"""Paths and shared constants for search evaluation."""
  
  from pathlib import Path
  
  _PKG_DIR = Path(__file__).resolve().parent
  _SCRIPTS_EVAL_DIR = _PKG_DIR.parent
  PROJECT_ROOT = _SCRIPTS_EVAL_DIR.parents[1]
  

  # Canonical English labels (must match LLM prompt output in prompts._CLASSIFY_TEMPLATE_EN)

  RELEVANCE_LV3 = "Fully Relevant"
  RELEVANCE_LV2 = "Mostly Relevant"
  RELEVANCE_LV1 = "Weakly Relevant"
  RELEVANCE_LV0 = "Irrelevant"

  VALID_LABELS = frozenset({RELEVANCE_LV3, RELEVANCE_LV2, RELEVANCE_LV1, RELEVANCE_LV0})

  # Useful label sets for binary diagnostic slices layered on top of graded ranking metrics.

  RELEVANCE_NON_IRRELEVANT = frozenset({RELEVANCE_LV3, RELEVANCE_LV2, RELEVANCE_LV1})
  RELEVANCE_STRONG = frozenset({RELEVANCE_LV3, RELEVANCE_LV2})

  
  # Graded relevance for ranking evaluation.
  # We use rel grades 3/2/1/0 and gain = 2^rel - 1, which is standard for NDCG-style metrics.
  RELEVANCE_GRADE_MAP = {

      RELEVANCE_LV3: 3,
      RELEVANCE_LV2: 2,
      RELEVANCE_LV1: 1,
      RELEVANCE_LV0: 0,

  }

  # 标准的gain计算方法：2^rel - 1
  # 但是是因为标注质量不是特别精确，因此适当降低 exact 和 high 的区分度

  RELEVANCE_GAIN_MAP = {

      # label: (2 ** grade) - 1
      label: grade

      for label, grade in RELEVANCE_GRADE_MAP.items()
  }

  # P(stop | relevance) for ERR (Expected Reciprocal Rank); cascade model (Chapelle et al., 2009).

  # p(t) = (2^t - 1) / 2^{max_grade}

  STOP_PROB_MAP = {

      RELEVANCE_LV3: 0.875,
      RELEVANCE_LV2: 0.375,
      RELEVANCE_LV1: 0.125,

      RELEVANCE_LV0: 0.0,

  }
  

  DEFAULT_ARTIFACT_ROOT = PROJECT_ROOT / "artifacts" / "search_evaluation"
  DEFAULT_QUERY_FILE = _SCRIPTS_EVAL_DIR / "queries" / "queries.txt"
  

  # Logging (``build_annotation_set.py`` / ``serve_eval_web.py`` → ``eval_framework.cli.main``)
  EVAL_LOG_DIR = PROJECT_ROOT / "logs"
  EVAL_VERBOSE_LOG_DIR = EVAL_LOG_DIR / "verbose"
  EVAL_LOG_FILE = EVAL_LOG_DIR / "eval.log"
  EVAL_VERBOSE_LOG_FILE = EVAL_VERBOSE_LOG_DIR / "eval_verbose.log"
  

  # Judge LLM (eval_framework only; override via CLI --judge-model / constructor kwargs)

  DEFAULT_JUDGE_MODEL = "qwen3.5-plus"
  DEFAULT_JUDGE_ENABLE_THINKING = False

  DEFAULT_JUDGE_DASHSCOPE_BATCH = False

  
  # Query-intent LLM (separate from judge; used once per query, injected into relevance prompts)

  DEFAULT_INTENT_MODEL = "qwen3-max"

  DEFAULT_INTENT_ENABLE_THINKING = True

  DEFAULT_JUDGE_BATCH_COMPLETION_WINDOW = "24h"
  DEFAULT_JUDGE_BATCH_POLL_INTERVAL_SEC = 10.0

  # --- Rebuild annotation pool (``build --force-refresh-labels``) ---
  # Flow: search recall pool (rerank_score=1, no rerank API) + rerank rest of corpus +
  # LLM labels in fixed-size batches along global order (see ``framework._annotate_rebuild_batches``).
  DEFAULT_SEARCH_RECALL_TOP_K = 200

  DEFAULT_RERANK_HIGH_THRESHOLD = 0.5
  DEFAULT_RERANK_HIGH_SKIP_COUNT = 1000
  DEFAULT_REBUILD_LLM_BATCH_SIZE = 50

  # At least this many LLM batches run before early-stop is considered.
  DEFAULT_REBUILD_MIN_LLM_BATCHES = 10
  # Hard cap on LLM batches per query (each batch labels up to ``DEFAULT_REBUILD_LLM_BATCH_SIZE`` docs).

  DEFAULT_REBUILD_MAX_LLM_BATCHES = 40

  
  # LLM early-stop (only after ``DEFAULT_REBUILD_MIN_LLM_BATCHES`` completed):

  # A batch is "bad" when **both** hold (strict inequalities; see ``framework._annotate_rebuild_batches``):
  #   - irrelevant_ratio > DEFAULT_REBUILD_IRRELEVANT_STOP_RATIO  (default 93.9%),

  #   - (Irrelevant + Weakly Relevant) / n > DEFAULT_REBUILD_IRREL_LOW_COMBINED_STOP_RATIO  (default 95.9%).

  # ``irrelevant_ratio`` = Irrelevant count / n; weak relevance is ``RELEVANCE_LV1`` ("Weakly Relevant").

  # Increment streak on consecutive bad batches; reset on any non-bad batch. Stop when streak
  # reaches ``DEFAULT_REBUILD_IRRELEVANT_STOP_STREAK`` (default 3).

  DEFAULT_REBUILD_IRRELEVANT_STOP_RATIO = 0.799

  DEFAULT_REBUILD_IRREL_LOW_COMBINED_STOP_RATIO = 0.959
  DEFAULT_REBUILD_IRRELEVANT_STOP_STREAK = 3