services.py

import os
import json
from dotenv import load_dotenv
from google import genai
from google.genai import types
from typing import List, Dict, Any, Optional

# 載入環境變數
load_dotenv()

class GeminiService:
    def __init__(self):
        api_key = os.getenv("GEMINI_API_KEY")
        if not api_key:
            print("警告：找不到 GEMINI_API_KEY")
        
        self.client = genai.Client(api_key=api_key) if api_key else None
        # 建議使用最新模型以獲得最佳分析能力
        self.model_id = os.getenv("GEMINI_MODEL_ID", "gemini-2.0-flash")

    def _check_client(self):
        if not self.client:
            raise ValueError("API Key 未設定")

    def search_companies(self, query: str, exclude_names: List[str] = []) -> List[Dict]:
        """
        Step 1: 搜尋台灣公司
        """
        self._check_client()
        exclusion_prompt = ""
        if exclude_names:
            exclusion_prompt = f"IMPORTANT: Do not include: {', '.join(exclude_names)}."

        # Phase 1: Google Search (廣泛搜尋)
        search_prompt = f"""
        Using Google Search, find 5 to 10 prominent companies in Taiwan related to the query: "{query}".
        
        **CRITICAL INSTRUCTIONS:**
        1. **TARGET:** Focus on Taiwanese companies (or global companies with a major branch in Taiwan).
        2. **IDENTIFIERS:** Try to find their distinct "Company Name" (e.g., 台積電 / 台灣積體電路製造股份有限公司).
        {exclusion_prompt}
        
        List them (Full Name - Industry/Sector) in Traditional Chinese.
        """

        search_response = self.client.models.generate_content(
            model=self.model_id,
            contents=search_prompt,
            config=types.GenerateContentConfig(
                tools=[types.Tool(google_search=types.GoogleSearch())]
            )
        )
        raw_text = search_response.text

        # Phase 2: Extract JSON (結構化)
        extract_prompt = f"""
        From the text below, extract company names and their industry.
        Calculate a Relevance Score (0-100) based on query: "{query}".
        
        Return ONLY a JSON array: [{{"name": "...", "industry": "...", "relevanceScore": 85}}]
        
        Text:
        ---
        {raw_text}
        ---
        """

        extract_response = self.client.models.generate_content(
            model=self.model_id,
            contents=extract_prompt,
            config=types.GenerateContentConfig(
                response_mime_type='application/json'
            )
        )

        try:
            return json.loads(extract_response.text)
        except Exception as e:
            print(f"JSON Parse Error: {e}")
            return []

    def get_company_details(self, company: Dict) -> Dict:
        """
        Step 2: 進行商業徵信調查 (Deep Dive)
        """
        self._check_client()
        name = company.get('name')
        
        prompt = f"""
        Act as a professional "Business Due Diligence Analyst" (商業徵信分析師). 
        Conduct a comprehensive investigation on the Taiwanese company: "{name}".

        **Investigation Targets (Must search for these specifically):**

        1.  **Corporate Identity (基本資料)**:
            - Find the **Tax ID (統一編號)**.
            - **Registered Capital (資本額)**.
            - **Representative (代表人)**.
            - **Establishment Date (設立日期)**.
            - *Source Hint: Ministry of Economic Affairs (經濟部商業司), Datagovtw.*

        2.  **Scale & Business (規模與業務)**:
            - **Employee Count**: Estimated number of employees.
            - **Core Products/Services**: What do they actually sell or do?
            - *Source Hint: 104 Job Bank, Company Website, LinkedIn.*

        3.  **Market Reputation & Culture (評價與文化 - KEY PART)**:
            - Search for employee reviews on **PTT (Tech_Job, Soft_Job, Salary)**, **Dcard (Work board)**, **Qollie (求職天眼通)**, or **Google Maps**.
            - Summarize the **Pros** (e.g., high pay, free snacks) and **Cons** (e.g., toxic management, forced overtime, family business style).
            - *Tone:* Be objective but highlight recurring complaints.

        4.  **Legal & Risk Assessment (法律與風險 - CRITICAL)**:
            - Search for keywords: "{name} 判決", "{name} 勞資糾紛", "{name} 違反勞基法", "{name} 詐騙", "{name} 吸金", "{name} 罰款".
            - List any major lawsuits, fines, or controversies found in news or government records.
            - If clean, state "No major public legal disputes found."

        **Format Requirements**:
        - Structure the output as a clean, readable report using Markdown.
        - Use clear headings.
        - **Language**: Traditional Chinese (繁體中文).
        """

        response = self.client.models.generate_content(
            model=self.model_id,
            contents=prompt,
            config=types.GenerateContentConfig(
                tools=[types.Tool(google_search=types.GoogleSearch())]
            )
        )
        
        # Extract Sources
        sources = []
        if response.candidates[0].grounding_metadata and response.candidates[0].grounding_metadata.grounding_chunks:
            for chunk in response.candidates[0].grounding_metadata.grounding_chunks:
                if chunk.web and chunk.web.uri and chunk.web.title:
                    sources.append({"title": chunk.web.title, "uri": chunk.web.uri})
        
        # Deduplicate
        unique_sources = {v['uri']: v for v in sources}.values()

        return {
            "text": response.text,
            "sources": list(unique_sources)
        }

    def chat_with_ai(self, history: List[Dict], new_message: str, context: str) -> str:
        self._check_client()
        system_instruction = f"You are a sharp Business Analyst. Answer based on this due diligence report:\n{context}"
        
        chat_history = []
        for h in history:
            role = "user" if h["role"] == "user" else "model"
            chat_history.append(types.Content(role=role, parts=[types.Part(text=h["content"])]))

        chat = self.client.chats.create(
            model=self.model_id,
            history=chat_history,
            config=types.GenerateContentConfig(
                system_instruction=system_instruction
            )
        )
        
        response = chat.send_message(new_message)
        return response.text