src/scoring_engine.py

"""
Scoring Engine

Takes match analysis and produces raw probability scores.
v1: LLM-based scoring with structured prompts
v2: Will use trained ML model with LLM features as inputs
"""

import json
from typing import Optional

from .feature_extractor import LLMClient, _extract_json
from .prompts.scoring import PROBABILITY_SCORING_PROMPT


class ScoringEngine:
    """Produces probability scores from match analysis."""

    def __init__(self, llm_client: Optional[LLMClient] = None):
        self.llm = llm_client or LLMClient()

    def score(self, match_analysis: dict) -> dict:
        """Generate probability scores from match analysis."""
        prompt = PROBABILITY_SCORING_PROMPT.format(
            match_analysis=json.dumps(match_analysis, indent=2),
        )
        response = self.llm.complete(prompt, temperature=0.1)
        return _extract_json(response)

    def score_deterministic(self, match_analysis: dict) -> dict:
        """
        Rule-based scoring fallback. No LLM needed.
        Useful for testing, offline mode, and as v2 baseline.
        """
        skill_match = match_analysis.get("skill_match_analysis", {})
        seniority = match_analysis.get("seniority_alignment", {})
        experience = match_analysis.get("experience_depth", {})
        context = match_analysis.get("context_fit", {})
        risks = match_analysis.get("risk_flags", [])

        # Shortlist scoring
        coverage = skill_match.get("coverage_ratio", 0.0)
        missing_critical = len(skill_match.get("missing_must_haves", []))

        shortlist_base = coverage * 70  # 0-70 from skill coverage

        # Seniority alignment bonus/penalty
        alignment = seniority.get("alignment", "aligned")
        seniority_mod = {
            "aligned": 15,
            "slightly_under": 5,
            "slightly_over": 0,
            "underqualified": -20,
            "overqualified": -10,
        }.get(alignment, 0)

        # Experience depth bonus
        depth = experience.get("depth_assessment", "adequate")
        depth_mod = {
            "exceptional": 15,
            "strong": 10,
            "adequate": 0,
            "insufficient": -20,
        }.get(depth, 0)

        shortlist_raw = max(5, min(92, shortlist_base + seniority_mod + depth_mod))

        # Apply hard caps
        shortlist_caps = []
        if missing_critical >= 3:
            shortlist_raw = min(shortlist_raw, 10)
            shortlist_caps.append("3+ missing critical skills -> cap 10%")
        elif missing_critical >= 2:
            shortlist_raw = min(shortlist_raw, 25)
            shortlist_caps.append("2 missing critical skills -> cap 25%")
        elif missing_critical >= 1:
            shortlist_raw = min(shortlist_raw, 45)
            shortlist_caps.append("1 missing critical skill -> cap 45%")

        # Offer acceptance scoring
        comp_fit = context.get("compensation_alignment_estimate", "unclear")
        comp_score = {
            "likely_aligned": 70,
            "unclear": 50,
            "likely_below": 30,
            "likely_above": 25,
        }.get(comp_fit, 50)

        location_fit = context.get("remote_fit", "possible")
        location_score = {
            "compatible": 80,
            "possible": 50,
            "incompatible": 15,
        }.get(location_fit, 50)

        stage_fit = context.get("company_stage_fit", "some_experience")
        stage_score = {
            "strong_experience": 75,
            "some_experience": 55,
            "no_experience": 35,
        }.get(stage_fit, 55)

        offer_raw = (comp_score * 0.35 + location_score * 0.25
                     + stage_score * 0.20 + 50 * 0.20)  # 50 for unknowns
        offer_raw = max(5, min(92, offer_raw))

        offer_caps = []
        if location_fit == "incompatible":
            offer_raw = min(offer_raw, 15)
            offer_caps.append("location incompatible -> cap 15%")
        if comp_fit == "likely_above":
            offer_raw = min(offer_raw, 35)
            offer_caps.append("comp likely above band -> cap 35%")

        # Retention scoring (uses risk flags)
        retention_raw = 65  # Base rate assumption

        high_risks = [r for r in risks if r.get("severity") == "high" and r.get("category") == "retention"]
        medium_risks = [r for r in risks if r.get("severity") == "medium" and r.get("category") == "retention"]

        retention_raw -= len(high_risks) * 20
        retention_raw -= len(medium_risks) * 10

        if alignment == "overqualified":
            retention_raw -= 15

        retention_raw = max(5, min(92, retention_raw))
        retention_caps = []

        # Overall hire probability (weighted blend, not multiplicative)
        interview_pass = min(80, shortlist_raw * 0.6 + 15)
        overall_raw = (
            shortlist_raw * 0.45
            + interview_pass * 0.25
            + offer_raw * 0.30
        )
        overall_raw = max(5, min(92, overall_raw))

        # Confidence
        data_signals = (
            len(skill_match.get("matched_must_haves", []))
            + len(skill_match.get("matched_preferred", []))
            + len(match_analysis.get("positive_signals", []))
            + len(match_analysis.get("risk_flags", []))  # risks are still data
            + (1 if experience.get("years_relevant") else 0)
            + (1 if context.get("industry_overlap") else 0)
            + (1 if context.get("compensation_alignment_estimate", "unclear") != "unclear" else 0)
        )

        if data_signals >= 12 and coverage >= 0.7:
            confidence = "high"
        elif data_signals >= 5 and coverage >= 0.4:
            confidence = "medium"
        else:
            confidence = "low"

        return {
            "shortlist_probability": {
                "value": round(shortlist_raw, 1),
                "hard_caps_applied": shortlist_caps,
            },
            "interview_pass_estimate": {
                "value": round(interview_pass, 1),
            },
            "offer_acceptance_probability": {
                "value": round(offer_raw, 1),
                "hard_caps_applied": offer_caps,
            },
            "retention_6m_probability": {
                "value": round(retention_raw, 1),
                "hard_caps_applied": retention_caps,
            },
            "overall_hire_probability": {
                "value": round(overall_raw, 1),
                "formula_inputs": {
                    "p_shortlist": round(shortlist_raw, 1),
                    "p_interview_pass": round(interview_pass, 1),
                    "p_offer_accept": round(offer_raw, 1),
                },
            },
            "confidence_level": confidence,
        }