keyword_boosting_layer.py

"""
FILE: keyword_boosting_layer.py (ENHANCED VERSION)

PURPOSE:
- Advanced keyword-based boosting for final score refinement
- Exact match detection for precise results
- Location-based boosting
- Expanded category coverage
"""

import re
from difflib import SequenceMatcher

# ---------------- KEYWORD CATEGORIES ----------------

ROMANTIC_BOOST_WORDS = [
    "love", "romantic", "couple", "honeymoon", "beautiful", "sunset",
    "palace", "view", "lake", "memories", "historic", "architecture"
]

SPIRITUAL_BOOST = [
    "peace", "meditation", "spiritual", "calm", "holy", "divine", "quiet",
    "temple", "church", "mosque", "monastery", "shrine", "sacred"
]

FOOD_SPICY = [
    "spicy", "masala", "hot", "tangy", "flavour", "chilli", "pepper"
]

ADVENTURE_BOOST = [
    "trek", "hike", "camping", "rafting", "adventure", "mountain", "climb",
    "rappelling", "kayaking", "paragliding", "safari", "jungle"
]

NATURE_BOOST = [
    "waterfall", "lake", "forest", "valley", "river", "hill", "mountain",
    "beach", "nature", "wildlife", "scenic", "green", "meadow", "canyon"
]

HERITAGE_BOOST = [
    "fort", "palace", "museum", "ruins", "ancient", "heritage", "historic",
    "archaeological", "monument", "tomb", "temple", "architecture"
]

SHOPPING_BOOST = [
    "market", "bazaar", "silk", "saree", "shopping", "handicraft", "souvenir",
    "textile", "jewellery", "craft"
]

FOOD_GENERAL = [
    "authentic", "street food", "cafe", "bakery", "restaurant", "cuisine",
    "traditional", "local food", "breakfast", "lunch", "dinner", "eatery", "mess",
    "bhavan", "tiffin", "dhaba", "canteen", "bistro"
]

SWEET_BOOST = [
    "sweet", "dessert", "halwa", "mysore pak", "payasam", "kaja", "laddu",
    "barfi", "ghevar", "petha", "rasgulla", "rosogolla", "sandesh", "mishti",
    "jalebi", "gulab jamun", "double ka meetha", "qubani", "chhena poda", "bebinca"
]

SPECIFIC_DISH_BOOST = [
    "biryani", "dosa", "idli", "vada", "sambar", "fish curry", "thali", "meals",
    "kebab", "tikka", "tandoori", "butter chicken", "dal makhani", "chole bhature",
    "paratha", "kulcha", "pulla heady", "haleem", "nihari", "galouti", "kachori",
    "dhokla", "khandvi", "pav bhaji", "vada pav", "misal", "poh", "upma", "bisi bele bath",
    "ragi mudde", "appam", "stew", "puttu", "beef fry", "porotta", "litti chokha",
    "momos", "thukpa", "fish fry", "bamboo chicken", "pongal", "avial"
]

# Indian states and major cities for location matching
INDIAN_LOCATIONS = [
    "delhi", "mumbai", "bangalore", "bengaluru", "kolkata", "chennai", "hyderabad",
    "pune", "ahmedabad", "jaipur", "lucknow", "varanasi", "banaras", "agra",
    "goa", "kerala", "karnataka", "tamil nadu", "maharashtra", "rajasthan",
    "gujarat", "punjab", "himachal pradesh", "uttarakhand", "kashmir", "jammu",
    "meghalaya", "assam", "nagaland", "manipur", "tripura", "mizoram", "sikkim",
    "west bengal", "odisha", "chhattisgarh", "madhya pradesh", "uttar pradesh",
    "bihar", "jharkhand", "andhra pradesh", "telangana", "ladakh", "arunachal pradesh",
    "shillong", "guwahati", "imphal", "kohima", "gangtok", "darjeeling",
    "ongole", "tirupati", "vijayawada", "visakhapatnam", "vizag", "srisailam",
    "simhachalam", "lepakshi", "ahobilam", "mangalagiri", "srikalahasti", "kurnool",
    "warangal", "madurai", "rameshwaram", "thanjavur", "coimbatore", "mysore", "hampi",
    "coorg", "wayanad", "munnar", "alleppey", "alappuzha", "pondicherry", "puducherry"
]

# ---------------- HELPER FUNCTIONS ----------------

def similarity_ratio(a, b):
    """Calculate similarity between two strings (0.0 to 1.0)"""
    return SequenceMatcher(None, a.lower(), b.lower()).ratio()

def extract_location_from_query(query):
    """Extract potential location names from query"""
    query_lower = query.lower()
    found_locations = []
    for loc in INDIAN_LOCATIONS:
        if loc in query_lower:
            found_locations.append(loc)
    return found_locations

# ---------------- MAIN BOOSTING FUNCTION ----------------

def apply_keyword_boost(query, candidates):
    """
    Apply advanced keyword boosting to candidates.
    
    Args:
        query: User query string
        candidates: List of candidate dictionaries with 'name', 'text', 'region', 'state', etc.
    
    Returns:
        Sorted list of candidates with 'final_score' field
    """
    query_lower = query.lower()
    query_locations = extract_location_from_query(query_lower)
    
    for item in candidates:
        # Start with rerank score or embedding score
        base_score = float(item.get("rerank_score", item.get("embedding_score", 0)))
        boost = 0.0
        
        name = item.get("name", "").lower()
        text = item.get("text", "").lower()
        region = item.get("region", "").lower()
        state = str(item.get("state", "")).lower()
        
        # ============ EXACT/NEAR EXACT MATCH BOOST ============
        # Massive boost if query contains the exact name or very close variant
        if name in query_lower or query_lower in name:
            boost += 10.0
        elif similarity_ratio(query_lower, name) > 0.85:
            boost += 8.0
        
        # ============ LOCATION MATCH BOOST ============
        # Check if query mentions the location
        for loc in query_locations:
            if loc in region or loc in state:
                boost += 8.0
                break
        
        # ============ INTENT MATCH BOOST ============
        # "Soft Filter": Boost items that match the detected intent
        detected_intent = item.get("intent", "general")
        item_domain = str(item.get("domain", "")).lower()
        
        # Map 'nature' intent to 'nature' domain (and 'travel' as secondary)
        if detected_intent == "nature":
            if item_domain == "nature":
                boost += 3.0
            elif item_domain == "travel":
                boost += 1.5 # Secondary boost for travel items in nature query
        
        # Standard intent match
        elif detected_intent != "general" and detected_intent in item_domain:
            boost += 3.0
            
        # ============ CATEGORY KEYWORD BOOSTS ============
        # Count matching keywords and apply scaled boost
        
        # Romantic
        romantic_matches = sum(1 for word in ROMANTIC_BOOST_WORDS if word in query_lower and word in text)
        if romantic_matches > 0:
            boost += min(romantic_matches * 0.8, 3.0)
        
        # Spiritual
        spiritual_matches = sum(1 for word in SPIRITUAL_BOOST if word in query_lower and word in text)
        if spiritual_matches > 0:
            boost += min(spiritual_matches * 0.7, 2.5)
        
        # Spicy Food
        spicy_matches = sum(1 for word in FOOD_SPICY if word in query_lower and word in text)
        if spicy_matches > 0:
            boost += min(spicy_matches * 0.6, 2.0)
        
        # Adventure
        adventure_matches = sum(1 for word in ADVENTURE_BOOST if word in query_lower and word in text)
        if adventure_matches > 0:
            boost += min(adventure_matches * 0.7, 2.5)
        
        # Nature
        nature_matches = sum(1 for word in NATURE_BOOST if word in query_lower and word in text)
        if nature_matches > 0:
            boost += min(nature_matches * 0.6, 2.5)
        
        # Heritage
        heritage_matches = sum(1 for word in HERITAGE_BOOST if word in query_lower and word in text)
        if heritage_matches > 0:
            boost += min(heritage_matches * 0.6, 2.5)
        
        # Shopping
        shopping_matches = sum(1 for word in SHOPPING_BOOST if word in query_lower and word in text)
        if shopping_matches > 0:
            boost += min(shopping_matches * 0.5, 2.0)
        
        # Food General
        food_matches = sum(1 for word in FOOD_GENERAL if word in query_lower and word in text)
        if food_matches > 0:
            boost += min(food_matches * 0.5, 2.0)

        # Sweets/Desserts
        sweet_matches = sum(1 for word in SWEET_BOOST if word in query_lower and word in text)
        if sweet_matches > 0:
            boost += min(sweet_matches * 0.8, 3.0)

        # Specific Dishes
        dish_matches = sum(1 for word in SPECIFIC_DISH_BOOST if word in query_lower and word in text)
        if dish_matches > 0:
            boost += min(dish_matches * 1.0, 4.0)
        
        # Calculate final score
        item["final_score"] = base_score + boost
        item["boost_applied"] = boost
    
    # Sort by final score (descending)
    return sorted(candidates, key=lambda x: x["final_score"], reverse=True)