app.py

import os

# Force writable home and cache paths before other imports
os.environ.setdefault("HOME", "/app")
os.environ.setdefault("HF_HOME", "/app/hf_cache")
os.environ.setdefault("HF_HUB_CACHE", "/app/hf_cache")
os.environ.setdefault("XDG_CACHE_HOME", "/app/.cache")
os.makedirs(os.environ["HF_HOME"], exist_ok=True)
os.makedirs(os.environ["XDG_CACHE_HOME"], exist_ok=True)

import json
from typing import Dict, List, Generator

import gradio as gr
import requests
from dotenv import load_dotenv

load_dotenv()

# -------- Keys (multi-key support) --------
FINNHUB_KEYS_RAW = os.getenv("FINNHUB_KEYS", "")
FINNHUB_KEYS = [k.strip() for k in FINNHUB_KEYS_RAW.split("\n") if k.strip()] if FINNHUB_KEYS_RAW else []
FINNHUB_API_KEY = os.getenv("FINNHUB_API_KEY", "")
if FINNHUB_API_KEY and FINNHUB_API_KEY.strip():
	FINNHUB_KEYS = FINNHUB_KEYS or [FINNHUB_API_KEY.strip()]

RAPIDAPI_KEYS_RAW = os.getenv("RAPIDAPI_KEYS", "")
RAPIDAPI_KEYS = [k.strip() for k in RAPIDAPI_KEYS_RAW.split("\n") if k.strip()] if RAPIDAPI_KEYS_RAW else []
RAPIDAPI_KEY = os.getenv("RAPIDAPI_KEY", "")
if RAPIDAPI_KEY and RAPIDAPI_KEY.strip():
	RAPIDAPI_KEYS = RAPIDAPI_KEYS or [RAPIDAPI_KEY.strip()]

RAPIDAPI_HOST = "alpha-vantage.p.rapidapi.com"

# -------- llama.cpp GGUF model --------
MODEL_REPO = "mradermacher/Fin-o1-8B-GGUF"
GGUF_OVERRIDE = os.getenv("GGUF_FILENAME", "").strip()
N_THREADS = int(os.getenv("LLAMA_CPP_THREADS", str(os.cpu_count() or 4)))
CTX_LEN = int(os.getenv("LLAMA_CPP_CTX", "3072"))  # CPU-friendly default
N_BATCH = int(os.getenv("LLAMA_CPP_BATCH", "128"))

from huggingface_hub import snapshot_download
from llama_cpp import Llama

_llm = None


def _pick_gguf_file(root_dir: str, override: str | None) -> str:
	import glob
	if override:
		path = os.path.join(root_dir, override)
		if os.path.isfile(path) and os.path.getsize(path) > 0:
			return path
		candidates = glob.glob(os.path.join(root_dir, "**", override), recursive=True)
		for c in candidates:
			if os.path.getsize(c) > 0:
				return c
	preferred: List[str] = [
		"Fin-o1-8B.Q4_K_M.gguf",  # explicit 8B file name first
		"Q4_K_M", "Q4_K_S", "Q4_0", "Q3_K_M", "Q3_K_S", "Q3_0", "Q2_K", "Q2_0",
	]
	import glob as _glob
	ggufs = _glob.glob(os.path.join(root_dir, "**", "*.gguf"), recursive=True)
	if not ggufs:
		raise FileNotFoundError("No .gguf files found in snapshot")
	for key in preferred:
		for f in ggufs:
			if key in os.path.basename(f):
				return f
	return ggufs[0]


def load_model():
	global _llm
	if _llm is not None:
		return _llm
	repo_dir = snapshot_download(
		repo_id=MODEL_REPO,
		allow_patterns=["*.gguf"],
		cache_dir=os.getenv("HF_HOME", "/app/hf_cache"),
		local_files_only=False,
		resume_download=True,
	)
	try:
		model_path = _pick_gguf_file(repo_dir, GGUF_OVERRIDE or None)
	except Exception as e:
		raise RuntimeError(f"GGUF not found: {e}")
	try:
		_llm = Llama(
			model_path=model_path,
			n_ctx=CTX_LEN,
			n_threads=N_THREADS,
			n_batch=N_BATCH,
			use_mlock=False,
			use_mmap=True,
			verbose=False,
		)
	except Exception as e:
		raise RuntimeError(f"Failed to load GGUF: {e}. Set GGUF_FILENAME to an available 8B file if needed.")
	return _llm


def generate_response_stream(prompt: str, temperature: float = 0.2, max_new_tokens: int = 384) -> Generator[str, None, None]:
	yield "Initializing model..."
	llm = load_model()
	yield "Model loaded. Generating..."
	accum = ""
	for chunk in llm(prompt=prompt, max_tokens=max_new_tokens, temperature=temperature, stream=True):
		text = chunk.get("choices", [{}])[0].get("text", "")
		if text:
			accum += text
			yield accum


def generate_response(prompt: str, temperature: float = 0.2, max_new_tokens: int = 384) -> str:
	# non-streaming fallback
	llm = load_model()
	res = llm(
		prompt=prompt,
		max_tokens=max_new_tokens,
		temperature=temperature,
		stop=["</s>", "<|eot_id|>"],
	)
	return res.get("choices", [{}])[0].get("text", "")


# -------- Robust requests session with retry --------
from requests.adapters import HTTPAdapter
from urllib3.util.retry import Retry

def create_session() -> requests.Session:
	session = requests.Session()
	retry_strategy = Retry(
		total=3,
		backoff_factor=1.0,
		status_forcelist=[429, 500, 502, 503, 504],
	)
	adapter = HTTPAdapter(max_retries=retry_strategy)
	session.mount("http://", adapter)
	session.mount("https://", adapter)
	return session

http = create_session()

# -------- Helpers for mock candles --------

def _create_mock_candles(symbol: str, count: int = 60) -> Dict:
	import time as tmod
	import random
	now = int(tmod.time())
	day = 86400
	t, o, h, l, c, v = [], [], [], [], [], []
	price = 100.0 + (hash(symbol) % 50)
	for i in range(count, 0, -1):
		ts = now - i * day
		op = max(1.0, price + random.uniform(-2, 2))
		hi = op + random.uniform(0, 2)
		lo = max(0.5, op - random.uniform(0, 2))
		cl = max(0.5, lo + random.uniform(0, (hi - lo) or 1))
		vol = abs(int(random.gauss(1_000_000, 250_000)))
		t.append(ts); o.append(op); h.append(hi); l.append(lo); c.append(cl); v.append(vol)
		price = cl
	return {"s": "ok", "t": t, "o": o, "h": h, "l": l, "c": c, "v": v, "source": "mock"}

# -------- Data helpers (Finnhub with fallback to Alpha Vantage) --------

def fetch_finnhub_candles(symbol: str, resolution: str = "D", count: int = 60) -> Dict:
	"""Try Finnhub first cycling keys; on 401/403 or exhaustion, raise to caller."""
	if not FINNHUB_KEYS:
		raise ValueError("Missing FINNHUB_KEYS/FINNHUB_API_KEY")
	import time as _time
	end = int(__import__("time").time())
	start = end - count * 86400
	last_error: Exception | None = None
	for api_key in FINNHUB_KEYS:
		url = (
			f"https://finnhub.io/api/v1/stock/candle?symbol={symbol}"
			f"&resolution={resolution}&from={start}&to={end}&token={api_key}"
		)
		try:
			r = http.get(url, timeout=30)
			if r.status_code in (401, 403):
				last_error = requests.HTTPError(f"Finnhub auth error {r.status_code}")
				continue
			r.raise_for_status()
			data = r.json()
			if data.get("s") == "ok":
				data["source"] = "finnhub"
				return data
			last_error = RuntimeError(f"Finnhub returned status: {data.get('s')}")
		except Exception as e:
			last_error = e
			continue
		finally:
			_time.sleep(0.3)
	raise last_error or RuntimeError("Finnhub candles failed")


def fetch_alpha_vantage_series_daily(symbol: str, outputsize: str = "compact", count: int = 60) -> Dict:
	"""Fallback: Alpha Vantage TIME_SERIES_DAILY via RapidAPI, format like Finnhub candles."""
	if not RAPIDAPI_KEYS:
		return _create_mock_candles(symbol, count)
	import time as _time
	for api_key in RAPIDAPI_KEYS:
		try:
			url = f"https://{RAPIDAPI_HOST}/query"
			headers = {"X-RapidAPI-Key": api_key, "X-RapidAPI-Host": RAPIDAPI_HOST}
			params = {"function": "TIME_SERIES_DAILY", "symbol": symbol, "outputsize": outputsize}
			r = http.get(url, headers=headers, params=params, timeout=30)
			r.raise_for_status()
			data = r.json()
			if isinstance(data, dict) and any(k in data for k in ("Note", "Error Message", "Information")):
				# rate limit or error; try next key
				continue
			series = data.get("Time Series (Daily)") or {}
			if not series:
				continue
			dates = sorted(series.keys())[-count:]
			import time as tmod
			t, o, h, l, c, v = [], [], [], [], [], []
			for d in dates:
				row = series[d]
				try:
					op_v = float(row.get("1. open"))
					h_v = float(row.get("2. high"))
					l_v = float(row.get("3. low"))
					c_v = float(row.get("4. close"))
					v_v = float(row.get("5. volume"))
					ts = int(tmod.mktime(tmod.strptime(d, "%Y-%m-%d")))
				except Exception:
					continue
				t.append(ts); o.append(op_v); h.append(h_v); l.append(l_v); c.append(c_v); v.append(v_v)
			return {"s": "ok", "t": t, "o": o, "h": h, "l": l, "c": c, "v": v, "source": "alpha_vantage"}
		except Exception:
			continue
		finally:
			_time.sleep(0.5)
	# If all keys failed or no data, return mock to keep UI responsive
	return _create_mock_candles(symbol, count)


def fetch_alpha_vantage_overview(symbol: str) -> Dict:
	if not RAPIDAPI_KEYS:
		raise ValueError("Missing RAPIDAPI_KEYS/RAPIDAPI_KEY")
	for api_key in RAPIDAPI_KEYS:
		try:
			url = f"https://{RAPIDAPI_HOST}/query"
			headers = {"x-rapidapi-key": api_key, "x-rapidapi-host": RAPIDAPI_HOST}
			params = {"function": "OVERVIEW", "symbol": symbol}
			r = http.get(url, headers=headers, params=params, timeout=30)
			r.raise_for_status()
			data = r.json()
			if data:
				return data
		except Exception:
			continue
	raise RuntimeError("Alpha Vantage OVERVIEW failed")


# -------- Prompts --------

def build_price_prediction_prompt(symbol: str, candles: Dict) -> str:
	context = json.dumps(candles)[:10000]
	source = candles.get("source", "finnhub")
	return (
		f"You are a financial analyst agent. Analyze recent OHLCV candles for {symbol} (source: {source}) and provide a short-term price prediction. "
		f"Explain key drivers in bullet points and give a 1-2 sentence forecast.\n\nData JSON: {context}\n\n"
	)


def build_equity_research_prompt(symbol: str, overview: Dict) -> str:
	context = json.dumps(overview)[:10000]
	return (
		"You are an equity research analyst. Using the fundamentals overview, write a concise equity research note including: "
		"Business summary, recent performance, profitability, leverage, valuation multiples, key risks, and an investment view (Buy/Hold/Sell) with rationale.\n\n"
		f"Ticker: {symbol}\nFundamentals JSON: {context}\n"
	)


# -------- Gradio UI --------

def ui_app():
	with gr.Blocks(title="Fin-o1-8B Tools") as demo:
		gr.Markdown("""# Fin-o1-8B Tools
Two tabs: Price Prediction (Finnhub with Alpha Vantage fallback) and Equity Research (Alpha Vantage via RapidAPI).""")

		with gr.Tab("Price Prediction"):
			symbol = gr.Textbox(label="Ticker (e.g., AAPL)", value="AAPL")
			resolution = gr.Dropdown(["D", "60", "30", "15", "5"], value="D", label="Resolution")
			count = gr.Slider(20, 160, value=60, step=5, label="Num candles")
			temp = gr.Slider(0.0, 1.0, value=0.2, step=0.05, label="Temperature")
			max_new = gr.Slider(64, 768, value=384, step=16, label="Max new tokens")
			btn = gr.Button("Predict")
			out = gr.Textbox(lines=30, show_copy_button=True)

			def on_predict(sym, res, cnt, temperature, max_tokens):
				try:
					candles = fetch_finnhub_candles(sym, res, int(cnt))
				except Exception:
					try:
						candles = fetch_alpha_vantage_series_daily(sym, outputsize="compact")
					except Exception as e2:
						yield f"Error fetching candles: {e2}"
						return
				prompt = build_price_prediction_prompt(sym, candles)
				for text in generate_response_stream(prompt, temperature=temperature, max_new_tokens=int(max_tokens)):
					yield text

			btn.click(on_predict, inputs=[symbol, resolution, count, temp, max_new], outputs=out, show_progress=True)

		with gr.Tab("Equity Research Report"):
			symbol2 = gr.Textbox(label="Ticker (e.g., MSFT)", value="MSFT")
			temp2 = gr.Slider(0.0, 1.0, value=0.2, step=0.05, label="Temperature")
			max_new2 = gr.Slider(64, 768, value=384, step=16, label="Max new tokens")
			btn2 = gr.Button("Generate Report")
			out2 = gr.Textbox(lines=30, show_copy_button=True)

			def on_report(sym, temperature, max_tokens):
				try:
					overview = fetch_alpha_vantage_overview(sym)
				except Exception as e:
					yield f"Error fetching fundamentals: {e}"
					return
				prompt = build_equity_research_prompt(sym, overview)
				for text in generate_response_stream(prompt, temperature=temperature, max_new_tokens=int(max_tokens)):
					yield text

			btn2.click(on_report, inputs=[symbol2, temp2, max_new2], outputs=out2, show_progress=True)

		# Enable queue with default settings for current Gradio version
		demo.queue()
	return demo


if __name__ == "__main__":
	app = ui_app()
	app.launch(server_name=os.getenv("GRADIO_SERVER_NAME", "0.0.0.0"), server_port=int(os.getenv("GRADIO_SERVER_PORT", "7860")))