inference_local.py

import json
import unicodedata
from llama_cpp import Llama

# ─────────────────────────────────────────────
# CONFIG
# ─────────────────────────────────────────────
GGUF_PATH = "./football-extractor-q4.gguf"

SYSTEM_PROMPT = (
    "You are a football data extraction assistant. "
    "Extract structured data from the message and return ONLY a valid JSON array. "
    "Each object in the array must have exactly these keys: "
    "league, team_1, team_2, prediction, date, odds. "
    "If a field is missing, use null. No extra text, no markdown."
)

# ─────────────────────────────────────────────
# LOAD MODEL (runs on Mac Metal / CPU)
# ─────────────────────────────────────────────
llm = Llama(
    model_path=GGUF_PATH,
    n_ctx=2048,        # context window
    n_gpu_layers=-1,   # offload all layers to Metal GPU
    verbose=False,
)
print("✅ Model loaded")

# ─────────────────────────────────────────────
# HELPERS
# ─────────────────────────────────────────────
def clean_input(text: str) -> str:
    """Strip bold unicode characters (e.g. Telegram bold)."""
    return ''.join(
        c for c in unicodedata.normalize('NFKD', text)
        if not unicodedata.combining(c)
    )

def fix_keys(results: list) -> list:
    """Fix 'match' key → team_1 / team_2 if model returns it."""
    for item in results:
        if "match" in item and "team_1" not in item:
            parts = item.pop("match").split(" - ", 1)
            item["team_1"] = parts[0].strip() if len(parts) > 0 else None
            item["team_2"] = parts[1].strip() if len(parts) > 1 else None
    return results

def normalize(result: list) -> list:
    keys = ["league", "team_1", "team_2", "prediction", "date", "odds"]
    if result and not isinstance(result[0], (dict, list)):
        return [dict(zip(keys, result))]
    normalized = []
    for item in result:
        if isinstance(item, str):
            try:
                item = json.loads(item)
            except:
                continue
        if isinstance(item, list):
            item = dict(zip(keys, item))
        if isinstance(item, dict):
            normalized.append(item)
    return normalized

# ─────────────────────────────────────────────
# INFERENCE
# ─────────────────────────────────────────────
def extract(text: str, debug: bool = False) -> list:
    text = clean_input(text)

    response = llm.create_chat_completion(
        messages=[
            {"role": "system", "content": SYSTEM_PROMPT},
            {"role": "user",   "content": text},
        ],
        temperature=0.0,
        max_tokens=512,
        stop=["<|im_end|>", "<|endoftext|>"],
    )

    raw = response["choices"][0]["message"]["content"].strip()

    if debug:
        print(f"[raw] {repr(raw)}")

    try:
        result = json.loads(raw)
        result = normalize(result if isinstance(result, list) else [result])
        result = fix_keys(result)
        return result
    except json.JSONDecodeError:
        print(f"[!] Could not parse JSON:\n{raw}")
        return []

# ─────────────────────────────────────────────
# TEST
# ─────────────────────────────────────────────
if __name__ == "__main__":
    tests = [
        # single tip
        """⚽️ Prediction of the Day ⚽️
Date: 24/03/2026
League: Eerste divisie Netherlands
Match: FC Emmen - SC Cambuur
Kick off: 20:00 WAT
✅Over 1.5
✅Odds @1.13 on BETANO""",

        # multi tip real format
        """⚽️ 𝐏𝐫𝐞𝐝𝐢𝐜𝐭𝐢𝐨𝐧 𝐨𝐟 𝐭𝐡𝐞 𝐃𝐚𝐲 ⚽️
𝐃𝐚𝐭𝐞: 24/03/2026
𝐋𝐞𝐚𝐠𝐮𝐞: League 1 England
𝐌𝐚𝐭𝐜𝐡: Doncaster Rovers - Port Vale
𝐊𝐢𝐜𝐤 𝐨𝐟𝐟: 20:45 WAT
✅Under 3.5
✅Odds @1.36 on BETANO
⚽️ 𝗙𝗼𝗼𝘁𝗯𝗮𝗹𝗹 𝗧𝗶𝗽 𝟮 ⚽️
𝐃𝐚𝐭𝐞: 24/03/2026
𝐋𝐞𝐚𝐠𝐮𝐞: La Liga
𝐌𝐚𝐭𝐜𝐡: Real Madrid - Barcelona
𝐊𝐢𝐜𝐤 𝐨𝐟𝐟: 21:00 WAT
✅1X
✅Odds @1.42 on BETANO""",

        # noisy missing date
        """wow predictions
MATCH: Juventus VS Napoli
League: Serie A
we forecast Over 2.5
Odds 1.75""",
    ]

    for i, test in enumerate(tests, 1):
        print(f"\n{'='*50}")
        print(f"TEST {i}: {test[:80]}...")
        result = extract(test)
        print(json.dumps(result, indent=2, ensure_ascii=False))