import tempfile
import unittest
import zipfile
from pathlib import Path
import numpy as np
from finalign.core import (
chunks,
clean_query,
dense_top_k,
fuse_passages,
hybrid_scores,
metrics,
min_max,
validate_query,
)
from finalign.index import unpack
from finalign.presentation import results_html
class RankingTests(unittest.TestCase):
def test_fusion_happens_before_document_aggregation(self):
# A has different best chunks in each variant; combining document maxima
# would incorrectly put A above B (1.0 vs 0.8).
scores, passages = fuse_passages([
(.7, {"A%0": 1., "A%1": 0., "B%0": .8}),
(.3, {"A%0": 0., "A%1": 1., "B%0": .8})])
self.assertAlmostEqual(scores["B"], .8)
self.assertAlmostEqual(scores["A"], .7)
self.assertEqual(passages["A"], "A%0")
def test_hybrid_retains_lexical_only_documents(self):
output = hybrid_scores({"a": 10, "b": 0}, {"b": 10, "c": 5, "d": 0})
self.assertEqual(output, {"a": 1., "b": 1., "c": .5, "d": 0.})
def test_constant_scores_and_missing_runs(self):
self.assertEqual(min_max({"a": 4, "b": 4}), {"a": 1., "b": 1.})
self.assertEqual(fuse_passages([]), ({}, {}))
with self.assertRaises(ValueError):
min_max({"a": float("nan")})
def test_exact_chunk_boundaries_and_empty_text(self):
items = list(chunks({"a": " ".join(str(i) for i in range(501)), "empty": ""}))
self.assertEqual([p for p, _ in items], ["a%0", "a%1", "a%2"])
self.assertEqual(items[-1][1], "500")
self.assertEqual(len(items[1][1].split()), 250)
def test_top_k_orders_dot_products(self):
matrix = np.array([[1, 0], [0, 1], [.5, .5]], dtype=np.float32)
scores = dense_top_k(matrix, np.array([0, 1], dtype=np.float32), ["a", "b", "c"], 2)
self.assertEqual(list(scores), ["b", "c"])
def test_metrics_match_independent_trec_evaluation(self):
import ir_measures
from ir_measures import AP, R, nDCG
qrels = [ir_measures.Qrel("q", d, rel) for d, rel in {"a": 1, "b": 1, "d": 0}.items()]
run = [ir_measures.ScoredDoc("q", d, 10 - i) for i, d in enumerate(["c", "a", "b"])]
expected = ir_measures.calc_aggregate([AP, R@10, nDCG@10], qrels, run)
actual = metrics(["c", "a", "b"], {"a": 1, "b": 1, "d": 0})
for name, metric in (("AP", AP), ("Recall@10", R@10), ("nDCG@10", nDCG@10)):
self.assertAlmostEqual(actual[name], expected[metric])
self.assertIsNone(metrics(["c"], {}))
def test_query_validation_preserves_notebook_normalization(self):
self.assertEqual(clean_query(" bonds:\n interest? "), "bonds interest?")
for query in ("", " ", "?!", "a" * 601):
with self.assertRaises(ValueError):
validate_query(query)
def test_corpus_html_is_escaped(self):
class FakeIndex:
def __init__(self):
self.documents = {"a": ' &
'}
def snippet(self, docno, passage=None):
return self.documents[docno]
output = results_html(FakeIndex(), [("a", 1.)], known={"qrels": {}})
self.assertNotIn("