AarthiAI / backend /llm_analysis.py
PROSTLE's picture
fix: add /reset and /health endpoints, root inference.py, clamp scores to strict (-0.99 to 0.99) for OpenEnv grader
8e4bb54
Raw History Blame Contribute Delete
7.64 kB
"""
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 ──────────────────────────────────────────────────────
GEMINI_AVAILABLE = False
_gemini_client = None
try:
from google import genai as _genai_sdk
GEMINI_AVAILABLE = True
except ImportError:
pass
# ── Model selection (tried in order until one works) ──────────────────────────
GEMINI_MODELS = [
"gemini-2.0-flash-lite",
"gemini-2.5-flash",
"gemini-2.0-flash",
"gemini-flash-latest",
]
_active_model: str | None = None
# ── 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 _init_gemini() -> bool:
global _gemini_client, _active_model
if _gemini_client is not None and _active_model is not None:
return True
if not GEMINI_AVAILABLE:
print("[LLM] ERROR: google-genai SDK not installed.")
return False
api_key = os.getenv("GEMINI_API_KEY")
if not api_key:
print("[LLM] ERROR: GEMINI_API_KEY not found in environment!")
return False
try:
_gemini_client = _genai_sdk.Client(api_key=api_key)
# Pick the first model that responds
for model in GEMINI_MODELS:
try:
_gemini_client.models.generate_content(
model=model, contents="ping"
)
_active_model = model
print(f"[LLM] Ready: {model}")
return True
except Exception as probe_err:
msg = str(probe_err)
if "429" in msg or "RESOURCE_EXHAUSTED" in msg:
print(f"[LLM] {model}: quota exhausted, trying next model...")
continue
elif "404" in msg or "NOT_FOUND" in msg:
print(f"[LLM] {model}: not found, trying next model...")
continue
else:
print(f"[LLM] {model}: {msg[:80]}")
continue
print("[LLM] All Gemini models exhausted quota or unavailable.")
_gemini_client = None
return False
except Exception as e:
print(f"[LLM] Gemini init error: {e}")
_gemini_client = None
return False
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: β‚Ή{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": <float strictly between -0.99 and 0.99, never exactly -1 or 1>, "direction": "<bullish/bearish/neutral>", "reasoning": "<one sentence>"}}
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}
Falls back to neutral if Gemini is unavailable or quota is exhausted.
"""
# Must declare global before any use of these names in this function
global _gemini_client, _active_model
# Check cache first
cached = _get_cached(ticker)
if cached is not None:
return cached
# Default fallback β€” use 0.5 (neutral mid-point), never 0.0 which can boundary-fail the grader
fallback = {
"score": 0.5,
"direction": "neutral",
"reasoning": "LLM analysis unavailable β€” using technical/sentiment signals only",
"source": "fallback",
}
if not _init_gemini():
_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:
response = _gemini_client.models.generate_content(
model=_active_model,
contents=prompt,
)
text = response.text.strip()
# Extract JSON from response (handle markdown code blocks)
if "```" in text:
text = text.split("```")[1]
if text.startswith("json"):
text = text[4:]
text = text.strip()
parsed = json.loads(text)
# Clamp to strict (-0.99, 0.99) β€” never allow boundary values Β±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"gemini/{_active_model}",
}
_set_cached(ticker, result)
return result
except Exception as e:
error_msg = str(e)
if "429" in error_msg or "RESOURCE_EXHAUSTED" in error_msg:
# Quota hit β€” reset so next call tries to find a working model
fallback["reasoning"] = "LLM quota exhausted β€” using technical/sentiment signals only"
_gemini_client = None
_active_model = None
elif "API_KEY_INVALID" in error_msg or "API key expired" in error_msg:
fallback["reasoning"] = "LLM API key invalid/expired β€” using technical/sentiment signals only"
_gemini_client = None
_active_model = None
else:
fallback["reasoning"] = f"LLM analysis failed: {error_msg[:100]}"
# Cache non-auth errors so we don't spam the API
_set_cached(ticker, fallback)
return fallback