backend/app/ai/intelligence.py

"""
AI Intelligence Engine — three high-value features:
  1. Financial Coach Mode  — proactive weekly recommendations, nudges, anomaly explanations
  2. Fraud Explanation     — human-readable AI explanation of why a transaction was flagged
  3. Spending Narrative    — monthly story: what changed, what improved, what to watch
"""
import os
from datetime import datetime, timedelta
from collections import defaultdict
from sqlalchemy.orm import Session

from app.database.models import (
    User, Account, Transaction, Goal, Investment,
    Subscription, FraudLog, Notification
)
from app.ai.coaching import calculate_financial_health_score
from app.ai.forecasting import get_cashflow_metrics


# ─── Shared Groq/OpenAI caller ────────────────────────────────────────────────

def _call_llm(messages: list, max_tokens: int = 500) -> str | None:
    """
    Calls OpenAI → Groq in priority order with the exact messages list provided.
    Returns None if both fail.
    """
    openai_key = os.environ.get("OPENAI_API_KEY", "")
    groq_key = os.environ.get("GROQ_API_KEY", "") or os.environ.get("GROQ_KEY", "")

    if openai_key:
        try:
            from openai import OpenAI
            client = OpenAI(api_key=openai_key)
            res = client.chat.completions.create(
                model="gpt-4o-mini",
                messages=messages,
                temperature=0.2,
                max_tokens=max_tokens,
            )
            return res.choices[0].message.content.strip()
        except Exception as e:
            print(f"[intelligence] OpenAI error: {e}")

    if groq_key:
        try:
            from groq import Groq
            client = Groq(api_key=groq_key)
            res = client.chat.completions.create(
                model="llama-3.3-70b-versatile",
                messages=messages,
                temperature=0.2,
                max_tokens=max_tokens,
            )
            return res.choices[0].message.content.strip()
        except Exception as e:
            print(f"[intelligence] Groq error: {e}")

    return None


# ─── 1. FINANCIAL COACH MODE ──────────────────────────────────────────────────

def generate_weekly_coaching(db: Session, user_id: str) -> dict:
    """
    Generates a proactive weekly coaching report with:
    - Weekly summary (what happened this week)
    - Budget coaching (where money went vs targets)
    - Savings nudge (specific, actionable)
    - Anomaly explanations (unusual patterns this week)
    - Top 3 recommendations
    """
    user = db.query(User).filter(User.id == user_id).first()
    if not user:
        return {"error": "User not found"}

    # ── Gather this week's data ───────────────────────────────────────────────
    now = datetime.utcnow()
    week_start = now - timedelta(days=7)

    accounts = db.query(Account).filter(Account.user_id == user_id).all()
    account_ids = [a.id for a in accounts]
    total_balance = sum(a.balance for a in accounts)
    savings_balance = sum(a.balance for a in accounts if a.type == "savings")

    week_txns = (
        db.query(Transaction)
        .filter(
            Transaction.account_id.in_(account_ids),
            Transaction.timestamp >= week_start,
        )
        .all()
    ) if account_ids else []

    week_spend = sum(t.amount for t in week_txns if t.type == "debit")
    week_income = sum(t.amount for t in week_txns if t.type == "credit")

    # Category breakdown this week
    cat_spend: dict = defaultdict(float)
    for t in week_txns:
        if t.type == "debit":
            cat_spend[t.category or "Other"] += t.amount
    top_cats = sorted(cat_spend.items(), key=lambda x: x[1], reverse=True)[:5]

    # Compare to prior week
    prior_start = week_start - timedelta(days=7)
    prior_txns = (
        db.query(Transaction)
        .filter(
            Transaction.account_id.in_(account_ids),
            Transaction.timestamp >= prior_start,
            Transaction.timestamp < week_start,
        )
        .all()
    ) if account_ids else []
    prior_spend = sum(t.amount for t in prior_txns if t.type == "debit")
    spend_delta_pct = ((week_spend - prior_spend) / max(prior_spend, 1)) * 100

    # Goals progress
    goals = db.query(Goal).filter(Goal.user_id == user_id).all()
    goals_summary = [
        f"{g.title}: ${g.current_amount:,.0f}/${g.target_amount:,.0f} "
        f"({g.current_amount/max(g.target_amount,1)*100:.0f}%)"
        for g in goals
    ]

    # Health score
    score_data = calculate_financial_health_score(db, user_id)
    health_score = score_data.get("overall_score", 50)
    improvements = score_data.get("actionable_improvements", [])

    # Anomalies this week
    anomalies = []
    if week_txns:
        all_amounts = [t.amount for t in week_txns if t.type == "debit"]
        if all_amounts:
            import numpy as np
            avg = float(np.mean(all_amounts))
            for t in week_txns:
                if t.type == "debit" and t.amount > avg * 3:
                    anomalies.append(
                        f"${t.amount:,.2f} at {t.merchant or 'Unknown'} "
                        f"({(t.amount/avg):.1f}× your weekly average)"
                    )

    # Late-night count
    late_night = [
        t for t in week_txns
        if t.type == "debit" and (t.timestamp.hour >= 23 or t.timestamp.hour < 4)
    ]

    # ── Build LLM prompt ──────────────────────────────────────────────────────
    system = (
        "You are a personal AI financial coach. You have access to the user's real financial data. "
        "Be specific, use exact numbers, and give actionable advice. "
        "Never be generic. Format your response with clear sections."
    )

    user_prompt = f"""Generate a weekly financial coaching report for {user.profile_data.get('name', 'the user')}.

FINANCIAL DATA:
- Total Balance: ${total_balance:,.2f} (Savings: ${savings_balance:,.2f})
- Health Score: {health_score:.0f}/100
- This Week Spending: ${week_spend:,.2f} ({spend_delta_pct:+.1f}% vs last week)
- This Week Income: ${week_income:,.2f}
- Top Spending Categories: {', '.join(f'{c}: ${a:,.0f}' for c, a in top_cats)}
- Goals: {'; '.join(goals_summary) if goals_summary else 'None set'}
- Anomalies: {'; '.join(anomalies) if anomalies else 'None detected'}
- Late-night transactions: {len(late_night)}
- Key improvements needed: {'; '.join(improvements[:2]) if improvements else 'None'}

Write a coaching report with these exact sections:
1. WEEKLY SUMMARY (2 sentences, use exact dollar figures)
2. BUDGET COACHING (what's on track, what's overspent — use real category names and amounts)
3. SAVINGS NUDGE (one specific, actionable savings recommendation with a dollar target)
4. ANOMALY ALERT (explain any unusual spending in plain English, or say "No anomalies this week")
5. TOP 3 ACTIONS (numbered list, each under 15 words, specific and actionable)

Be direct. No filler phrases."""

    messages = [
        {"role": "system", "content": system},
        {"role": "user", "content": user_prompt},
    ]

    coaching_text = _call_llm(messages, max_tokens=600)

    if not coaching_text:
        # Rule-based fallback
        direction = "up" if spend_delta_pct > 0 else "down"
        coaching_text = (
            f"WEEKLY SUMMARY\nYou spent ${week_spend:,.2f} this week, "
            f"{abs(spend_delta_pct):.0f}% {direction} from last week. "
            f"Your balance stands at ${total_balance:,.2f}.\n\n"
            f"BUDGET COACHING\nTop category: {top_cats[0][0] if top_cats else 'N/A'} "
            f"(${top_cats[0][1]:,.2f} if top_cats else '$0').\n\n"
            f"SAVINGS NUDGE\nTransfer ${min(week_income * 0.1, 200):,.0f} to savings this week.\n\n"
            f"ANOMALY ALERT\n{'Unusual: ' + anomalies[0] if anomalies else 'No anomalies this week.'}\n\n"
            f"TOP 3 ACTIONS\n"
            + "\n".join(f"{i+1}. {imp}" for i, imp in enumerate(improvements[:3]))
        )

    return {
        "generated_at": now.isoformat(),
        "week_start": week_start.date().isoformat(),
        "week_end": now.date().isoformat(),
        "health_score": health_score,
        "week_spend": round(week_spend, 2),
        "week_income": round(week_income, 2),
        "spend_delta_pct": round(spend_delta_pct, 1),
        "top_categories": [{"name": c, "amount": round(a, 2)} for c, a in top_cats],
        "anomalies": anomalies,
        "coaching_report": coaching_text,
        "improvements": improvements[:3],
    }


# ─── 2. AI FRAUD EXPLANATION ──────────────────────────────────────────────────

def explain_fraud_alert(db: Session, fraud_log_id: str) -> dict:
    """
    Generates a human-readable AI explanation of exactly why a transaction was flagged,
    with context about the user's normal patterns.
    """
    fraud_log = db.query(FraudLog).filter(FraudLog.id == fraud_log_id).first()
    if not fraud_log:
        return {"error": "Fraud log not found"}

    txn = fraud_log.transaction
    if not txn:
        return {"error": "Transaction not found"}

    account = db.query(Account).filter(Account.id == txn.account_id).first()
    if not account:
        return {"error": "Account not found"}

    user_id = account.user_id

    # Get user's historical baseline
    history = (
        db.query(Transaction)
        .filter(
            Transaction.account_id == txn.account_id,
            Transaction.type == "debit",
            Transaction.id != txn.id,
        )
        .order_by(Transaction.timestamp.desc())
        .limit(30)
        .all()
    )

    import numpy as np
    avg_amount = float(np.mean([t.amount for t in history])) if history else 0
    typical_hour = "daytime"
    if txn.timestamp.hour >= 23 or txn.timestamp.hour < 4:
        typical_hour = f"{txn.timestamp.strftime('%I:%M %p')} (late night)"
    else:
        typical_hour = txn.timestamp.strftime("%I:%M %p")

    risk_pct = round(fraud_log.risk_score * 100)
    raw_reasons = fraud_log.suspicious_activity_details or ""

    system = (
        "You are a fraud analyst AI. Explain in plain English why this transaction was flagged. "
        "Be specific, cite exact numbers, and help the user understand the risk. "
        "Write 2-3 sentences maximum. Do not use bullet points."
    )

    user_prompt = f"""Explain why this transaction was flagged as suspicious:

Transaction: ${txn.amount:,.2f} at {txn.merchant or 'Unknown merchant'}
Time: {typical_hour}
Risk Score: {risk_pct}/100
Detection reasons: {raw_reasons}
User's average transaction: ${avg_amount:,.2f}
Amount vs average: {txn.amount/max(avg_amount,1):.1f}×

Write a clear, human-readable explanation of why this was flagged. 
Start with "This payment was flagged because..." and be specific."""

    messages = [
        {"role": "system", "content": system},
        {"role": "user", "content": user_prompt},
    ]

    explanation = _call_llm(messages, max_tokens=150)

    if not explanation:
        # Rule-based fallback
        parts = []
        if txn.amount > avg_amount * 2:
            parts.append(
                f"the amount (${txn.amount:,.2f}) is {txn.amount/max(avg_amount,1):.1f}× "
                f"your typical transaction of ${avg_amount:,.2f}"
            )
        if txn.timestamp.hour >= 23 or txn.timestamp.hour < 4:
            parts.append(f"it occurred at {txn.timestamp.strftime('%I:%M %p')} (unusual hours)")
        if not parts:
            parts = [raw_reasons or "multiple anomaly signals were detected simultaneously"]
        explanation = f"This payment was flagged because {' and '.join(parts)}."

    return {
        "fraud_log_id": fraud_log_id,
        "transaction_id": txn.id,
        "merchant": txn.merchant,
        "amount": txn.amount,
        "timestamp": txn.timestamp.isoformat(),
        "risk_score": risk_pct,
        "raw_reasons": raw_reasons.split("; ") if raw_reasons else [],
        "ai_explanation": explanation,
        "user_avg_amount": round(avg_amount, 2),
        "amount_vs_avg": round(txn.amount / max(avg_amount, 1), 1),
    }


# ─── 3. SPENDING NARRATIVE ────────────────────────────────────────────────────

def generate_spending_narrative(db: Session, user_id: str) -> dict:
    """
    Generates a monthly spending narrative:
    - What changed vs last month (category-level)
    - Investment performance insight
    - Savings performance
    - Subscription waste analysis
    - One-paragraph human story of the month
    """
    user = db.query(User).filter(User.id == user_id).first()
    if not user:
        return {"error": "User not found"}

    now = datetime.utcnow()
    month_start = now.replace(day=1, hour=0, minute=0, second=0, microsecond=0)
    last_month_start = (month_start - timedelta(days=1)).replace(day=1)

    accounts = db.query(Account).filter(Account.user_id == user_id).all()
    account_ids = [a.id for a in accounts]

    def get_category_spend(start, end):
        txns = (
            db.query(Transaction)
            .filter(
                Transaction.account_id.in_(account_ids),
                Transaction.type == "debit",
                Transaction.timestamp >= start,
                Transaction.timestamp < end,
            )
            .all()
        ) if account_ids else []
        cats: dict = defaultdict(float)
        for t in txns:
            cats[t.category or "Other"] += t.amount
        return dict(cats), txns

    this_cats, this_txns = get_category_spend(month_start, now)
    last_cats, last_txns = get_category_spend(last_month_start, month_start)

    this_total = sum(this_cats.values())
    last_total = sum(last_cats.values())
    total_delta_pct = ((this_total - last_total) / max(last_total, 1)) * 100

    # Category changes
    all_cats = set(this_cats) | set(last_cats)
    cat_changes = []
    for cat in all_cats:
        this_amt = this_cats.get(cat, 0)
        last_amt = last_cats.get(cat, 0)
        if last_amt > 0:
            delta_pct = ((this_amt - last_amt) / last_amt) * 100
            cat_changes.append({
                "category": cat,
                "this_month": round(this_amt, 2),
                "last_month": round(last_amt, 2),
                "delta_pct": round(delta_pct, 1),
            })
    cat_changes.sort(key=lambda x: abs(x["delta_pct"]), reverse=True)

    # Investment performance
    investments = db.query(Investment).filter(Investment.user_id == user_id).all()
    inv_total_invested = sum(i.amount_invested for i in investments)
    inv_total_value = sum(i.current_value for i in investments)
    inv_gain = inv_total_value - inv_total_invested
    inv_gain_pct = (inv_gain / max(inv_total_invested, 1)) * 100

    # Savings
    savings_acct = next((a for a in accounts if a.type == "savings"), None)
    savings_balance = savings_acct.balance if savings_acct else 0

    # Subscriptions
    subs = db.query(Subscription).filter(
        Subscription.user_id == user_id, Subscription.active == True
    ).all()
    monthly_sub_cost = sum(
        s.amount if s.billing_cycle == "monthly" else s.amount / 12
        for s in subs
    )

    # Income this month
    this_income = sum(
        t.amount for t in this_txns
        if t.type == "credit"
    ) if this_txns else 0
    # Also check credit transactions
    income_txns = (
        db.query(Transaction)
        .filter(
            Transaction.account_id.in_(account_ids),
            Transaction.type == "credit",
            Transaction.timestamp >= month_start,
        )
        .all()
    ) if account_ids else []
    this_income = sum(t.amount for t in income_txns)

    # ── Build narrative prompt ────────────────────────────────────────────────
    top_changes = cat_changes[:4]
    changes_text = "\n".join(
        f"  - {c['category']}: ${c['this_month']:,.0f} this month vs ${c['last_month']:,.0f} last month "
        f"({c['delta_pct']:+.0f}%)"
        for c in top_changes
    )

    system = (
        "You are a personal finance narrator. Write in a warm, intelligent, first-person style "
        "as if you're a financial advisor summarizing the month for your client. "
        "Use exact numbers. Be insightful, not generic."
    )

    user_prompt = f"""Write a monthly financial narrative for {user.profile_data.get('name', 'the user')}.

THIS MONTH'S DATA:
- Total Spending: ${this_total:,.2f} ({total_delta_pct:+.1f}% vs last month)
- Total Income: ${this_income:,.2f}
- Savings Balance: ${savings_balance:,.2f}
- Investment Portfolio: ${inv_total_value:,.2f} ({inv_gain_pct:+.1f}% return)
- Monthly Subscriptions: ${monthly_sub_cost:,.2f}/month ({len(subs)} active)

CATEGORY CHANGES:
{changes_text if changes_text else '  - No prior month data available'}

Write 4 short sections:
1. MONTH IN REVIEW (1-2 sentences: overall spending story, use exact figures)
2. WHAT IMPROVED (1 sentence: best positive change with exact numbers)
3. WATCH OUT (1 sentence: biggest concern or overspend with exact numbers)
4. NEXT MONTH GOAL (1 sentence: one specific, measurable target)

Keep each section to 1-2 sentences. Use real numbers throughout."""

    messages = [
        {"role": "system", "content": system},
        {"role": "user", "content": user_prompt},
    ]

    narrative = _call_llm(messages, max_tokens=400)

    if not narrative:
        direction = "increased" if total_delta_pct > 0 else "decreased"
        best_cat = min(cat_changes, key=lambda x: x["delta_pct"], default=None)
        worst_cat = max(cat_changes, key=lambda x: x["delta_pct"], default=None)
        narrative = (
            f"MONTH IN REVIEW\nYour spending {direction} by {abs(total_delta_pct):.0f}% "
            f"to ${this_total:,.2f} this month.\n\n"
            f"WHAT IMPROVED\n"
            + (f"{best_cat['category']} spending dropped {abs(best_cat['delta_pct']):.0f}%." if best_cat and best_cat['delta_pct'] < 0 else "Spending patterns are stable.")
            + f"\n\nWATCH OUT\n"
            + (f"{worst_cat['category']} increased {worst_cat['delta_pct']:.0f}% — review this category." if worst_cat and worst_cat['delta_pct'] > 10 else "No major overspends detected.")
            + f"\n\nNEXT MONTH GOAL\nAim to keep total spending under ${this_total * 0.95:,.0f}."
        )

    return {
        "month": month_start.strftime("%B %Y"),
        "generated_at": now.isoformat(),
        "summary": {
            "total_spend": round(this_total, 2),
            "total_income": round(this_income, 2),
            "spend_delta_pct": round(total_delta_pct, 1),
            "savings_balance": round(savings_balance, 2),
            "investment_value": round(inv_total_value, 2),
            "investment_gain_pct": round(inv_gain_pct, 1),
            "monthly_subscriptions": round(monthly_sub_cost, 2),
        },
        "category_changes": cat_changes[:6],
        "narrative": narrative,
    }