None, price_usd=None, activities=None):
    """
    Upserts a client_trips record from a PM's trip-intent description
    (CLAUDE.md §8's TripIntentForm — destination/duration/travel_style,
    optionally a product name). Every /research submission calls this,
    which is the only place client_trips ever gets populated — without
    it, find_comparable_trips() has nothing on the client side to match
    competitor products against.

    url/extracted/price_usd/activities are optional and only ever passed
    by scrape_own_trip() below, once the tenant's own trip page has
    actually been scraped and extracted — the original typed-values-only
    call site in api.py's /research route passes none of these, so its
    behavior is unchanged.

    Dedupes on (tenant_id, trip_name) so re-running analysis for the
    same product doesn't create duplicate rows — it just keeps the
    existing one current.

    Data flow:
      destination/duration/travel_style/product_name (+ optional
      url/extracted/price_usd/activities from a real scrape) →
      trip_name resolved (product_name if given, else a
      destination+style fallback) →
      client_trips_repository.get_by_tenant_and_name() →
      found: update in place → not found: create →
      client_trips record returned
    """
    trip_name = product_name or f'{destination} {travel_style}'.strip()
    payload = {
        'tenant_id': tenant_id,
        'trip_name': trip_name,
        'destination': destination,
        'duration_days': duration,
    }
    if url is not None:
        payload['url'] = url
    if extracted is not None:
        payload['extracted'] = extracted
    if price_usd is not None:
        payload['price_usd'] = price_usd
    if activities is not None:
        payload['activities'] = activities

    existing = client_trips_repository.get_by_tenant_and_name(tenant_id, trip_name)
    if existing:
        # A stale embedding would silently never get regenerated —
        # /internal/embed-products only embeds trips missing a vector —
        # so clear it whenever the text embed_trip() hashes actually changed.
        if existing.get('destination') != destination or existing.get('duration_days') != duration:
            payload['embedding'] = None
        return client_trips_repository.update(existing['id'], payload)
    return client_trips_repository.create(payload)


def scrape_own_trip(tenant_id, website_url, context):
    """
    Scrapes and extracts the tenant's OWN confirmed/matched trip page —
    the same Firecrawl-matched-or-manually-swapped URL pattern already
    used for competitors, pointed at the tenant's own site instead. Gives
    build_prompt()'s own_trip param and client_trips real ground-truth
    data instead of only the PM's typed destination/duration/style.

    Reuses the same CACHE_STALENESS_HOURS freshness window as competitors
    (core/scraper.py) so the tenant's own site isn't rescraped on every
    single analysis run — only when the cached client_trips row is stale
    or context['force_refresh'] is set, both exactly mirroring
    analysis_service.needs_refresh()'s competitor logic.

    Never raises — a side-effect failure here (site blocked, proxy
    error, extraction failure) must not fail the whole batch (CLAUDE.md
    §18). Logs and returns None on any failure, which build_prompt()
    treats identically to a tenant who never confirmed an own-trip URL.

    The returned dict always carries a `cached` bool (True when served
    from client_trips.extracted without a live scrape, False when just
    freshly extracted) — surfaced to the dashboard via jobs.py's
    results['own_trip'] so ResultsTable.vue can show the same "from
    PocketBase vs new scrape" pill on the own-trip column as every
    competitor column. `link` is always overwritten with the real
    website_url actually scraped, rather than trusting the LLM's own
    echoed `"link"` field verbatim — same reasoning as
    analysis_service._save_scrape_result() using page_data['url'], not
    extracted.get('link'), for a competitor's stored product URL.

    Data flow:
      website_url + context → trip_name resolved exactly as
      save_client_trip() does (product_name, else destination+style) →
      existing client_trips row fresh (and not force_refresh)? → return
      its cached `extracted` blob (cached=True) → otherwise:
      core.scraper.scrape() → build_own_trip_prompt() →
      core.extractor.extract_with_openrouter() → link corrected to the
      real website_url → save_client_trip() with
      url/extracted/price_usd/activities → extracted dict returned
      (cached=False)
    """
    from datetime import datetime, timedelta, timezone

    trip_name = context.get('product_name') or f"{context.get('destination')} {context.get('travel_style')}".strip()
    existing = client_trips_repository.get_by_tenant_and_name(tenant_id, trip_name)
    force_refresh = context.get('force_refresh', False)

    if existing and existing.get('extracted') and not force_refresh:
        last_scraped = existing.get('last_scraped')
        if last_scraped:
            try:
                from core.scraper import CACHE_STALENESS_HOURS
                scraped_at = datetime.fromisoformat(last_scraped.replace('Z', '+00:00'))
                if datetime.now(timezone.utc) - scraped_at <= timedelta(hours=CACHE_STALENESS_HOURS):
                    logger.info(f'Own trip for tenant {tenant_id} — cache fresh, reusing extracted data')
                    return {**existing['extracted'], 'cached': True}
            except (ValueError, AttributeError):
                pass  # unparseable timestamp — fall through and rescrape

    try:
        from core.scraper import scrape
        from core.extractor import extract_with_openrouter
        from industries.adventure_travel.prompt import build_own_trip_prompt

        page_data = scrape(website_url)
        if not page_data.get('success'):
            logger.warning(f'Own-trip scrape failed for tenant {tenant_id}: {page_data.get("text")}')
            return None

        prompt = build_own_trip_prompt(
            page_data, context.get('product_name'), context.get('destination'),
            context.get('duration'), context.get('travel_style'),
            context.get('company_name'), is_ngs=context.get('tab_type') == 'NGS',
        )
        extracted = extract_with_openrouter(prompt, content=page_data['text'])
        extracted['link'] = website_url

        saved = save_client_trip(
            tenant_id, context.get('destination'), context.get('duration'),
            context.get('travel_style'), product_name=context.get('product_name'),
            url=website_url, extracted=extracted,
            price_usd=extracted.get('majorityPrice'), activities=extracted.get('activities'),
        )
        client_trips_repository.update(saved['id'], {'last_scraped': datetime.now(timezone.utc).isoformat()})
        return {**extracted, 'cached': False}
    except Exception as e:
        logger.error(f'scrape_own_trip failed for tenant {tenant_id}: {str(e)}')
        return None


def embed_trip(trip):
    """
    Generates a semantic embedding for a trip using OpenRouter's
    text-embedding-3-small, combining trip name, destination, duration,
    and activities into a single text representation first.

    Data flow:
      trip dict → formatted text string → OpenRouter embeddings API →
      vector list returned for storage in products.embedding /
      client_trips.embedding
    """
    # Imported lazily so this module has no hard dependency on the
    # `openai` package at import time for callers that only need
    # find_comparable_trips() (pure math, no API calls).
    from openai import OpenAI

    text = (
        f"{trip.get('trip_name', '')}\n"
        f"Destination: {trip.get('destination', '')}\n"
        f"Duration: {trip.get('duration_days', '')} days\n"
        f"Activities: {trip.get('activities', '')}\n"
        f"Start: {trip.get('start_location', '')}"
    )

    client = OpenAI(api_key=OPENROUTER_API_KEY, base_url='https://openrouter.ai/api/v1')
    response = client.embeddings.create(model=EMBEDDING_MODEL, input=text)
    return response.data[0].embedding


def _cosine_similarity(a, b):
    """Standard cosine similarity between two equal-length embedding vectors."""
    a, b = np.array(a), np.array(b)
    denom = (np.linalg.norm(a) * np.linalg.norm(b))
    if denom == 0:
        return 0.0
    return float(np.dot(a, b) / denom)


def find_comparable_trips(client_trips, competitor_trips, threshold=SIMILARITY_THRESHOLD):
    """
    Calculates cosine similarity between every client trip and every
    competitor trip, returning matches above threshold sorted by
    similarity score descending.

    Data flow:
      client_trips + competitor_trips (each with .embedding) →
      pairwise cosine similarity → filter by threshold →
      sort by score descending → list of match dicts returned
    """
    matches = []
    for client_trip in client_trips:
        client_embedding = client_trip.get('embedding')
        if not client_embedding:
            continue
        for competitor_trip in competitor_trips:
            competitor_embedding = competitor_trip.get('embedding')
            if not competitor_embedding:
                continue
            score = _cosine_similarity(client_embedding, competitor_embedding)
            if score >= threshold:
                matches.append({
                    'client_product': client_trip.get('trip_name'),
                    'competitor_product': competitor_trip.get('id'),
                    'similarity_score': round(score, 4),
                    'price_difference': _safe_diff(
                        competitor_trip.get('majority_price_usd'), client_trip.get('price_usd')
                    ),
                    'duration_difference': _safe_diff(
                        competitor_trip.get('duration_days'), client_trip.get('duration_days')
                    ),
                })
    matches.sort(key=lambda m: m['similarity_score'], reverse=True)
    return matches


def _safe_diff(a, b):
    """Returns a - b if both are present, else None — keeps the diff fields nullable per schema."""
    if a is None or b is None:
        return None
    return a - b