scripts/verify_metrics.py

"""One-off cross-check of mcr's self-implemented metrics against pytrec_eval.

Run once to gain confidence, then we never depend on pytrec_eval again:

    pip install pytrec_eval
    PYTHONPATH=src python scripts/verify_metrics.py

Generates random runs/qrels and asserts nDCG@k / MAP@k / Recall@k / P@k agree
with pytrec_eval within a tolerance.
"""

from __future__ import annotations

import random
import sys

from mcr.eval.metrics import compute_metrics

K_VALUES = [1, 5, 10, 20, 100]
TOL = 1e-6


def random_case(n_queries=50, n_docs=200, seed=0):
    rng = random.Random(seed)
    qrels, run = {}, {}
    doc_ids = [f"d{i}" for i in range(n_docs)]
    for q in range(n_queries):
        qid = f"q{q}"
        n_rel = rng.randint(1, 5)
        rels = rng.sample(doc_ids, n_rel)
        qrels[qid] = {d: 1 for d in rels}
        # a run that retrieves a random subset (sometimes missing positives)
        retrieved = rng.sample(doc_ids, rng.randint(50, n_docs))
        run[qid] = {d: rng.random() for d in retrieved}
    return run, qrels


def pytrec_metrics(run, qrels, k_values):
    import pytrec_eval

    measures = set()
    for k in k_values:
        measures |= {f"ndcg_cut_{k}", f"map_cut_{k}", f"recall_{k}", f"P_{k}"}
    evaluator = pytrec_eval.RelevanceEvaluator(qrels, measures)
    per_q = evaluator.evaluate(run)
    agg = {}
    for k in k_values:
        for src, dst in [
            (f"ndcg_cut_{k}", f"ndcg@{k}"),
            (f"map_cut_{k}", f"map@{k}"),
            (f"recall_{k}", f"recall@{k}"),
            (f"P_{k}", f"precision@{k}"),
        ]:
            agg[dst] = sum(v[src] for v in per_q.values()) / len(per_q)
    return agg


def main() -> int:
    try:
        import pytrec_eval  # noqa: F401
    except ImportError:
        print("pytrec_eval not installed: pip install pytrec_eval", file=sys.stderr)
        return 2

    ok = True
    for seed in range(5):
        run, qrels = random_case(seed=seed)
        mine = compute_metrics(run, qrels, K_VALUES)
        theirs = pytrec_metrics(run, qrels, K_VALUES)
        for key, ref in theirs.items():
            diff = abs(mine[key] - ref)
            if diff > TOL:
                ok = False
                print(f"[seed {seed}] MISMATCH {key}: mine={mine[key]:.6f} pytrec={ref:.6f} (Δ={diff:.2e})")
    print("ALL METRICS MATCH ✓" if ok else "MISMATCHES FOUND ✗")
    return 0 if ok else 1


if __name__ == "__main__":
    raise SystemExit(main())