src/agentcore/classify.py

"""Stage 2-3 of the pipeline: rule-based pre-classifier and the GPT-4o-mini intent classifier."""
import asyncio
import html
import json
import logging
import os
import re
import time
import uuid
from datetime import datetime
from typing import Any, Dict, List, Literal, Optional, Tuple
from urllib.parse import quote

import httpx
from pydantic import BaseModel, ConfigDict, Field

from src.config import get_settings, LIBBEE_VERSION

from src.agentcore.models import ChatMessage
from src.agentcore.constants import (
    CURRENT_INFO_RE,
    GREETING_FOLLOWUP_RE,
    HISTORY_WINDOW,
    HOURS_RE,
    KU_CAMPUS_RE,
    LIBRARY_CUE_RE,
    MEDICAL_KEYWORDS,
    MEDICAL_SEARCH_RE,
    RESEARCH_CUE_RE,
    SOCIAL_RE,
    SUMMARY_RE,
    _CLASSIFIER_SYSTEM,
)
from src.agentcore.utils import _get_llm

logger = logging.getLogger(__name__)

def _is_summary_request(question: str) -> bool:
    return bool(SUMMARY_RE.search(question or ""))


def _is_greeting_menu_followup(question: str, history: List[ChatMessage]) -> bool:
    if not GREETING_FOLLOWUP_RE.match((question or "").strip()):
        return False
    for m in reversed(history):
        if m.role == "assistant":
            last = m.content.lower()
            return "i'm libbee" in last or "are you looking for one of these" in last
    return False


def _looks_library_hours_question(question: str) -> bool:
    q = (question or "").strip().lower()
    return bool(HOURS_RE.search(q)) and any(
        kw in q for kw in (
            "library", "campus", "habshan", "san", "hours", "open", "close",
            "schedule", "till", "until", "today", "tomorrow", "friday",
            "saturday", "sunday", "week", "ramadan",
        )
    )


def _looks_campus_question(question: str) -> bool:
    return bool(KU_CAMPUS_RE.search(question or ""))


def _looks_medical_search(question: str) -> bool:
    q = (question or "").lower()
    if not any(kw in q for kw in MEDICAL_KEYWORDS):
        return False
    return bool(MEDICAL_SEARCH_RE.search(q) or RESEARCH_CUE_RE.search(q) or SUMMARY_RE.search(q))


def _looks_research_question(question: str) -> bool:
    q = question or ""
    return bool(RESEARCH_CUE_RE.search(q) or SUMMARY_RE.search(q))


def _rule_based_classify(question: str) -> Dict[str, str]:
    q = (question or "").lower()
    if SOCIAL_RE.match((question or "").strip()):
        return {
            "intent": "social",
            "casual_answer": (
                "Hello! I'm LibBee, the KU Library AI Assistant. "
                "I'm here to help you with articles, books, databases, and library services. "
                "What would you like to find today?"
            ),
        }
    if _looks_medical_search(question):
        return {"intent": "search_medical"}
    if _looks_research_question(question):
        return {"intent": "search_academic"}
    if LIBRARY_CUE_RE.search(q):
        return {"intent": "library_info"}
    if CURRENT_INFO_RE.search(q):
        return {"intent": "general_recent"}
    return {"intent": "general"}


async def _llm_classify(question: str, history: List[ChatMessage], model: str) -> Dict[str, str]:
    settings = get_settings()
    if not settings.openai_api_key and not settings.anthropic_api_key:
        return _rule_based_classify(question)
    try:
        ctx_parts = [
            f"{m.role}: {m.content[:100]}"
            for m in history[-HISTORY_WINDOW:]
            if m.role in ("user", "assistant")
        ]
        ctx = " | ".join(ctx_parts) if ctx_parts else "none"
        classify_model = "gpt" if settings.openai_api_key else model
        llm = _get_llm(classify_model, temperature=0, max_tokens=120)
        response = await llm.ainvoke([
            {"role": "system", "content": _CLASSIFIER_SYSTEM},
            {"role": "user", "content": f"Message: {question[:400]}\nContext: {ctx}\nJSON:"},
        ])
        raw = response.content.strip()
        if raw.startswith("```"):
            raw = raw.split("\n", 1)[1].rsplit("```", 1)[0].strip()
        result = json.loads(raw)
        intent = result.get("intent", "general")
        if intent not in {"social", "library_info", "search_academic", "search_medical", "general_recent", "general", "sensitive"}:
            intent = "general"
        return {"intent": intent, "casual_answer": result.get("casual_answer", "")}
    except Exception as e:
        logger.warning(f"LLM classifier error: {e} — falling back to rule-based")
        return _rule_based_classify(question)