# QC Élection Forecast — Plateforme de prévision électorale du Québec 2026
# Auteur : Simon-Pierre Boucher
# Contact : contact@spboucher.ai
# https://www.qc-election.com
"""« Interroger le modèle » — explications traçables, fondées sur les données.

Chaque affirmation provient d'une comparaison chiffrée entre deux runs du
modèle (probabilités, sièges, vote), des sondages ajoutés entre les deux, du
sentiment médiatique et des événements détectés. L'éventuel LLM ne fait que
reformuler ces faits — jamais générer une prédiction.
"""
from __future__ import annotations

from datetime import date, datetime, timedelta, timezone

from sqlalchemy.orm import Session

from ..config import settings
from .. import models as Mo


def _latest_run(db: Session, before: date | None = None) -> Mo.ForecastRun | None:
    q = db.query(Mo.ForecastRun).filter(Mo.ForecastRun.is_backtest.is_(False))
    if before:
        q = q.filter(Mo.ForecastRun.as_of <= before)
    return q.order_by(Mo.ForecastRun.as_of.desc(), Mo.ForecastRun.id.desc()).first()


def gather_facts(db: Session, days_back: int = 7) -> dict:
    """Faits comparables entre le run actuel et celui d'il y a ~days_back jours."""
    cur = _latest_run(db)
    if cur is None:
        return {"error": "aucun forecast disponible"}
    prev = _latest_run(db, before=cur.as_of - timedelta(days=days_back))
    facts: dict = {"as_of": cur.as_of.isoformat(), "run_id": cur.id,
                   "compared_to": prev.as_of.isoformat() if prev else None,
                   "changes": [], "new_polls": [], "events": [], "sentiment": {}}

    parties = [p for p in settings.parties if p != "AUT"]
    for p in parties:
        cur_s = cur.seats["per_party"][p]
        cur_f = cur.national["forecast"][p]
        change = {"party": p, "prob_most": cur_s["prob_most"],
                  "prob_majority": cur_s["prob_majority"],
                  "seats_mean": cur_s["mean"], "vote_mean": cur_f["mean"]}
        if prev:
            prev_s = prev.seats["per_party"].get(p, {})
            prev_f = prev.national["forecast"].get(p, {})
            change["d_prob_most"] = round(cur_s["prob_most"] - prev_s.get("prob_most", 0), 4)
            change["d_seats"] = round(cur_s["mean"] - prev_s.get("mean", 0), 1)
            change["d_vote"] = round(cur_f["mean"] - prev_f.get("mean", 0), 2)
        facts["changes"].append(change)

    since = cur.as_of - timedelta(days=days_back)
    polls = (db.query(Mo.Poll).filter(Mo.Poll.field_end > since,
                                      Mo.Poll.field_end <= cur.as_of,
                                      Mo.Poll.excluded.is_(False)).all())
    for poll in polls:
        facts["new_polls"].append({
            "pollster": poll.pollster.name, "field_end": poll.field_end.isoformat(),
            "n": poll.sample_size,
            "shares": {r.party: r.normalized_value for r in poll.results if r.party != "AUT"}})

    for ev in (db.query(Mo.NewsEvent).filter(Mo.NewsEvent.event_date > since)
               .order_by(Mo.NewsEvent.importance.desc()).limit(6)):
        facts["events"].append({"date": ev.event_date.isoformat(), "title": ev.title,
                                "kind": ev.kind, "parties": ev.parties})

    cutoff = datetime.now(timezone.utc) - timedelta(days=days_back)
    for p in parties:
        rows = (db.query(Mo.SentimentScore).join(Mo.SentimentDocument)
                .filter(Mo.SentimentScore.entity == p,
                        Mo.SentimentDocument.fetched_at >= cutoff).all())
        if rows:
            facts["sentiment"][p] = {
                "volume": len(rows),
                "moyenne": round(sum(r.sentiment for r in rows) / len(rows), 3)}
    return facts


def compose_answer(question: str, facts: dict) -> dict:
    """Réponse déterministe en français, chaque phrase traçable aux faits."""
    if "error" in facts:
        return {"answer": "Le modèle n'a pas encore produit de forecast.", "facts": facts}
    q = question.lower()
    target = next((p for p in settings.parties
                   if p.lower() in q or settings.party_names[p].lower() in q), None)
    lines = []
    changes = {c["party"]: c for c in facts["changes"]}
    subjects = [target] if target else [c["party"] for c in
                sorted(facts["changes"], key=lambda c: -abs(c.get("d_prob_most", 0)))[:2]]
    for p in subjects:
        c = changes[p]
        name = settings.party_names[p]
        lines.append(
            f"{name} : {c['prob_most']*100:.0f} % de chances de remporter le plus de "
            f"sièges ({c['seats_mean']:.0f} sièges attendus, {c['vote_mean']:.1f} % du vote).")
        if "d_prob_most" in c and facts["compared_to"]:
            verb = "a augmenté" if c["d_prob_most"] > 0 else "a diminué"
            lines.append(
                f"Depuis le {facts['compared_to']}, cette probabilité {verb} de "
                f"{abs(c['d_prob_most'])*100:.0f} point(s) "
                f"(vote {c['d_vote']:+.1f} pp, sièges {c['d_seats']:+.0f}).")
    if facts["new_polls"]:
        ps = ", ".join(f"{x['pollster']} ({x['field_end']})" for x in facts["new_polls"][:4])
        lines.append(f"Sondages intégrés récemment : {ps}.")
        if target:
            vals = [x["shares"].get(target) for x in facts["new_polls"] if x["shares"].get(target)]
            if vals:
                lines.append(f"Ces sondages placent {settings.party_names[target]} en moyenne "
                             f"à {sum(vals)/len(vals):.1f} %.")
    if facts["events"]:
        lines.append("Événements récents : "
                     + "; ".join(f"{e['title']} ({e['date']})" for e in facts["events"][:3]) + ".")
    if target and target in facts["sentiment"]:
        s = facts["sentiment"][target]
        tone = "plutôt positif" if s["moyenne"] > 0.1 else (
            "plutôt négatif" if s["moyenne"] < -0.1 else "neutre")
        lines.append(f"Signal médiatique (auxiliaire, hors modèle) : {s['volume']} mentions, "
                     f"ton {tone} ({s['moyenne']:+.2f}).")
    lines.append("Rappel : ces chiffres sont des probabilités, pas des certitudes.")
    return {"answer": " ".join(lines), "facts": facts}