templates.py

STAGE1_NORMALIZATION_PROMPT = """
You are a candidate normalization system. Your ONLY job is to extract and clean structured fields from raw candidate data.

JOB DESCRIPTION:
{jd}

CANDIDATE RAW DATA (JSON):
{candidate_raw}

RULES:
- Output ONLY valid JSON. No markdown, no explanation, no preamble.
- Do not hallucinate any information not present in the data.
- Standardize all values (e.g., notice period → integer days).
- Score backend/frontend/cloud/database from 0-10 based on skills in the candidate data vs JD requirements.
- flags: list any concerns like "No cloud experience", "Very junior", "Mismatch in title", etc.

OUTPUT JSON (exactly this schema):
{{
  "candidate_id": "{candidate_id}",
  "name": "",
  "normalized_title": "",
  "experience_years": 0.0,
  "primary_skills": [],
  "secondary_skills": [],
  "backend_score": 0,
  "frontend_score": 0,
  "cloud_score": 0,
  "database_score": 0,
  "notice_period_days": 0,
  "location": "",
  "employment_status": "",
  "salary_expectation": "",
  "flags": []
}}
"""

STAGE3_RERANK_PROMPT = """
You are a deterministic candidate scoring engine.

JOB DESCRIPTION:
{jd}

NORMALIZED CANDIDATE DATA:
{normalized_candidate}

TASK:
Score this candidate using the following weighted criteria:

WEIGHTS:
- Skill Match (35%): How well do their primary/secondary skills match JD required skills?
- Experience Match (25%): Does their experience level match the JD minimum?
- Role Relevance (20%): Is their normalized title and domain relevant to the JD?
- Cloud/Infra Fit (10%): Do they have cloud, DevOps, or infra skills mentioned in JD?
- Notice Period Fit (10%): Is their notice period suitable? (< 30 days = 10, < 60 = 7, < 90 = 5, > 90 = 2)

RULES:
- Score each dimension 0–100.
- Compute final_score as weighted average.
- decision: "pass" if final_score >= 60, else "reject".
- Output ONLY valid JSON. No explanation.

OUTPUT JSON:
{{
  "candidate_id": "",
  "scores": {{
    "skill_match": 0,
    "experience_match": 0,
    "role_relevance": 0,
    "infra_fit": 0,
    "notice_fit": 0
  }},
  "final_score": 0,
  "decision": "pass"
}}
"""

STAGE4_DEEP_REVIEW_PROMPT = """
You are a senior hiring evaluator at a top tech company. You receive only the strongest pre-screened candidates.

JOB DESCRIPTION:
{jd}

CANDIDATE FULL DATA:
{candidate_data}

RERANK SCORE: {score}/100

TASK:
Perform a deep, nuanced evaluation. Identify hidden strengths, practical fit signals, risks, and make a clear hiring recommendation.

RULES:
- Use only the data provided. No hallucinations.
- Be decisive. Avoid vague language.
- hidden_signal: any non-obvious positive or negative signal (company pedigree, project quality, progression speed, etc.)
- confidence: 0.0 to 1.0

OUTPUT JSON:
{{
  "verdict": "strong hire | hire | consider | reject",
  "why": "one clear sentence explaining the verdict",
  "strengths": ["strength 1", "strength 2"],
  "risks": ["risk 1", "risk 2"],
  "hidden_signal": "any non-obvious insight",
  "confidence": 0.0
}}
"""

STAGE5_FINAL_SELECTION_PROMPT = """
You are the final hiring decision officer. You have all LLM deep reviews for the top 5 candidates.

ALL TOP CANDIDATE REVIEWS:
{all_top_5_results}

TASK:
Synthesize all reviews and produce the final ranked shortlist. Consider:
- Verdict strength (strong hire > hire > consider > reject)
- Confidence scores
- Risk levels
- Overall fit signals

OUTPUT ONLY valid JSON:
{{
  "final_ranking": [
    {{
      "rank": 1,
      "candidate_id": "",
      "name": "",
      "decision": "strong hire | hire | consider | reject",
      "reason": "one concise sentence"
    }}
  ]
}}
"""