backend/intelligence/sentiment_analyzer.py

"""
Sentiment Analyzer — Batched target-level financial sentiment via OpenAI gpt-4o-mini.
Processes headlines in a single API call for maximum token efficiency.
"""
import json
import logging
import os

from openai import OpenAI

logger = logging.getLogger(__name__)

SENTIMENT_SYSTEM_PROMPT = """You are a financial sentiment classifier. You will receive numbered headlines.
Return a JSON object with key "results" containing an array. Each element must have these exact keys:
- "idx": the headline number (integer)
- "overall_sentiment": "positive", "negative", or "neutral"
- "confidence": float 0.0 to 1.0
- "targets": array of {"name": string, "sentiment": string} for specific companies mentioned
- "impact_summary": one sentence, max 20 words, explaining financial impact

Distinguish target-level sentiment. Example:
"Netflix dips despite Apple's surge" → Netflix=negative, Apple=positive

You MUST return exactly one result per headline, in order."""

SENTIMENT_USER_TEMPLATE = """Analyze these {count} financial headlines:

{headlines}

Return: {{"results": [{{"idx": 0, "overall_sentiment": "positive", "confidence": 0.8, "targets": [{{"name": "AAPL", "sentiment": "positive"}}], "impact_summary": "Apple gains on strong earnings"}}]}}"""


def analyze_sentiment_batch(headlines: list[dict]) -> list[dict]:
    """
    Analyze sentiment for a batch of news items using OpenAI.
    Each item should have at least 'title' and optionally 'summary'.
    Returns list of sentiment results in same order as input.
    """
    if not headlines:
        return []

    api_key = os.getenv("OPENAI_API_KEY", "")
    if not api_key:
        logger.warning("No OPENAI_API_KEY set, returning neutral sentiment for all")
        return [_neutral("No API key") for _ in headlines]

    model = os.getenv("OPENAI_MODEL", "gpt-4o-mini")
    client = OpenAI(api_key=api_key)

    # Format headlines for the prompt
    formatted = "\n".join(
        f"[{i}] {item['title']}"
        + (f" — {item.get('summary', '')[:80]}" if item.get("summary") else "")
        for i, item in enumerate(headlines)
    )

    try:
        response = client.chat.completions.create(
            model=model,
            messages=[
                {"role": "system", "content": SENTIMENT_SYSTEM_PROMPT},
                {"role": "user", "content": SENTIMENT_USER_TEMPLATE.format(
                    count=len(headlines), headlines=formatted
                )},
            ],
            temperature=0.1,
            max_tokens=max(1500, len(headlines) * 80),
            response_format={"type": "json_object"},
        )

        content = response.choices[0].message.content or "{}"
        logger.debug(f"Raw sentiment response: {content[:500]}")

        raw = json.loads(content)

        # Extract the results array from various possible wrapper keys
        results = []
        if isinstance(raw, dict):
            for key in ("results", "headlines", "analysis", "data", "sentiments"):
                if key in raw and isinstance(raw[key], list):
                    results = raw[key]
                    break
            if not results:
                # Try first list value in the dict
                for v in raw.values():
                    if isinstance(v, list) and len(v) > 0:
                        results = v
                        break
        elif isinstance(raw, list):
            results = raw

        # Normalize each result to ensure expected keys exist
        normalized = []
        for r in results:
            if not isinstance(r, dict):
                continue
            normalized.append({
                "overall_sentiment": r.get("overall_sentiment", r.get("sentiment", "neutral")),
                "confidence": float(r.get("confidence", 0.5)),
                "targets": r.get("targets", []),
                "impact_summary": r.get("impact_summary", r.get("summary", "")),
            })

        # Pad if model returned fewer than expected
        while len(normalized) < len(headlines):
            normalized.append(_neutral("Model returned fewer results"))

        tokens = response.usage.total_tokens if response.usage else "?"
        logger.info(f"Sentiment analyzed: {len(headlines)} headlines, "
                     f"{len(results)} results parsed (tokens: {tokens})")
        return normalized[:len(headlines)]

    except Exception as e:
        logger.error(f"Sentiment analysis error: {e}")
        return [_neutral(f"Error: {str(e)[:50]}") for _ in headlines]


def _neutral(reason: str = "") -> dict:
    return {
        "overall_sentiment": "neutral",
        "confidence": 0.3,
        "targets": [],
        "impact_summary": reason or "Unable to analyze",
    }