tagging_schema.yaml

# LLM Tagging Schema
# Defines the output format for each problem during LLM tagging passes

output_per_problem:
  id: int                          # LeetCode problem ID

  # Pattern tagging — must use exact subtopic names from taxonomy.yaml
  # All weights sum to 1.0
  primary_subtopic:
    name: string                   # e.g. "Complement Search (Two Sum Pattern)"
    weight: float                  # e.g. 0.85

  secondary_subtopics:             # can be empty
    - name: string
      weight: float

  # Difficulty — numerical rating on Zerotrac scale
  # ~800 = trivial, ~1200 = easy, ~1600 = medium, ~2000 = hard, ~2400+ = elite
  # Calibrated against Zerotrac anchor problems included in each batch
  difficulty: int

  # Importance — how generalizable this specific problem is (0-1)
  # High: teaches a broadly applicable pattern/technique
  # Low: tests a niche trick or very specific variant
  importance: float

  # Interview plausibility — likelihood of appearing in a real interview (0-1)
  # High: classic interview problem, clean problem statement, tests core skills
  # Low: contest-style, obscure edge cases, unlikely to be asked
  interview_plausibility: float

  # Company type plausibility — how likely each company type asks this (0-1 each)
  company_plausibility:
    quant: float                   # trading firms, hedge funds
    faang: float                   # big tech
    mid: float                     # mid-size tech companies
    startup: float                 # startups, smaller companies

# Batch configuration
batching:
  batch_size: 25                   # problems per batch
  anchors_per_batch: 6             # Zerotrac-rated problems included for calibration
  anchor_spread:                   # one anchor near each difficulty tier
    - 800
    - 1200
    - 1600
    - 2000
    - 2400
    - 2800

# Pass 1 — Batch tagging
# System prompt: full taxonomy, subtopic names, scoring definitions, output format
# User message: batch of problems (content_clean + solution_clean + topics) + anchors
# Output: structured JSON array of tagged problems

# Pass 2 — Consistency check
# Group pass 1 results by subtopic, send each group back
# LLM checks: are difficulty ratings consistent within this subtopic?
# Are weights reasonable? Flag outliers for recalibration.