app/services/objection.py

"""Author RAG Chatbot SaaS — Purchase Objection Detector.

Classifies reader pushback so the LLM can counter the SPECIFIC concern
instead of giving a generic reply. An objection is a buying signal —
the reader is weighing the purchase, not leaving.

RULE: Objections are modifiers on top of intent — they never block the pipeline.
RULE: Objection turns are never cached (responses are session-personalized).
"""

import re

import structlog

logger = structlog.get_logger(__name__)

# Objection type → trigger phrases. Ordered most-specific first inside each list.
_OBJECTION_PATTERNS: dict[str, tuple[str, ...]] = {
    # "Too expensive / not worth the money"
    "price": (
        "too expensive", "too pricey", "costs too much", "cost too much",
        "can't afford", "cannot afford", "cheaper", "expensive for",
        "not worth the money", "waste of money", "that price", "overpriced",
    ),
    # "No time to read / too long"
    "time": (
        "no time to read", "don't have time", "do not have time",
        "too long", "too many pages", "never finish", "takes too long",
        "busy to read", "too busy",
    ),
    # "Not my topic / no interest in the subject"
    "relevance": (
        "no interest in", "not interested in", "not into",
        "don't care about", "do not care about", "not my thing",
        "not for me", "doesn't apply to me", "why would i need",
        "don't like politics", "boring topic", "sounds boring",
    ),
    # "Is it actually good? Prove it."
    "skepticism": (
        "is it actually good", "is it any good", "why should i trust",
        "prove it", "sounds like hype", "everyone says that",
        "heard that before", "what makes it different", "how is it different",
        "why is it special", "what's so special",
    ),
    # "I only read <other genre>"
    "genre": (
        "i only read", "i usually read", "i prefer fiction", "i prefer non-fiction",
        "not my genre", "don't read this genre", "do not read this genre",
    ),
    # "I already read something like this"
    "alternatives": (
        "already read", "read something similar", "similar book",
        "same as", "just like every other", "nothing new",
    ),
}

# Generic value-challenge phrasing → relevance objection when no specific type hits
_GENERIC_OBJECTION_MARKERS: tuple[str, ...] = (
    "why should i buy", "why would i buy", "why should i get",
    "convince me", "why bother", "give me one reason",
    "not sure if i should", "should i really",
)

_WS_RE = re.compile(r"\s+")


def detect_objection(query: str) -> str | None:
    """Detect and classify a purchase objection in the reader's message.

    Args:
        query: Sanitized user message.

    Returns:
        Objection type ('price', 'time', 'relevance', 'skepticism',
        'genre', 'alternatives') or None if no objection detected.
    """
    q = _WS_RE.sub(" ", query.lower().strip())

    for objection_type, phrases in _OBJECTION_PATTERNS.items():
        if any(phrase in q for phrase in phrases):
            logger.debug("Objection detected", type=objection_type, query=q[:60])
            return objection_type

    if any(marker in q for marker in _GENERIC_OBJECTION_MARKERS):
        logger.debug("Generic value objection detected", query=q[:60])
        return "relevance"

    return None


def count_prior_objections(history: list[dict]) -> int:
    """Count how many previous user messages contained an objection.

    Used to vary the counter-angle — never repeat the same persuasion
    twice, and soften after repeated hesitation.

    Args:
        history: Session history as [{role, content}] dicts.

    Returns:
        Number of prior user turns that contained an objection.
    """
    count = 0
    for message in history:
        if message.get("role") != "user":
            continue
        if detect_objection(message.get("content", "")):
            count += 1
    return count