1#!/usr/bin/env python3
2"""Smoke tests for evidence_store.py CLI."""
3 
4import json
5import os
6import shutil
7import subprocess
8import sys
9import tempfile
10import unittest
11 
12SCRIPT = os.path.join(os.path.dirname(__file__), '..', 'scripts', 'evidence_store.py')
13 
14 
15def run_es(*args: str) -> dict | list:
16    """Run evidence_store.py with args, return parsed JSON from stdout."""
17    result = subprocess.run(
18        [sys.executable, SCRIPT, *args],
19        capture_output=True, text=True,
20    )
21    if result.returncode != 0:
22        raise RuntimeError(f'Exit {result.returncode}: {result.stderr}')
23    return json.loads(result.stdout)
24 
25 
26class TestInit(unittest.TestCase):
27    def test_creates_empty_file(self):
28        with tempfile.TemporaryDirectory() as d:
29            out = run_es('init', '--dir', d)
30            self.assertEqual(out['status'], 'ok')
31            path = os.path.join(d, 'evidence.jsonl')
32            self.assertTrue(os.path.exists(path))
33            self.assertEqual(os.path.getsize(path), 0)
34 
35 
36class TestAddEvidence(unittest.TestCase):
37    def setUp(self):
38        self.tmpdir = tempfile.mkdtemp()
39        run_es('init', '--dir', self.tmpdir)
40 
41    def tearDown(self):
42        shutil.rmtree(self.tmpdir, ignore_errors=True)
43 
44    def test_add_and_dedup(self):
45        ev = json.dumps({
46            'source_id': 'abcdef0123456789',
47            'quote': 'FActScore decomposes generation into atomic facts.',
48            'evidence_type': 'direct_quote',
49            'locator': 'page 3',
50            'retrieval_query': 'factuality evaluation methods',
51        })
52        out1 = run_es('add', '--json', ev, '--dir', self.tmpdir)
53        self.assertEqual(out1['status'], 'added')
54        self.assertEqual(len(out1['evidence_id']), 16)
55 
56        # Same quote -> duplicate
57        out2 = run_es('add', '--json', ev, '--dir', self.tmpdir)
58        self.assertEqual(out2['status'], 'duplicate')
59        self.assertEqual(out2['evidence_id'], out1['evidence_id'])
60 
61    def test_whitespace_normalization(self):
62        ev1 = json.dumps({
63            'source_id': 'abcdef0123456789',
64            'quote': '  FActScore   decomposes   generation  into atomic facts.  ',
65            'evidence_type': 'direct_quote',
66        })
67        ev2 = json.dumps({
68            'source_id': 'abcdef0123456789',
69            'quote': 'FActScore decomposes generation into atomic facts.',
70            'evidence_type': 'direct_quote',
71        })
72        out1 = run_es('add', '--json', ev1, '--dir', self.tmpdir)
73        out2 = run_es('add', '--json', ev2, '--dir', self.tmpdir)
74        # Should be same ID due to normalization
75        self.assertEqual(out1['evidence_id'], out2['evidence_id'])
76        self.assertEqual(out2['status'], 'duplicate')
77 
78    def test_different_sources_different_ids(self):
79        ev1 = json.dumps({
80            'source_id': 'aaaaaaaaaaaaaaaa',
81            'quote': 'Same quote text.',
82            'evidence_type': 'paraphrase',
83        })
84        ev2 = json.dumps({
85            'source_id': 'bbbbbbbbbbbbbbbb',
86            'quote': 'Same quote text.',
87            'evidence_type': 'paraphrase',
88        })
89        out1 = run_es('add', '--json', ev1, '--dir', self.tmpdir)
90        out2 = run_es('add', '--json', ev2, '--dir', self.tmpdir)
91        self.assertNotEqual(out1['evidence_id'], out2['evidence_id'])
92        self.assertEqual(out2['status'], 'added')
93 
94 
95class TestListAndExport(unittest.TestCase):
96    def setUp(self):
97        self.tmpdir = tempfile.mkdtemp()
98        run_es('init', '--dir', self.tmpdir)
99        # Add 3 evidence items from 2 sources
100        for src, quote in [
101            ('src_aaa', 'First quote from source A.'),
102            ('src_aaa', 'Second quote from source A.'),
103            ('src_bbb', 'Quote from source B.'),
104        ]:
105            run_es('add', '--json', json.dumps({
106                'source_id': src,
107                'quote': quote,
108                'evidence_type': 'direct_quote',
109            }), '--dir', self.tmpdir)
110 
111    def tearDown(self):
112        shutil.rmtree(self.tmpdir, ignore_errors=True)
113 
114    def test_list_all(self):
115        out = run_es('list', '--dir', self.tmpdir)
116        self.assertEqual(out['count'], 3)
117 
118    def test_list_filtered(self):
119        out = run_es('list', '--dir', self.tmpdir, '--source-id', 'src_aaa')
120        self.assertEqual(out['count'], 2)
121 
122        out = run_es('list', '--dir', self.tmpdir, '--source-id', 'src_bbb')
123        self.assertEqual(out['count'], 1)
124 
125    def test_export(self):
126        out = run_es('export', '--dir', self.tmpdir)
127        self.assertIsInstance(out, list)
128        self.assertEqual(len(out), 3)
129        # Each has required fields
130        for row in out:
131            self.assertIn('evidence_id', row)
132            self.assertIn('source_id', row)
133            self.assertIn('quote', row)
134            self.assertIn('evidence_type', row)
135            self.assertIn('captured_at', row)
136 
137 
138class TestEvidenceID(unittest.TestCase):
139    """Unit tests for compute_evidence_id."""
140 
141    @classmethod
142    def setUpClass(cls):
143        sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..', 'scripts'))
144        from evidence_store import compute_evidence_id, normalize_quote
145        cls.compute_id = staticmethod(compute_evidence_id)
146        cls.normalize = staticmethod(normalize_quote)
147 
148    def test_deterministic(self):
149        id1 = self.compute_id('src_a', 'test quote', 'page 1')
150        id2 = self.compute_id('src_a', 'test quote', 'page 1')
151        self.assertEqual(id1, id2)
152 
153    def test_locator_matters(self):
154        id1 = self.compute_id('src_a', 'test quote', 'page 1')
155        id2 = self.compute_id('src_a', 'test quote', 'page 2')
156        self.assertNotEqual(id1, id2)
157 
158    def test_normalize_whitespace(self):
159        self.assertEqual(
160            self.normalize('  hello   world  '),
161            'hello world',
162        )
163 
164 
165if __name__ == '__main__':
166    unittest.main()
167