""" LLM Market Analysis via Google Gemini API (google-genai SDK). Provides directional market insight (-1 to +1) with reasoning. Falls back gracefully to neutral (0) on API errors or quota exhaustion. """ from __future__ import annotations import os import json import time from dotenv import load_dotenv # Load .env first load_dotenv() # ── New google-genai SDK ────────────────────────────────────────────────────── from llm_proxy import get_llm_client # ── Model/Client state ──────────────────────────────────────────────────────── _llm_client = None def _init_client() -> bool: global _llm_client if _llm_client is not None: return True _llm_client = get_llm_client() if _llm_client.mode == "none": print("[LLM] ERROR: No valid LLM client (Proxy or Gemini) could be initialized.") return False return True # ── Cache: 30-minute TTL per ticker ────────────────────────────────────────── _LLM_CACHE_TTL = 30 * 60 _llm_cache: dict = {} # {ticker: {"result": ..., "timestamp": ...}} def _get_cached(ticker: str) -> dict | None: entry = _llm_cache.get(ticker) if entry and (time.time() - entry["timestamp"]) < _LLM_CACHE_TTL: return entry["result"] return None def _set_cached(ticker: str, result: dict): _llm_cache[ticker] = {"result": result, "timestamp": time.time()} if len(_llm_cache) > 30: oldest = min(_llm_cache, key=lambda k: _llm_cache[k]["timestamp"]) del _llm_cache[oldest] def _build_prompt( ticker: str, current_price: float, indicators: dict, sentiment_score: float, sentiment_label: str, recent_prices: list[float], ) -> str: """Build a concise prompt for market analysis.""" price_change_5d = ((current_price - recent_prices[0]) / recent_prices[0] * 100) if recent_prices else 0 return f"""You are a stock market analyst. Analyze this Indian stock and give a prediction bias. STOCK: {ticker} CURRENT PRICE: ₹{current_price:.2f} 5-DAY PRICE CHANGE: {price_change_5d:+.2f}% RECENT 5 CLOSES: {[round(p,2) for p in recent_prices[-5:]]} TECHNICAL INDICATORS: - RSI: {indicators.get('RSI', 'N/A')} (>70 overbought, <30 oversold) - MACD: {indicators.get('MACD', 'N/A')} (Signal: {indicators.get('MACD_Signal', 'N/A')}) - SMA_20: {indicators.get('SMA_20', 'N/A')} | SMA_50: {indicators.get('SMA_50', 'N/A')} - BB_Width: {indicators.get('BB_Width', 'N/A')} | ATR: {indicators.get('ATR', 'N/A')} NEWS SENTIMENT: {sentiment_label} (score: {sentiment_score:+.3f}) Respond ONLY with this exact JSON format, no other text: {{"score": , "direction": "", "reasoning": ""}} Score guide: -0.99 = very bearish, 0 = neutral, +0.99 = very bullish. NEVER return exactly -1.0 or 1.0.""" def analyze_with_llm( ticker: str, current_price: float, indicators: dict, sentiment_score: float = 0.5, sentiment_label: str = "neutral", recent_prices: list[float] | None = None, ) -> dict: """ Get LLM market analysis. Returns: {"score": float, "direction": str, "reasoning": str, "source": str} """ global _llm_client # Check cache first cached = _get_cached(ticker) if cached is not None: return cached # Default fallback — use 0.5 (neutral mid-point) fallback = { "score": 0.5, "direction": "neutral", "reasoning": "LLM analysis unavailable — using technical/sentiment signals only", "source": "fallback", } if not _init_client(): _set_cached(ticker, fallback) return fallback if recent_prices is None: recent_prices = [current_price] prompt = _build_prompt( ticker, current_price, indicators, sentiment_score, sentiment_label, recent_prices, ) try: text = _llm_client.generate_content(prompt) text = text.strip() # Extract JSON from response (handle markdown code blocks) if "```" in text: # Try to grab the first code block parts = text.split("```") if len(parts) >= 2: text = parts[1] if text.startswith("json"): text = text[4:] text = text.strip() parsed = json.loads(text) # Clamp to strict (-0.99, 0.99) — never allow boundary values ±1.0 raw_score = float(parsed.get("score", 0.5)) score = max(-0.99, min(0.99, raw_score)) result = { "score": round(score, 3), "direction": parsed.get("direction", "neutral"), "reasoning": parsed.get("reasoning", ""), "source": f"llm_proxy/{_llm_client.mode}", } _set_cached(ticker, result) return result except Exception as e: error_msg = str(e) fallback["reasoning"] = f"LLM analysis failed: {error_msg[:100]}" _set_cached(ticker, fallback) return fallback