src/prompts/feature_extraction.py

"""
LLM prompt templates for structured feature extraction.

These prompts are designed to extract factual, observable features
from job descriptions and resumes WITHOUT making subjective judgments.
The scoring happens in a separate layer.
"""

ROLE_FEATURE_EXTRACTION_PROMPT = """You are a structured data extraction system. Extract factual features from this job description.
Do NOT make subjective quality judgments. Only extract what is explicitly stated or clearly implied.

JOB DESCRIPTION:
{job_description}

COMPANY CONTEXT:
- Stage: {company_stage}
- Industry: {industry}
- Compensation Band: {compensation_band}
- Location: {location}
- Remote Type: {remote_type}

Extract the following as JSON. Use null for anything not determinable:

{{
  "inferred_title": "string - job title",
  "seniority_band": "IC1_entry | IC2_mid | IC3_senior | IC4_staff | IC5_principal | M1_manager | M2_director | M3_vp",
  "core_competencies": [
    {{
      "skill": "string",
      "proficiency_required": "basic | intermediate | advanced | expert",
      "criticality": "must_have | strong_prefer | nice_to_have",
      "years_implied": null or number
    }}
  ],
  "scope_complexity": {{
    "system_scale": "small | medium | large | massive",
    "ambiguity_level": "low | medium | high | very_high",
    "cross_functional_dependency": "none | low | medium | high",
    "decision_authority": "executor | contributor | driver | owner"
  }},
  "domain_specificity": 0.0 to 1.0,
  "team_context": {{
    "team_size": null or integer,
    "team_maturity": "new | growing | established | restructuring",
    "reporting_structure": "individual_contributor | tech_lead | people_manager"
  }},
  "growth_trajectory_required": true or false,
  "hard_requirements": ["list of absolute must-haves mentioned"],
  "location_requirement": "strict_onsite | hybrid_required | remote_ok | fully_remote"
}}

Rules:
- Only mark "must_have" if the JD explicitly says required/must/essential
- If years of experience mentioned, extract the number
- domain_specificity: 0.0 = generic role, 1.0 = requires deep niche expertise
- Be conservative: when uncertain between two levels, choose the lower one
"""

CANDIDATE_FEATURE_EXTRACTION_PROMPT = """You are a structured data extraction system. Extract factual features from this resume.
Do NOT make subjective quality judgments. Only extract what is explicitly stated or clearly implied.

RESUME:
{resume_text}

Extract the following as JSON. Use null for anything not determinable.
For experience entries, extract ALL positions listed.

{{
  "candidate_name": "string or null",
  "experience_profile": {{
    "total_years": number or null,
    "positions": [
      {{
        "title": "string",
        "company": "string",
        "duration_months": number or null,
        "start_year": number or null,
        "end_year": number or null,
        "is_current": boolean,
        "seniority_estimate": "IC1_entry | IC2_mid | IC3_senior | IC4_staff | IC5_principal | M1_manager | M2_director | M3_vp",
        "key_technologies": ["list"],
        "quantified_achievements": ["list of achievements with numbers"],
        "scope_indicators": ["team size", "user count", "revenue impact", etc.]
      }}
    ]
  }},
  "skills_mentioned": ["complete list of technical and professional skills"],
  "education": [
    {{
      "degree": "string",
      "field": "string",
      "institution": "string",
      "year": number or null
    }}
  ],
  "certifications": ["list"],
  "trajectory_indicators": {{
    "career_velocity": "slow | steady | fast | exceptional",
    "scope_progression": "flat | linear | accelerating | decelerating",
    "role_type_consistency": "consistent | pivoting | scattered"
  }},
  "stability_indicators": {{
    "avg_tenure_months": number,
    "shortest_tenure_months": number,
    "longest_tenure_months": number,
    "tenure_trend": "decreasing | stable | increasing",
    "total_moves": number
  }},
  "impact_evidence": {{
    "quantified_achievements_count": number,
    "max_scope_of_impact": "individual | team | org | company | industry",
    "technical_depth_signals": number,
    "leadership_signals": number
  }},
  "industry_experience": ["list of industries worked in"],
  "company_stage_experience": ["startup | scaleup | enterprise | public | government | etc."],
  "remote_experience_evident": boolean,
  "employment_gaps_months": [list of gap durations, empty if none detected]
}}

Rules:
- Calculate tenure from dates when available, estimate otherwise
- career_velocity: based on how quickly titles/scope increased
- Count only EXPLICIT numbers as quantified_achievements
- Do not infer skills not mentioned
- For scope_of_impact, use the HIGHEST level evidenced
- Do not use university name/prestige as a signal (bias mitigation)
"""

MATCH_ANALYSIS_PROMPT = """You are an analytical matching system. Given extracted features from a role and a candidate,
identify specific matches, gaps, and risks. Be factual and conservative.

ROLE FEATURES:
{role_features}

CANDIDATE FEATURES:
{candidate_features}

COMPANY CONTEXT:
- Stage: {company_stage}
- Industry: {industry}
- Compensation Band: {compensation_band}
- Remote Type: {remote_type}

Produce the following JSON:

{{
  "skill_match_analysis": {{
    "matched_must_haves": [
      {{"skill": "...", "candidate_proficiency": "...", "required_proficiency": "...", "gap": "none | minor | significant"}}
    ],
    "missing_must_haves": ["list of must-have skills not found in resume"],
    "matched_preferred": ["list of nice-to-have skills found"],
    "adjacent_skills": ["skills candidate has that are related but not exact match"],
    "coverage_ratio": 0.0 to 1.0
  }},
  "seniority_alignment": {{
    "role_seniority": "...",
    "candidate_seniority": "...",
    "alignment": "underqualified | slightly_under | aligned | slightly_over | overqualified",
    "band_difference": integer
  }},
  "experience_depth": {{
    "years_required": number or null,
    "years_relevant": number,
    "depth_assessment": "insufficient | adequate | strong | exceptional",
    "recency_of_relevant_experience": "current | recent_2yr | moderate_5yr | stale"
  }},
  "context_fit": {{
    "company_stage_fit": "no_experience | some_experience | strong_experience",
    "industry_overlap": boolean,
    "remote_fit": "incompatible | possible | compatible",
    "compensation_alignment_estimate": "likely_below | unclear | likely_aligned | likely_above"
  }},
  "risk_flags": [
    {{"risk": "string", "severity": "low | medium | high", "category": "shortlist | offer_acceptance | retention", "evidence": "string"}}
  ],
  "positive_signals": [
    {{"signal": "string", "strength": "moderate | strong | exceptional", "evidence": "string"}}
  ],
  "missing_information": ["list of important data points that could not be determined"]
}}

Rules:
- coverage_ratio = matched_must_haves / total_must_haves
- Be strict: adjacent skills are NOT matches
- Only flag risks with specific evidence from the data
- Do NOT use demographic information, university prestige, or personal characteristics
- Focus on: skills, experience, trajectory, stability, scope, impact
"""