import numpy as np import pandas as pd import pytest from drape.index import CatalogIndex, mmr, zscore from fakes import product def test_finds_same_colour_within_group(catalog, embedder): q = embedder.embed_images([product((205, 35, 25))])[0] res = catalog.search(q, k=3, groups=("shoes",)) assert res[0].item_id == "shoes-red" assert all(m.group == "shoes" for m in res) assert 0.9 < res[0].similarity <= 1.0001 def test_group_filter_union(catalog, embedder): q = embedder.embed_images([product((30, 60, 200))])[0] res = catalog.search(q, k=6, groups=("top", "outerwear"), diversify=False) assert {m.group for m in res} == {"top", "outerwear"} def test_unknown_group_returns_nothing(catalog, embedder): q = embedder.embed_images([product((30, 60, 200))])[0] assert catalog.search(q, groups=("spaceships",)) == [] def test_exclude(catalog, embedder): q = embedder.embed_images([product((200, 30, 30))])[0] first = catalog.search(q, k=1, groups=("bag",))[0] again = catalog.search(q, k=1, groups=("bag",), exclude=[first.row])[0] assert again.row != first.row def test_text_fusion_can_change_ranking(catalog, embedder): q = embedder.embed_images([product((200, 30, 30))])[0] t = embedder.embed_texts(["anything"])[0] base = [m.row for m in catalog.search(q, k=3, groups=("dress",), diversify=False)] heavy_text = [m.row for m in catalog.search(q, k=3, groups=("dress",), text_query=t, alpha=1.0, diversify=False)] expected = list(catalog.rows_for(("dress",))[np.argsort(-(catalog.embeddings[catalog.rows_for(("dress",))] @ t))]) assert heavy_text == expected assert base[0] == catalog.meta.index[catalog.meta.item_id == "dress-red"][0] def test_text_search_has_no_fake_percentages(catalog, embedder): t = embedder.embed_texts(["red dress"])[0] res = catalog.search_text(t, k=4) assert len(res) == 4 and all(m.similarity is None for m in res) assert "%" not in res[0].caption() def test_mmr_no_duplicates(): emb = np.eye(5, dtype=np.float32) emb = np.vstack([emb, emb[0]]) # row 5 duplicates row 0 picked = mmr(np.arange(6), np.array([1.0, 0.2, 0.2, 0.2, 0.2, 0.99]), emb, k=3) assert len(set(picked)) == 3 and not {0, 5} <= set(picked) def test_zscore_constant(): assert np.allclose(zscore(np.ones(4)), 0) def test_save_load_roundtrip(catalog, tmp_path): catalog.save(tmp_path) loaded = CatalogIndex.load(tmp_path) assert len(loaded) == len(catalog) and loaded.dim == catalog.dim assert np.allclose(loaded.embeddings, catalog.embeddings, atol=2e-3) assert loaded.manifest["embed_model"] == "fake" assert loaded.thumbnail(0).size == (64, 64) def test_rejects_mismatched_rows(): with pytest.raises(ValueError): CatalogIndex(np.zeros((2, 4)), pd.DataFrame({"item_id": ["a"], "title": ["a"], "group": ["top"]})) def test_caption_truncates(): from drape.index import Match m = Match(0, "x", "a" * 80, "top", 0.9, 0.876) assert m.caption().endswith("88% similar") and "…" in m.caption()