app/analyzer.py

"""
Conversation analyzer: slot extraction and intent classification.

Extracts structured slots from conversation history (stateless — re-extracts
every call) and classifies the user's intent to determine agent behavior.
"""

import re
from app.models import Message


# Slot definitions with extraction patterns
SENIORITY_KEYWORDS = {
    "entry": ["entry-level", "entry level", "junior", "fresher", "graduate", "intern", "trainee"],
    "mid": ["mid-level", "mid level", "mid-professional", "3-5 years", "4 years", "intermediate"],
    "senior": ["senior", "lead", "principal", "staff", "5+ years", "7+ years", "10+ years", "experienced"],
    "executive": ["director", "executive", "cxo", "c-suite", "vp", "vice president", "head of", "15 years"],
    "manager": ["manager", "supervisor", "team lead", "front line manager"],
}

ASSESSMENT_PURPOSE = {
    "selection": ["hiring", "screening", "selection", "recruit", "candidate", "fill", "hire", "assess candidates"],
    "development": ["development", "re-skill", "reskill", "upskill", "talent audit", "training", "coaching", "growth"],
}


class ConversationSlots:
    """Extracted slots from conversation history."""
    
    def __init__(self):
        self.role: str | None = None           # e.g., "Java developer", "contact center agent"
        self.seniority: str | None = None       # entry, mid, senior, executive, manager
        self.technical_skills: list[str] = []   # e.g., ["java", "spring", "sql"]
        self.behavioral_skills: list[str] = []  # e.g., ["leadership", "communication"]
        self.industry: str | None = None        # e.g., "healthcare", "manufacturing"
        self.language: str | None = None        # e.g., "English", "Spanish"
        self.volume: str | None = None          # e.g., "500", "high volume"
        self.purpose: str | None = None         # selection or development
        self.constraints: list[str] = []        # e.g., ["quick", "short", "no personality"]
        self.specific_requests: list[str] = []  # e.g., ["add personality", "remove REST"]
    
    @property
    def confidence(self) -> float:
        """
        Confidence score (0-1) based on how many slots are filled.
        Used to decide whether to clarify or recommend.
        """
        filled = 0
        total = 5  # role, seniority, skills (tech or behavioral), purpose are most important
        
        if self.role:
            filled += 1
        if self.seniority:
            filled += 1
        if self.technical_skills or self.behavioral_skills:
            filled += 1
        if self.purpose:
            filled += 0.5
        if self.industry:
            filled += 0.5
        
        return min(filled / total, 1.0)
    
    def to_summary(self) -> str:
        """Summarize slots for LLM context."""
        parts = []
        if self.role:
            parts.append(f"Role: {self.role}")
        if self.seniority:
            parts.append(f"Seniority: {self.seniority}")
        if self.technical_skills:
            parts.append(f"Technical Skills: {', '.join(self.technical_skills)}")
        if self.behavioral_skills:
            parts.append(f"Behavioral Skills: {', '.join(self.behavioral_skills)}")
        if self.industry:
            parts.append(f"Industry: {self.industry}")
        if self.language:
            parts.append(f"Language: {self.language}")
        if self.volume:
            parts.append(f"Volume: {self.volume}")
        if self.purpose:
            parts.append(f"Purpose: {self.purpose}")
        if self.constraints:
            parts.append(f"Constraints: {', '.join(self.constraints)}")
        if self.specific_requests:
            parts.append(f"Specific Requests: {', '.join(self.specific_requests)}")
        
        return "\n".join(parts) if parts else "No slots extracted yet."
    
    def build_search_queries(self) -> list[str]:
        """Build search queries from filled slots for retrieval."""
        queries = []
        
        # Primary query: role + skills
        primary_parts = []
        if self.role:
            primary_parts.append(self.role)
        if self.technical_skills:
            primary_parts.append(" ".join(self.technical_skills))
        if primary_parts:
            queries.append(" ".join(primary_parts))
        
        # Technical skills query
        if self.technical_skills:
            queries.append(" ".join(self.technical_skills))
        
        # Behavioral query
        if self.behavioral_skills:
            queries.append(" ".join(self.behavioral_skills))
        
        # Seniority + role query
        if self.seniority and self.role:
            queries.append(f"{self.seniority} {self.role}")
        
        # Industry-specific query
        if self.industry:
            queries.append(self.industry)
        
        # Specific requests
        for req in self.specific_requests:
            queries.append(req)
        
        # Fallback: use the latest user message as-is
        if not queries:
            queries.append("assessment")
        
        return queries


class ConversationAnalyzer:
    """
    Analyzes conversation history to extract slots and classify intent.
    Stateless: re-analyzes from scratch every call.
    """
    
    def extract_slots(self, messages: list[Message]) -> ConversationSlots:
        """Extract all slots from the full conversation history."""
        slots = ConversationSlots()
        
        # Combine all user messages for analysis
        user_texts = [m.content for m in messages if m.role == "user"]
        combined = " ".join(user_texts).lower()
        
        # Extract seniority
        for level, keywords in SENIORITY_KEYWORDS.items():
            for kw in keywords:
                if kw in combined:
                    slots.seniority = level
                    break
            if slots.seniority:
                break
        
        # Extract purpose
        for purpose, keywords in ASSESSMENT_PURPOSE.items():
            for kw in keywords:
                if kw in combined:
                    slots.purpose = purpose
                    break
        
        # Extract language requirements
        language_patterns = [
            (r"spanish|español", "Spanish"),
            (r"french|français", "French"),
            (r"german|deutsch", "German"),
            (r"portuguese|português", "Portuguese"),
            (r"chinese|mandarin", "Chinese"),
            (r"english", "English"),
        ]
        for pattern, lang in language_patterns:
            if re.search(pattern, combined):
                slots.language = lang
        
        # Extract volume
        volume_match = re.search(r'(\d+)\s+(candidates|agents|people|staff|hires)', combined)
        if volume_match:
            slots.volume = volume_match.group(0)
        elif "high.volume" in combined or "volume" in combined:
            slots.volume = "high volume"
        
        # The role, technical skills, and specific requests are best extracted
        # by the LLM as part of the agent call — rule-based extraction is too brittle
        # for diverse job descriptions. We provide the raw text to the LLM prompt.
        
        return slots
    
    def classify_intent(self, messages: list[Message]) -> str:
        """
        Classify the user's intent from the latest message + context.
        
        Returns one of:
            'initial' — first message, may need clarification
            'providing_info' — user is answering a clarification question
            'refine' — user wants to add/remove/change recommendations
            'compare' — user is asking about differences
            'confirm' — user is accepting the current list
            'off_topic' — not about SHL assessments
        """
        if not messages:
            return "initial"
        
        latest = messages[-1].content.lower()
        
        # Check for comparison intent
        compare_patterns = [
            r"what('?s|\s+is)\s+the\s+difference",
            r"compare\b",
            r"difference\s+between",
            r"how\s+(does|do|is)\s+.+\s+(differ|compare|versus|vs)",
            r"\bvs\.?\b",
        ]
        for pattern in compare_patterns:
            if re.search(pattern, latest):
                return "compare"
        
        # Check for refinement intent
        refine_patterns = [
            r"\badd\b",
            r"\bremove\b",
            r"\bdrop\b",
            r"\breplace\b",
            r"\bswap\b",
            r"\bupdate\b",
            r"\bchange\b",
            r"\bactually\b",
            r"\binstead\b",
            r"\balso\b.*\b(include|test|assess)",
        ]
        for pattern in refine_patterns:
            if re.search(pattern, latest):
                return "refine"
        
        # Check for confirmation
        confirm_patterns = [
            r"^(perfect|great|good|thanks|thank|ok|okay|confirmed?|locking|lock\s+it|that('?s| is)\s+(it|what|good|perfect|fine))",
            r"keep\s+(it|the|this)",
            r"that\s+covers\s+it",
            r"we('?ll|\s+will)\s+use",
            r"final\s+list",
        ]
        for pattern in confirm_patterns:
            if re.search(pattern, latest):
                return "confirm"
        
        # Check if it's the first message
        user_messages = [m for m in messages if m.role == "user"]
        if len(user_messages) <= 1:
            return "initial"
        
        return "providing_info"
    
    def has_recommendations_in_history(self, messages: list[Message]) -> bool:
        """Check if the agent has already provided recommendations."""
        for msg in messages:
            if msg.role == "assistant":
                # Look for recommendation patterns in assistant messages
                if any(
                    indicator in msg.content.lower()
                    for indicator in ["here are", "shortlist", "battery", "recommend", "assessment"]
                ):
                    return True
        return False
    
    def get_latest_user_message(self, messages: list[Message]) -> str:
        """Get the most recent user message."""
        for msg in reversed(messages):
            if msg.role == "user":
                return msg.content
        return ""