import pandas as pd

from uvw.data import leakage
from uvw.data.sources import MINOR_PATTERN, _artwork_key, _named_subject, church_rows, vangogh_rows
from uvw.data.splits import assign


def _fake_ner(entities):
    return lambda text: entities(text)


def test_minor_pattern_catches_minors_but_not_adult_terms():
    for p in ["a teen at the beach", "17 year old", "16yo", "little girl", "schoolgirl outfit",
              "young beautiful girl posing", "young girls"]:
        assert MINOR_PATTERN.search(p), p
    for p in ["a woman at the beach", "girlfriend portrait", "grown man"]:
        assert not MINOR_PATTERN.search(p), p


def test_named_subject_distinguishes_credit_from_subject():
    def ner(text):
        i = text.lower().find("jane doe")
        return [{"entity_group": "PER", "score": 0.99, "start": i, "end": i + 8}] if i >= 0 else []
    assert _named_subject("portrait of Jane Doe on a beach", ner)
    assert _named_subject("jane doe lying in a bathtub", ner)  # lowercase subject
    assert not _named_subject("a beach scene, art by Jane Doe", ner)
    assert not _named_subject("by jane doe, a collage", ner)  # credit at the start
    assert not _named_subject("a beach scene by greg rutkowski and Jane Doe", ner)
    assert not _named_subject("in the style of artgerm, greg rutkowski, jane doe", ner)


def test_artwork_aliases_group_alternative_titles():
    assert _artwork_key("Starry Night by Vincent van Gogh") == _artwork_key("The Starry Night by Vincent van Gogh")
    assert _artwork_key("Bedroom in Arles by Vincent van Gogh") == _artwork_key("The Bedroom by Vincent van Gogh")
    assert _artwork_key("original replica of The Starry Night") == _artwork_key("Starry Night by X")


def _frame():
    rows = [r.__dict__ for r in vangogh_rows() + church_rows()]
    return pd.DataFrame(rows)


def test_splits_deterministic_and_template_disjoint():
    a, b = assign(_frame()), assign(_frame())
    pd.testing.assert_frame_equal(a, b)
    assert leakage.template_overlap(a).empty
    # Paired designs: every artist / class for one subject or context lands in the same split.
    assert a.groupby("template_id").split.nunique().max() == 1
    assert set(a.split) == {"dev", "val", "test"}


def test_leakage_detects_injected_leak():
    df = assign(_frame())
    leaked = df.iloc[[0]].copy()
    leaked["split"] = "test" if df.split.iloc[0] != "test" else "dev"
    bad = pd.concat([df, leaked.assign(prompt_id="leak")])
    assert not leakage.template_overlap(bad).empty
    assert not leakage.exact_duplicates(bad).empty


def test_summarise_clusters_and_agreement():
    from uvw.evaluate import summarise

    rows = [{"role": "target", "category": "a", "group_id": f"g{i // 3}", "target_hit": i % 2 == 0,
             "second_hit": i % 2 == 0, "clip_score": 30.0} for i in range(12)]
    s = summarise(rows)
    t = s["role:target"]
    assert t["n_images"] == 12 and t["n_clusters"] == 4
    assert t["target_hit"]["ci95"][0] <= t["target_hit"]["mean"] <= t["target_hit"]["ci95"][1]
    assert s["detector_agreement"] == 1.0


def test_expand_seeds_keeps_source_seed_first():
    from uvw.evaluate import expand_seeds

    df = pd.DataFrame([{"prompt_id": "p", "text": "x", "seed": 42}])
    seeds = [it.seed for _, it in expand_seeds(df, 3)]
    assert seeds[0] == 42 and len(set(seeds)) == 3
    assert seeds == [it.seed for _, it in expand_seeds(df, 3)]