app/services/intent.py

from typing import Literal
from pydantic import BaseModel
from app.services.llm import LLMService

class IntentResponse(BaseModel):
    category: Literal["SEARCH_REQUIRED", "CHAT_ONLY", "CODING_TASK", "DATA_ANALYSIS"]
    reasoning: str
    risk_level: Literal["LOW", "MEDIUM", "HIGH"]

class IntentService:
    def __init__(self, llm_service: LLMService):
        self.llm = llm_service
        self.system_prompt = """
        You are the 'Intent Analyzer' for an AI Operating System.
        Your job is to route the user's request to the correct module.
        
        Analyze the USER QUERY and return a JSON object.

        CATEGORIES:
        - SEARCH_REQUIRED: Query asks for current events, news, specific facts not in general knowledge, or research. (e.g., "Stock price of Apple", "Latest AI papers")
        - CHAT_ONLY: General greetings, philosophical questions, summaries of previous context, or logic puzzles. (e.g., "Hi", "Explain Stoicism")
        - CODING_TASK: Requests to write, debug, or explain code.
        - DATA_ANALYSIS: Requests involving CSVs, charts, or math aggregations.

        RISK LEVELS:
        - HIGH: Asking for dangerous/illegal content, PII, or financial advice.
        - MEDIUM: Ambiguous queries or potential controversies.
        - LOW: Safe, standard queries.

        Output format: {"category": "...", "reasoning": "...", "risk_level": "..."}
        """

    async def analyze(self, query: str) -> IntentResponse:
        # For efficiency, we can use a faster/smaller model or just the main one with strictly low temp
        prompt = f"{self.system_prompt}\n\nUSER QUERY: {query}"
        
        # We'll rely on the main LLM to parse this for now.
        # ideally this uses a 'router' model (cheap/fast)
        response_text = await self.llm._generate(
            messages=[{"role": "user", "content": prompt}],
            temperature=0.0
        )
        
        # Simple parsing logic (robustness would require structured output mode or regex)
        # Assuming LLM behaves well with JSON instructions.
        import json
        import re

        try:
            # Clean markdown code blocks if present
            clean_text = re.sub(r"```json|```", "", response_text).strip()
            data = json.loads(clean_text)
            return IntentResponse(**data)
        except Exception as e:
            # Fallback to Safe Default
            print(f"Intent Parsing Failed: {e}. Defaulting to SEARCH.")
            return IntentResponse(
                category="SEARCH_REQUIRED", 
                reasoning="Parsing error, defaulting to search.",
                risk_level="low"
            )

intent_service = None # initialized in main/deps