Spaces:

nexusbert
/

F1

Sleeping

F1 / app.py

Update app.py

047df21 verified about 1 year ago

18.6 kB

import os
import math
import asyncio
from datetime import date, timedelta
from typing import List, Optional

from fastapi import FastAPI, Query
from pydantic import BaseModel
import httpx
from bs4 import BeautifulSoup
from dotenv import load_dotenv

# Models are blocking -> run in threadpool
load_dotenv()

# === Config & environment ===
HF_CACHE = os.getenv("HF_HOME", os.path.expanduser("~/.cache/huggingface"))
os.environ["HF_HOME"] = HF_CACHE

FOOTBALL_API_KEY = os.getenv("FOOTBALL_API_KEY", "")
FOOTBALL_ENDPOINT = "https://api.football-data.org/v4/matches"
SOCCER_LEAGUES = {"EPL": 2021, "LaLiga": 2014, "Bundesliga": 2002}

BALLDONTLIE_ENDPOINT = "https://www.balldontlie.io/api/v1/games"
ESPN_SCOREBOARD_JSON = "https://site.web.api.espn.com/apis/v2/sports/basketball/nba/scoreboard?dates={ymd}"

# concurrency
MAX_CONCURRENT_REQUESTS = int(os.getenv("MAX_CONCURRENCY", "8"))
REQUEST_TIMEOUT = 12.0

# === FastAPI app ===
app = FastAPI(title="SafeBet AI v2", description="Soccer + NBA predictions (safe/aggressive) — async & robust")

# === Lazy model loading ===
_sentiment_model = None
_reasoning_model = None
_similarity_model = None
_model_lock = asyncio.Lock()


def _load_models_blocking():
    """Blocking load of models. Called within threadpool."""
    global _sentiment_model, _reasoning_model, _similarity_model
    from transformers import pipeline
    from sentence_transformers import SentenceTransformer

    if _sentiment_model is None:
        _sentiment_model = pipeline(
            "zero-shot-classification",
            model="valhalla/distilbart-mnli-12-1",
            device=-1
        )
    if _reasoning_model is None:
        _reasoning_model = pipeline(
            "text2text-generation",
            model="google/flan-t5-base",
            device=-1
        )
    if _similarity_model is None:
        _similarity_model = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")


async def ensure_models():
    async with _model_lock:
        if _sentiment_model is None or _reasoning_model is None or _similarity_model is None:
            await asyncio.to_thread(_load_models_blocking)


# === Utilities: async http + parsing ===
sem = asyncio.Semaphore(MAX_CONCURRENT_REQUESTS)


async def fetch_text(url: str, client: httpx.AsyncClient, params=None, headers=None) -> Optional[str]:
    try:
        async with sem:
            r = await client.get(url, params=params, headers=headers, timeout=REQUEST_TIMEOUT)
            r.raise_for_status()
            return r.text
    except Exception:
        return None


async def fetch_json(url: str, client: httpx.AsyncClient, params=None, headers=None):
    try:
        async with sem:
            r = await client.get(url, params=params, headers=headers, timeout=REQUEST_TIMEOUT)
            r.raise_for_status()
            return r.json()
    except Exception:
        return None


async def parse_google_news_headlines(html_text: str, max_h=5) -> str:
    # BeautifulSoup is blocking; wrap in thread
    def _parse(html):
        soup = BeautifulSoup(html, "html.parser")
        hs = [h.get_text(separator=" ", strip=True) for h in soup.select("h3")[:max_h]]
        return " ".join(hs)
    return await asyncio.to_thread(_parse, html_text) if html_text else ""


# === Aggressive news fetch chain ===
async def aggressive_team_news(team: str, sport: str = "football") -> str:
    """
    Best procedures:
    1. Google News search
    2. Bing News search
    3. ESPN (if sport-specific)
    4. Fallback: short default text
    """
    async with httpx.AsyncClient() as client:
        # Google News
        q = f"{team} {sport}"
        url = f"https://news.google.com/search"
        params = {"q": q, "hl": "en"}
        html = await fetch_text(url, client, params=params, headers={"User-Agent": "Mozilla/5.0"})
        text = await parse_google_news_headlines(html, max_h=6)
        if text and len(text.strip()) >= 20:
            return text

        # Bing News fallback (lightweight)
        try:
            bing_url = "https://www.bing.com/news/search"
            params = {"q": q}
            html = await fetch_text(bing_url, client, params=params, headers={"User-Agent": "Mozilla/5.0"})
            if html:
                parsed = await parse_google_news_headlines(html, max_h=6)
                if parsed and len(parsed.strip()) >= 20:
                    return parsed
        except Exception:
            pass

        # ESPN headlines (sport-specific)
        try:
            if sport.lower() in ("football", "soccer"):
                # ESPN soccer site search
                search_url = f"https://www.espn.com/search/results"
                params = {"search": q}
                html = await fetch_text(search_url, client, params=params, headers={"User-Agent": "Mozilla/5.0"})
                parsed = await parse_google_news_headlines(html, max_h=6)
                if parsed and len(parsed.strip()) >= 20:
                    return parsed
            else:
                # NBA-specific ESPN scoreboard headlines
                espn_search = f"https://www.espn.com/search/results?q={q}"
                html = await fetch_text(espn_search, client, headers={"User-Agent": "Mozilla/5.0"})
                parsed = await parse_google_news_headlines(html, max_h=6)
                if parsed and len(parsed.strip()) >= 20:
                    return parsed
        except Exception:
            pass

    # final fallback
    return f"No recent updates found for {team} ({sport})."


# === Sentiment & similarity wrappers (run in threadpool) ===
async def sentiment_and_signal(text: str):
    await ensure_models()
    labels = ["positive", "negative", "injury", "motivation", "transfer"]
    # call blocking pipeline in thread
    def _call():
        try:
            res = _sentiment_model(text, labels)
            # HF zero-shot returns dict with labels & scores
            labels_out = res.get("labels", [])
            scores_out = res.get("scores", [])
            table = {lbl: float(scores_out[i]) for i, lbl in enumerate(labels_out)}
            net = table.get("positive", 0.0) - table.get("negative", 0.0) - 0.4 * table.get("injury", 0.0) + 0.2 * table.get("motivation", 0.0)
            return {**table, "net": net}
        except Exception:
            return {"positive": 0.0, "negative": 0.0, "injury": 0.0, "motivation": 0.0, "transfer": 0.0, "net": 0.0}
    return await asyncio.to_thread(_call)


async def similarity_score(a: str, b: str) -> float:
    await ensure_models()

    def _call():
        try:
            emb_a = _similarity_model.encode(a, convert_to_tensor=True)
            emb_b = _similarity_model.encode(b, convert_to_tensor=True)
            from sentence_transformers import util as st_util
            return float(st_util.pytorch_cos_sim(emb_a, emb_b).item())
        except Exception:
            return 0.0
    return await asyncio.to_thread(_call)


async def reasoning_text(context: str, sport: str, mode: str) -> str:
    await ensure_models()

    def _call():
        try:
            prompt = (
                f"Mode: {mode}. Sport: {sport}.\n"
                f"Context:\n{context}\n\n"
                "Provide a concise betting-style verdict and short safe reasoning."
            )
            out = _reasoning_model(prompt, max_new_tokens=120)
            return out[0].get("generated_text", "").strip()
        except Exception:
            return ""
    return await asyncio.to_thread(_call)


# === Odds math helpers ===
def softmax_pair(h: float, a: float):
    """Simple softmax over two values to get probabilities."""
    exph = math.exp(h)
    expa = math.exp(a)
    s = exph + expa
    if s == 0:
        return 0.5, 0.5
    return exph / s, expa / s


def apply_bet_bias(p_home: float, p_draw: float, p_away: float, mode: str):
    """Apply bias: safe increases draw/home; aggressive favors wins."""
    if mode == "safe":
        p_draw += 0.15
        p_home += 0.10
    else:
        p_home += 0.20
        p_away += 0.05
    total = max(1e-6, p_home + p_draw + p_away)
    return p_home / total, p_draw / total, p_away / total


def confidence_to_decimal_odds(conf: float, cap_min: float = 0.02):
    # avoid division by zero, cap at min confidence
    c = max(conf, cap_min)
    odds = 1.0 / c
    # sensible rounding and bounds
    if odds < 1.01:
        odds = 1.01
    if odds > 1000:
        odds = 1000.0
    return round(odds, 2)


# === Soccer prediction pipeline ===
async def predict_soccer(home: str, away: str, mode: str):
    # aggressive news fetches in parallel
    home_news_task = asyncio.create_task(aggressive_team_news(home, "football"))
    away_news_task = asyncio.create_task(aggressive_team_news(away, "football"))
    home_news, away_news = await asyncio.gather(home_news_task, away_news_task)

    home_sig_task = asyncio.create_task(sentiment_and_signal(home_news))
    away_sig_task = asyncio.create_task(sentiment_and_signal(away_news))
    sim_task = asyncio.create_task(similarity_score(home + " " + home_news, away + " " + away_news))
    home_sig, away_sig, sim = await asyncio.gather(home_sig_task, away_sig_task, sim_task)

    # core probability (home vs away)
    diff = home_sig.get("net", 0.0) - away_sig.get("net", 0.0)
    p_home = 1 / (1 + math.exp(-3 * diff))
    p_away = 1 - p_home
    p_draw = max(0.05, 1 - abs(p_home - p_away))

    p_home, p_draw, p_away = apply_bet_bias(p_home, p_draw, p_away, mode)

    # determine pick (safe heuristics)
    if mode == "safe":
        if p_draw >= max(p_home, p_away) and p_draw >= 0.35:
            pick = "Draw / Under 2.5 (safe)"
            conf = p_draw
        elif p_home >= 0.45:
            pick = f"{home} or Draw (1X)"
            conf = p_home
        else:
            pick = f"Draw or {away} (X2)"
            conf = max(p_draw, p_away)
    else:
        # aggressive: pick the higher of home/away, but give reasoning if margin small
        if abs(p_home - p_away) < 0.07:
            pick = "Small margin — consider Double Chance"
            conf = max(p_home, p_away)
        else:
            pick = home if p_home > p_away else away
            conf = max(p_home, p_away)

    # reasoning via LLM
    context = (
        f"{home} vs {away}\nhome_net={home_sig.get('net'):.3f} away_net={away_sig.get('net'):.3f}\n"
        f"similarity={sim:.3f}\nhome_news_excerpt={home_news[:200]}\naway_news_excerpt={away_news[:200]}"
    )
    reasoning = await reasoning_text(context, "football", mode)

    # implied odds
    decimal_odds = confidence_to_decimal_odds(conf)

    return {
        "match": f"{home} vs {away}",
        "pick": pick,
        "confidence": round(conf, 3),
        "decimal_odds": decimal_odds,
        "similarity": round(sim, 3),
        "reasoning": reasoning,
        "home_news": home_news[:500],
        "away_news": away_news[:500],
        "home_signal": home_sig,
        "away_signal": away_sig,
    }


# === NBA pipeline with aggressive fixture fallback ===
async def fetch_nba_fixtures_for_date(ymd: str):
    """Try: balldontlie -> ESPN JSON scoreboard -> Google search fallback (light)."""
    fixtures = []

    async with httpx.AsyncClient() as client:
        # 1) balldontlie
        try:
            params = {"dates[]": ymd}
            r = await client.get(BALLDONTLIE_ENDPOINT, params=params, timeout=REQUEST_TIMEOUT)
            if r.status_code == 200:
                j = r.json()
                data = j.get("data", [])
                for g in data:
                    home = g.get("home_team", {}).get("full_name")
                    away = g.get("visitor_team", {}).get("full_name")
                    if home and away:
                        fixtures.append({"home": home, "away": away})
        except Exception:
            pass

        if fixtures:
            return fixtures

        # 2) ESPN scoreboard JSON (dates as YYYYMMDD)
        try:
            # ESPN endpoint expects YYYYMMDD
            ymd_es = ymd.replace("-", "")
            url = ESPN_SCOREBOARD_JSON.format(ymd=ymd_es)
            j = await fetch_json(url, client)
            events = (j or {}).get("events", []) if isinstance(j, dict) else []
            for ev in events:
                comps = (ev.get("competitions"), None) or ev.get("competitions") or []
                if comps:
                    comps0 = comps[0]
                    comps_list = comps0.get("competitors") or []
                    home = next((c for c in comps_list if c.get("homeAway") == "home"), None)
                    away = next((c for c in comps_list if c.get("homeAway") == "away"), None)
                    if home and away:
                        fixtures.append({
                            "home": home.get("team", {}).get("displayName") or home.get("team", {}).get("name"),
                            "away": away.get("team", {}).get("displayName") or away.get("team", {}).get("name"),
                        })
        except Exception:
            pass

        if fixtures:
            return fixtures

        # 3) Lightweight Google search fallback for the date
        try:
            search_q = f"NBA schedule {ymd}"
            google = await fetch_text("https://www.google.com/search", client, params={"q": search_q}, headers={"User-Agent": "Mozilla/5.0"})
            parsed = await parse_google_news_headlines(google, max_h=30)
            # extract "Team A vs Team B" lines
            lines = parsed.split("  ")
            from re import findall
            for ln in lines:
                matches = findall(r"([A-Za-z .&]+) (?:v|vs|vs.|at) ([A-Za-z .&]+)", ln, flags=0)
                for m in matches:
                    fixtures.append({"home": m[0].strip(), "away": m[1].strip()})
        except Exception:
            pass

    return fixtures


async def predict_nba(home: str, away: str, mode: str):
    # aggressive news fetch
    home_news_task = asyncio.create_task(aggressive_team_news(home, "NBA"))
    away_news_task = asyncio.create_task(aggressive_team_news(away, "NBA"))
    home_news, away_news = await asyncio.gather(home_news_task, away_news_task)

    home_sig_task = asyncio.create_task(sentiment_and_signal(home_news))
    away_sig_task = asyncio.create_task(sentiment_and_signal(away_news))
    home_sig, away_sig = await asyncio.gather(home_sig_task, away_sig_task)

    # compute probabilities
    diff = home_sig.get("net", 0.0) - away_sig.get("net", 0.0)
    p_home = 1 / (1 + math.exp(-3 * diff))
    p_away = 1 - p_home
    p_draw = 0.02

    p_home, p_draw, p_away = apply_bet_bias(p_home, p_draw, p_away, mode)

    # pick
    if mode == "safe":
        # safe: lean to close margin and safe margin suggestion
        pick = f"{home} Win (safe margin < 10 pts)"
        conf = p_home if p_home > p_away else p_away
    else:
        pick = home if p_home > p_away else away
        conf = max(p_home, p_away)

    context = (
        f"{home} vs {away}\nhome_net={home_sig.get('net'):.3f} away_net={away_sig.get('net'):.3f}\n"
        f"home_news_excerpt={home_news[:200]}\naway_news_excerpt={away_news[:200]}"
    )
    reasoning = await reasoning_text(context, "NBA", mode)
    decimal_odds = confidence_to_decimal_odds(conf)
    return {
        "match": f"{home} vs {away}",
        "pick": pick,
        "confidence": round(conf, 3),
        "decimal_odds": decimal_odds,
        "reasoning": reasoning,
        "home_news": home_news[:500],
        "away_news": away_news[:500],
        "home_signal": home_sig,
        "away_signal": away_sig,
    }


# === API models & endpoints ===
class PredictionsResponse(BaseModel):
    date: str
    mode: str
    sport: str
    predictions: List[dict]
    count: int


@app.get("/predictions", response_model=PredictionsResponse)
async def predictions(
    mode: str = Query("safe", enum=["safe", "aggressive"]),
    sport: str = Query("both", enum=["soccer", "nba", "both"]),
):
    await ensure_models()

    today = date.today()
    results = []

    # soccer
    if sport in ("soccer", "both"):
        date_from = (today - timedelta(days=1)).isoformat()
        date_to = (today + timedelta(days=1)).isoformat()
        headers = {"X-Auth-Token": FOOTBALL_API_KEY} if FOOTBALL_API_KEY else {}
        async with httpx.AsyncClient() as client:
            for league, code in SOCCER_LEAGUES.items():
                try:
                    params = {"competitions": code, "dateFrom": date_from, "dateTo": date_to}
                    r = await client.get(FOOTBALL_ENDPOINT, params=params, headers=headers, timeout=REQUEST_TIMEOUT)
                    if r.status_code == 200:
                        j = r.json()
                        matches = j.get("matches", [])
                        # limit to first 3 per league for speed
                        tasks = []
                        for m in matches[:3]:
                            home = m.get("homeTeam", {}).get("name")
                            away = m.get("awayTeam", {}).get("name")
                            if home and away:
                                tasks.append(asyncio.create_task(predict_soccer(home, away, mode)))
                        if tasks:
                            res = await asyncio.gather(*tasks)
                            results.extend(res)
                except Exception as e:
                    results.append({"league": league, "error": str(e)})

    # nba
    if sport in ("nba", "both"):
        # try balldontlie and aggressive fallback
        ymd = today.isoformat()
        fixtures = await fetch_nba_fixtures_for_date(ymd)
        if fixtures:
            # predict top N (limit)
            tasks = []
            for f in fixtures[:6]:
                tasks.append(asyncio.create_task(predict_nba(f["home"], f["away"], mode)))
            res = await asyncio.gather(*tasks)
            results.extend(res)
        else:
            # nothing found
            results.append({"league": "NBA", "notice": "No fixtures found for today via API or fallbacks."})

    return PredictionsResponse(
        date=today.isoformat(),
        mode=mode,
        sport=sport,
        predictions=results,
        count=len(results)
    )


@app.get("/")
def root():
    return {
        "message": "SafeBet AI v2 — async, aggressive news, NBA fallback, safe/aggressive modes",
        "endpoints": {
            "/predictions?mode=safe|aggressive&sport=soccer|nba|both": "Main predictions endpoint"
        },
    }


# === Run guard ===
if __name__ == "__main__":
    import uvicorn
    uvicorn.run("safebet_v2:app", host="0.0.0.0", port=int(os.getenv("PORT", "7860")), reload=False)

	import os
	import math
	import asyncio
	from datetime import date, timedelta
	from typing import List, Optional

	from fastapi import FastAPI, Query
	from pydantic import BaseModel
	import httpx
	from bs4 import BeautifulSoup
	from dotenv import load_dotenv

	# Models are blocking -> run in threadpool
	load_dotenv()

	# === Config & environment ===
	HF_CACHE = os.getenv("HF_HOME", os.path.expanduser("~/.cache/huggingface"))
	os.environ["HF_HOME"] = HF_CACHE

	FOOTBALL_API_KEY = os.getenv("FOOTBALL_API_KEY", "")
	FOOTBALL_ENDPOINT = "https://api.football-data.org/v4/matches"
	SOCCER_LEAGUES = {"EPL": 2021, "LaLiga": 2014, "Bundesliga": 2002}

	BALLDONTLIE_ENDPOINT = "https://www.balldontlie.io/api/v1/games"
	ESPN_SCOREBOARD_JSON = "https://site.web.api.espn.com/apis/v2/sports/basketball/nba/scoreboard?dates={ymd}"

	# concurrency
	MAX_CONCURRENT_REQUESTS = int(os.getenv("MAX_CONCURRENCY", "8"))
	REQUEST_TIMEOUT = 12.0

	# === FastAPI app ===
	app = FastAPI(title="SafeBet AI v2", description="Soccer + NBA predictions (safe/aggressive) — async & robust")

	# === Lazy model loading ===
	_sentiment_model = None
	_reasoning_model = None
	_similarity_model = None
	_model_lock = asyncio.Lock()


	def _load_models_blocking():
	"""Blocking load of models. Called within threadpool."""
	global _sentiment_model, _reasoning_model, _similarity_model
	from transformers import pipeline
	from sentence_transformers import SentenceTransformer

	if _sentiment_model is None:
	_sentiment_model = pipeline(
	"zero-shot-classification",
	model="valhalla/distilbart-mnli-12-1",
	device=-1
	)
	if _reasoning_model is None:
	_reasoning_model = pipeline(
	"text2text-generation",
	model="google/flan-t5-base",
	device=-1
	)
	if _similarity_model is None:
	_similarity_model = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")


	async def ensure_models():
	async with _model_lock:
	if _sentiment_model is None or _reasoning_model is None or _similarity_model is None:
	await asyncio.to_thread(_load_models_blocking)


	# === Utilities: async http + parsing ===
	sem = asyncio.Semaphore(MAX_CONCURRENT_REQUESTS)


	async def fetch_text(url: str, client: httpx.AsyncClient, params=None, headers=None) -> Optional[str]:
	try:
	async with sem:
	r = await client.get(url, params=params, headers=headers, timeout=REQUEST_TIMEOUT)
	r.raise_for_status()
	return r.text
	except Exception:
	return None


	async def fetch_json(url: str, client: httpx.AsyncClient, params=None, headers=None):
	try:
	async with sem:
	r = await client.get(url, params=params, headers=headers, timeout=REQUEST_TIMEOUT)
	r.raise_for_status()
	return r.json()
	except Exception:
	return None


	async def parse_google_news_headlines(html_text: str, max_h=5) -> str:
	# BeautifulSoup is blocking; wrap in thread
	def _parse(html):
	soup = BeautifulSoup(html, "html.parser")
	hs = [h.get_text(separator=" ", strip=True) for h in soup.select("h3")[:max_h]]
	return " ".join(hs)
	return await asyncio.to_thread(_parse, html_text) if html_text else ""


	# === Aggressive news fetch chain ===
	async def aggressive_team_news(team: str, sport: str = "football") -> str:
	"""
	Best procedures:
	1. Google News search
	2. Bing News search
	3. ESPN (if sport-specific)
	4. Fallback: short default text
	"""
	async with httpx.AsyncClient() as client:
	# Google News
	q = f"{team} {sport}"
	url = f"https://news.google.com/search"
	params = {"q": q, "hl": "en"}
	html = await fetch_text(url, client, params=params, headers={"User-Agent": "Mozilla/5.0"})
	text = await parse_google_news_headlines(html, max_h=6)
	if text and len(text.strip()) >= 20:
	return text

	# Bing News fallback (lightweight)
	try:
	bing_url = "https://www.bing.com/news/search"
	params = {"q": q}
	html = await fetch_text(bing_url, client, params=params, headers={"User-Agent": "Mozilla/5.0"})
	if html:
	parsed = await parse_google_news_headlines(html, max_h=6)
	if parsed and len(parsed.strip()) >= 20:
	return parsed
	except Exception:
	pass

	# ESPN headlines (sport-specific)
	try:
	if sport.lower() in ("football", "soccer"):
	# ESPN soccer site search
	search_url = f"https://www.espn.com/search/results"
	params = {"search": q}
	html = await fetch_text(search_url, client, params=params, headers={"User-Agent": "Mozilla/5.0"})
	parsed = await parse_google_news_headlines(html, max_h=6)
	if parsed and len(parsed.strip()) >= 20:
	return parsed
	else:
	# NBA-specific ESPN scoreboard headlines
	espn_search = f"https://www.espn.com/search/results?q={q}"
	html = await fetch_text(espn_search, client, headers={"User-Agent": "Mozilla/5.0"})
	parsed = await parse_google_news_headlines(html, max_h=6)
	if parsed and len(parsed.strip()) >= 20:
	return parsed
	except Exception:
	pass

	# final fallback
	return f"No recent updates found for {team} ({sport})."


	# === Sentiment & similarity wrappers (run in threadpool) ===
	async def sentiment_and_signal(text: str):
	await ensure_models()
	labels = ["positive", "negative", "injury", "motivation", "transfer"]
	# call blocking pipeline in thread
	def _call():
	try:
	res = _sentiment_model(text, labels)
	# HF zero-shot returns dict with labels & scores
	labels_out = res.get("labels", [])
	scores_out = res.get("scores", [])
	table = {lbl: float(scores_out[i]) for i, lbl in enumerate(labels_out)}
	net = table.get("positive", 0.0) - table.get("negative", 0.0) - 0.4 * table.get("injury", 0.0) + 0.2 * table.get("motivation", 0.0)
	return {**table, "net": net}
	except Exception:
	return {"positive": 0.0, "negative": 0.0, "injury": 0.0, "motivation": 0.0, "transfer": 0.0, "net": 0.0}
	return await asyncio.to_thread(_call)


	async def similarity_score(a: str, b: str) -> float:
	await ensure_models()

	def _call():
	try:
	emb_a = _similarity_model.encode(a, convert_to_tensor=True)
	emb_b = _similarity_model.encode(b, convert_to_tensor=True)
	from sentence_transformers import util as st_util
	return float(st_util.pytorch_cos_sim(emb_a, emb_b).item())
	except Exception:
	return 0.0
	return await asyncio.to_thread(_call)


	async def reasoning_text(context: str, sport: str, mode: str) -> str:
	await ensure_models()

	def _call():
	try:
	prompt = (
	f"Mode: {mode}. Sport: {sport}.\n"
	f"Context:\n{context}\n\n"
	"Provide a concise betting-style verdict and short safe reasoning."
	)
	out = _reasoning_model(prompt, max_new_tokens=120)
	return out[0].get("generated_text", "").strip()
	except Exception:
	return ""
	return await asyncio.to_thread(_call)


	# === Odds math helpers ===
	def softmax_pair(h: float, a: float):
	"""Simple softmax over two values to get probabilities."""
	exph = math.exp(h)
	expa = math.exp(a)
	s = exph + expa
	if s == 0:
	return 0.5, 0.5
	return exph / s, expa / s


	def apply_bet_bias(p_home: float, p_draw: float, p_away: float, mode: str):
	"""Apply bias: safe increases draw/home; aggressive favors wins."""
	if mode == "safe":
	p_draw += 0.15
	p_home += 0.10
	else:
	p_home += 0.20
	p_away += 0.05
	total = max(1e-6, p_home + p_draw + p_away)
	return p_home / total, p_draw / total, p_away / total


	def confidence_to_decimal_odds(conf: float, cap_min: float = 0.02):
	# avoid division by zero, cap at min confidence
	c = max(conf, cap_min)
	odds = 1.0 / c
	# sensible rounding and bounds
	if odds < 1.01:
	odds = 1.01
	if odds > 1000:
	odds = 1000.0
	return round(odds, 2)


	# === Soccer prediction pipeline ===
	async def predict_soccer(home: str, away: str, mode: str):
	# aggressive news fetches in parallel
	home_news_task = asyncio.create_task(aggressive_team_news(home, "football"))
	away_news_task = asyncio.create_task(aggressive_team_news(away, "football"))
	home_news, away_news = await asyncio.gather(home_news_task, away_news_task)

	home_sig_task = asyncio.create_task(sentiment_and_signal(home_news))
	away_sig_task = asyncio.create_task(sentiment_and_signal(away_news))
	sim_task = asyncio.create_task(similarity_score(home + " " + home_news, away + " " + away_news))
	home_sig, away_sig, sim = await asyncio.gather(home_sig_task, away_sig_task, sim_task)

	# core probability (home vs away)
	diff = home_sig.get("net", 0.0) - away_sig.get("net", 0.0)
	p_home = 1 / (1 + math.exp(-3 * diff))
	p_away = 1 - p_home
	p_draw = max(0.05, 1 - abs(p_home - p_away))

	p_home, p_draw, p_away = apply_bet_bias(p_home, p_draw, p_away, mode)

	# determine pick (safe heuristics)
	if mode == "safe":
	if p_draw >= max(p_home, p_away) and p_draw >= 0.35:
	pick = "Draw / Under 2.5 (safe)"
	conf = p_draw
	elif p_home >= 0.45:
	pick = f"{home} or Draw (1X)"
	conf = p_home
	else:
	pick = f"Draw or {away} (X2)"
	conf = max(p_draw, p_away)
	else:
	# aggressive: pick the higher of home/away, but give reasoning if margin small
	if abs(p_home - p_away) < 0.07:
	pick = "Small margin — consider Double Chance"
	conf = max(p_home, p_away)
	else:
	pick = home if p_home > p_away else away
	conf = max(p_home, p_away)

	# reasoning via LLM
	context = (
	f"{home} vs {away}\nhome_net={home_sig.get('net'):.3f} away_net={away_sig.get('net'):.3f}\n"
	f"similarity={sim:.3f}\nhome_news_excerpt={home_news[:200]}\naway_news_excerpt={away_news[:200]}"
	)
	reasoning = await reasoning_text(context, "football", mode)

	# implied odds
	decimal_odds = confidence_to_decimal_odds(conf)

	return {
	"match": f"{home} vs {away}",
	"pick": pick,
	"confidence": round(conf, 3),
	"decimal_odds": decimal_odds,
	"similarity": round(sim, 3),
	"reasoning": reasoning,
	"home_news": home_news[:500],
	"away_news": away_news[:500],
	"home_signal": home_sig,
	"away_signal": away_sig,
	}


	# === NBA pipeline with aggressive fixture fallback ===
	async def fetch_nba_fixtures_for_date(ymd: str):
	"""Try: balldontlie -> ESPN JSON scoreboard -> Google search fallback (light)."""
	fixtures = []

	async with httpx.AsyncClient() as client:
	# 1) balldontlie
	try:
	params = {"dates[]": ymd}
	r = await client.get(BALLDONTLIE_ENDPOINT, params=params, timeout=REQUEST_TIMEOUT)
	if r.status_code == 200:
	j = r.json()
	data = j.get("data", [])
	for g in data:
	home = g.get("home_team", {}).get("full_name")
	away = g.get("visitor_team", {}).get("full_name")
	if home and away:
	fixtures.append({"home": home, "away": away})
	except Exception:
	pass

	if fixtures:
	return fixtures

	# 2) ESPN scoreboard JSON (dates as YYYYMMDD)
	try:
	# ESPN endpoint expects YYYYMMDD
	ymd_es = ymd.replace("-", "")
	url = ESPN_SCOREBOARD_JSON.format(ymd=ymd_es)
	j = await fetch_json(url, client)
	events = (j or {}).get("events", []) if isinstance(j, dict) else []
	for ev in events:
	comps = (ev.get("competitions"), None) or ev.get("competitions") or []
	if comps:
	comps0 = comps[0]
	comps_list = comps0.get("competitors") or []
	home = next((c for c in comps_list if c.get("homeAway") == "home"), None)
	away = next((c for c in comps_list if c.get("homeAway") == "away"), None)
	if home and away:
	fixtures.append({
	"home": home.get("team", {}).get("displayName") or home.get("team", {}).get("name"),
	"away": away.get("team", {}).get("displayName") or away.get("team", {}).get("name"),
	})
	except Exception:
	pass

	if fixtures:
	return fixtures

	# 3) Lightweight Google search fallback for the date
	try:
	search_q = f"NBA schedule {ymd}"
	google = await fetch_text("https://www.google.com/search", client, params={"q": search_q}, headers={"User-Agent": "Mozilla/5.0"})
	parsed = await parse_google_news_headlines(google, max_h=30)
	# extract "Team A vs Team B" lines
	lines = parsed.split(" ")
	from re import findall
	for ln in lines:
	matches = findall(r"([A-Za-z .&]+) (?:v\|vs\|vs.\|at) ([A-Za-z .&]+)", ln, flags=0)
	for m in matches:
	fixtures.append({"home": m[0].strip(), "away": m[1].strip()})
	except Exception:
	pass

	return fixtures


	async def predict_nba(home: str, away: str, mode: str):
	# aggressive news fetch
	home_news_task = asyncio.create_task(aggressive_team_news(home, "NBA"))
	away_news_task = asyncio.create_task(aggressive_team_news(away, "NBA"))
	home_news, away_news = await asyncio.gather(home_news_task, away_news_task)

	home_sig_task = asyncio.create_task(sentiment_and_signal(home_news))
	away_sig_task = asyncio.create_task(sentiment_and_signal(away_news))
	home_sig, away_sig = await asyncio.gather(home_sig_task, away_sig_task)

	# compute probabilities
	diff = home_sig.get("net", 0.0) - away_sig.get("net", 0.0)
	p_home = 1 / (1 + math.exp(-3 * diff))
	p_away = 1 - p_home
	p_draw = 0.02

	p_home, p_draw, p_away = apply_bet_bias(p_home, p_draw, p_away, mode)

	# pick
	if mode == "safe":
	# safe: lean to close margin and safe margin suggestion
	pick = f"{home} Win (safe margin < 10 pts)"
	conf = p_home if p_home > p_away else p_away
	else:
	pick = home if p_home > p_away else away
	conf = max(p_home, p_away)

	context = (
	f"{home} vs {away}\nhome_net={home_sig.get('net'):.3f} away_net={away_sig.get('net'):.3f}\n"
	f"home_news_excerpt={home_news[:200]}\naway_news_excerpt={away_news[:200]}"
	)
	reasoning = await reasoning_text(context, "NBA", mode)
	decimal_odds = confidence_to_decimal_odds(conf)
	return {
	"match": f"{home} vs {away}",
	"pick": pick,
	"confidence": round(conf, 3),
	"decimal_odds": decimal_odds,
	"reasoning": reasoning,
	"home_news": home_news[:500],
	"away_news": away_news[:500],
	"home_signal": home_sig,
	"away_signal": away_sig,
	}


	# === API models & endpoints ===
	class PredictionsResponse(BaseModel):
	date: str
	mode: str
	sport: str
	predictions: List[dict]
	count: int


	@app.get("/predictions", response_model=PredictionsResponse)
	async def predictions(
	mode: str = Query("safe", enum=["safe", "aggressive"]),
	sport: str = Query("both", enum=["soccer", "nba", "both"]),
	):
	await ensure_models()

	today = date.today()
	results = []

	# soccer
	if sport in ("soccer", "both"):
	date_from = (today - timedelta(days=1)).isoformat()
	date_to = (today + timedelta(days=1)).isoformat()
	headers = {"X-Auth-Token": FOOTBALL_API_KEY} if FOOTBALL_API_KEY else {}
	async with httpx.AsyncClient() as client:
	for league, code in SOCCER_LEAGUES.items():
	try:
	params = {"competitions": code, "dateFrom": date_from, "dateTo": date_to}
	r = await client.get(FOOTBALL_ENDPOINT, params=params, headers=headers, timeout=REQUEST_TIMEOUT)
	if r.status_code == 200:
	j = r.json()
	matches = j.get("matches", [])
	# limit to first 3 per league for speed
	tasks = []
	for m in matches[:3]:
	home = m.get("homeTeam", {}).get("name")
	away = m.get("awayTeam", {}).get("name")
	if home and away:
	tasks.append(asyncio.create_task(predict_soccer(home, away, mode)))
	if tasks:
	res = await asyncio.gather(*tasks)
	results.extend(res)
	except Exception as e:
	results.append({"league": league, "error": str(e)})

	# nba
	if sport in ("nba", "both"):
	# try balldontlie and aggressive fallback
	ymd = today.isoformat()
	fixtures = await fetch_nba_fixtures_for_date(ymd)
	if fixtures:
	# predict top N (limit)
	tasks = []
	for f in fixtures[:6]:
	tasks.append(asyncio.create_task(predict_nba(f["home"], f["away"], mode)))
	res = await asyncio.gather(*tasks)
	results.extend(res)
	else:
	# nothing found
	results.append({"league": "NBA", "notice": "No fixtures found for today via API or fallbacks."})

	return PredictionsResponse(
	date=today.isoformat(),
	mode=mode,
	sport=sport,
	predictions=results,
	count=len(results)
	)


	@app.get("/")
	def root():
	return {
	"message": "SafeBet AI v2 — async, aggressive news, NBA fallback, safe/aggressive modes",
	"endpoints": {
	"/predictions?mode=safe\|aggressive&sport=soccer\|nba\|both": "Main predictions endpoint"
	},
	}


	# === Run guard ===
	if __name__ == "__main__":
	import uvicorn
	uvicorn.run("safebet_v2:app", host="0.0.0.0", port=int(os.getenv("PORT", "7860")), reload=False)