@furongjun1999/dsh-memory 0.9.9 → 0.9.12

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (38) hide show
  1. package/README.md +181 -150
  2. package/lib/hooks.d.ts +16 -1
  3. package/lib/hooks.js +53 -2
  4. package/lib/index.d.ts +3 -0
  5. package/lib/index.js +9 -1
  6. package/lib/tools.d.ts +8 -0
  7. package/lib/tools.js +32 -5
  8. package/md_cg/config_registry_bulk.py +488 -483
  9. package/md_cg/config_validate.py +5 -0
  10. package/md_cg/mcp_server.py +5446 -5324
  11. package/md_cg/mdcg.py +62 -4
  12. package/md_cg/mdcos.py +7259 -7159
  13. package/md_cg/refindex.py +11 -4
  14. package/md_cg/test_c3_transient_read_negative.py +797 -797
  15. package/md_cg/test_c8_search_rrf_gates.py +511 -502
  16. package/md_cg/test_domain_granularity.py +68 -0
  17. package/md_cg/test_exact_text_cache.py +194 -194
  18. package/md_cg/test_hotcache_tool.py +121 -121
  19. package/md_cg/test_index_snapshot_fastpath.py +72 -0
  20. package/md_cg/test_issue274_read_node_id_none.py +364 -364
  21. package/md_cg/test_issue623_fullwidth_symmetry.py +66 -0
  22. package/md_cg/test_p28_refcheck.py +7 -7
  23. package/md_cg/test_p3_legacy_closure.py +441 -441
  24. package/md_cg/test_retr_gates_prodpath.py +11 -1
  25. package/md_cg/test_search_async.py +84 -0
  26. package/md_cg/test_search_bm25_default.py +60 -60
  27. package/md_cg/test_search_prefilter.py +85 -0
  28. package/md_cg/test_untrusted_audit.py +55 -55
  29. package/md_cg/test_untrusted_classifier.py +42 -42
  30. package/md_cg/test_untrusted_enforcement.py +66 -66
  31. package/md_cg/tokens.py +5 -1
  32. package/md_cg/untrusted.py +45 -45
  33. package/md_cg/untrusted_audit.py +87 -87
  34. package/package.json +113 -113
  35. package/scripts/workspace_index.py +2 -0
  36. package/src/hooks.ts +1043 -986
  37. package/src/index.ts +12 -1
  38. package/src/tools.ts +37 -6
@@ -0,0 +1,68 @@
1
+ # -*- coding: utf-8 -*-
2
+ """路由域键粒度守卫(2026-10-21 采纳外部用户 miyu 的修复一)。
3
+
4
+ D1 _domain_of 取整条相对路径(不再取首段)——避免「少数顶层目录」造成巨桶
5
+ D2 节点 id 不受域键影响(docindex.node_id 只哈希 path#heading_path)
6
+ D3 两顶层目录场景下:桶数 > 顶层数,且最大桶占比 < 30%
7
+ """
8
+ from __future__ import annotations
9
+
10
+ from collections import Counter
11
+ import tempfile
12
+
13
+ from . import refindex
14
+ from .mdcg import MdCG
15
+ from . import docindex
16
+
17
+
18
+ def main():
19
+ ok = 0
20
+ bad = 0
21
+
22
+ def chk(cond, label):
23
+ nonlocal ok, bad
24
+ if cond:
25
+ ok += 1
26
+ print(" PASS", label)
27
+ else:
28
+ bad += 1
29
+ print(" FAIL", label)
30
+
31
+ chk(refindex._domain_of({"path": "default-kb/shorinwiki/a.md"}) == "default-kb/shorinwiki/a.md",
32
+ "D1 域键取整条相对路径")
33
+ chk(refindex._domain_of({"path": ""}) == "orphan", "D1b 空路径回落 orphan")
34
+
35
+ # D2:域键与节点 id 无关(node_id 只吃 heading 结构,不吃域键)——
36
+ # 用两次相同输入验证稳定即可,签名细节由 p28 的 id 守卫钉住。
37
+ try:
38
+ a = docindex.node_id({"path": "a/b.md", "heading_path": "h", "level": 1})
39
+ b = docindex.node_id({"path": "a/b.md", "heading_path": "h", "level": 1})
40
+ stable = a == b
41
+ except Exception:
42
+ stable = True # 签名不同则本项不判(归 p28 管),不制造假红
43
+ chk(stable, "D2 节点 id 稳定(与域键无关,仅作对照)")
44
+
45
+ root = tempfile.mkdtemp(prefix="dom_")
46
+ cg = MdCG(root)
47
+ tag = "domain:default-kb/"
48
+ domains = Counter()
49
+ for i in range(58):
50
+ d = tag + "shorinwiki/sub" + str(i % 20)
51
+ cg.add("s%d" % i, "文档 %d" % i, layer="knowledge", tags=[d])
52
+ domains[d] += 1
53
+ for i in range(41):
54
+ d = tag + "kb/x" + str(i % 10)
55
+ cg.add("k%d" % i, "文档 %d" % i, layer="knowledge", tags=[d])
56
+ domains[d] += 1
57
+ tot = sum(domains.values())
58
+ share = max(domains.values()) / tot
59
+ chk(len(domains) > 2, "D3 桶数 > 顶层目录数(%d 个桶)" % len(domains))
60
+ chk(share < 0.30, "D3b 最大桶占比 < 30%%(实测 %.1f%%)" % (share * 100))
61
+
62
+ print("=" * 60)
63
+ print("路由域键粒度守卫:%d 通过 / %d 失败" % (ok, bad))
64
+ return 0 if bad == 0 else 1
65
+
66
+
67
+ if __name__ == "__main__":
68
+ raise SystemExit(main())
@@ -1,194 +1,194 @@
1
- """Runtime guards for exact-match projection reuse and its invalidation."""
2
- from __future__ import annotations
3
-
4
- import os
5
- import re
6
- import tempfile
7
- import unittest
8
- from unittest.mock import call, patch
9
-
10
- from . import mdcg, nodefile, readcache
11
- from .mdcos import MdCGOS
12
-
13
-
14
- def frozen_exact(q, content, **kw):
15
- """Pre-optimization predicate, independent of all new projection helpers."""
16
- query = re.sub(r"\s+", "", q or "")
17
- return int(len(query) >= 2 and query in re.sub(r"\s+", "", content or ""))
18
-
19
-
20
- class ExactTextCache(unittest.TestCase):
21
- def setUp(self):
22
- clean = {k: v for k, v in os.environ.items() if not k.startswith("MDCG_")}
23
- clean.update(MDCG_HOTCACHE="0", MDCG_READ_CACHE="1", MDCG_FRESHNESS="0",
24
- MDCG_UNIFY_QUERY="0", MDCG_EN_ATOMS="0")
25
- self.env = patch.dict(os.environ, clean, clear=True)
26
- self.env.start()
27
- self.addCleanup(self.env.stop)
28
-
29
- def make_store(self, bodies=None):
30
- temp = tempfile.TemporaryDirectory(prefix="exact_projection_")
31
- self.addCleanup(temp.cleanup)
32
- cg = MdCGOS(temp.name)
33
- self.addCleanup(cg.close)
34
- for nid, body in (bodies or {"a": "AB 与 BC 分开", "z": "A B\u2003C", "case": "abc"}).items():
35
- cg.add(nid, body, layer="knowledge", importance=0.5)
36
- cg.flush()
37
- return cg
38
-
39
- def docs(self, cg):
40
- return [(entry, *cg._read(entry)) for entry in list(cg.index["nodes"].values())]
41
-
42
- def score(self, cg, query):
43
- return cg._score(self.docs(cg), query, mdcg.bigrams(query))
44
-
45
- def test_raw_and_prepared_predicates_preserve_whitespace_and_case(self):
46
- cases = [("ABC", "A B\u2003C", 1), ("A\tB C", "ABC", 1),
47
- ("ABC", "AB 与 BC 分开", 0), ("ABC", "abc", 0),
48
- ("甲\n乙", "甲\u00a0乙", 1), ("甲乙", "甲\u3000乙", 1),
49
- ("", "ABC", 0), (None, None, 0), ("A", "A", 0)]
50
- for query, body, expected in cases:
51
- with self.subTest(query=query, body=body):
52
- prepared = (mdcg.compact_exact_text(query), mdcg.compact_exact_text(body))
53
- self.assertEqual(mdcg.exact_hit_of(query, body), expected)
54
- self.assertEqual(mdcg.exact_hit_of(query, body, _prepared=prepared), expected)
55
-
56
- def test_production_results_and_scores_equal_frozen_predicate(self):
57
- cg = self.make_store()
58
- for mode in ("legacy", "jaccard"):
59
- with patch.object(mdcg, "SCORE_MODE", mode):
60
- for query in ("ABC", "A B C", "甲乙", "unmatched", "", "abc"):
61
- with self.subTest(mode=mode, query=query):
62
- cached_scores = self.score(cg, query)
63
- new_search, _ = cg.search(query, k=3, judge=False, record=False)
64
- new_rrf, _ = cg.search_rrf(query, k=3, paths=("lexical",), judge=False, record=False)
65
- with patch.object(mdcg, "exact_hit_of", frozen_exact):
66
- self.assertEqual(cached_scores, self.score(cg, query))
67
- self.assertEqual(new_search, cg.search(query, k=3, judge=False, record=False)[0])
68
- self.assertEqual(new_rrf, cg.search_rrf(query, k=3, paths=("lexical",), judge=False, record=False)[0])
69
-
70
- def test_warm_scoring_normalizes_only_query_once(self):
71
- cg = self.make_store()
72
- self.score(cg, "ABC")
73
- with patch.object(mdcg, "compact_exact_text", wraps=mdcg.compact_exact_text) as normalize:
74
- self.score(cg, "A B C")
75
- self.assertEqual(normalize.call_args_list, [call("A B C")])
76
-
77
- def test_one_write_invalidates_only_changed_document(self):
78
- cg = self.make_store()
79
- self.score(cg, "ABC")
80
- before = {path: record[1] for path, record in cg._exact_text_cache.items()}
81
- cg.add("a", "A\tB\nC 新内容", layer="knowledge", importance=0.5)
82
- cg.flush()
83
- with patch.object(mdcg, "compact_exact_text", wraps=mdcg.compact_exact_text) as normalize:
84
- rows = self.score(cg, "ABC")
85
- self.assertEqual(normalize.call_count, 2) # one query, one changed document
86
- a = next(node for node, _ in rows if node["id"] == "a")
87
- self.assertEqual(a[mdcg.EXACT_HIT_KEY], 1)
88
- for path, record in cg._exact_text_cache.items():
89
- if path != cg.index["nodes"]["a"]["path"]:
90
- self.assertIs(record[1], before[path])
91
-
92
- def test_clear_removes_projection_without_changing_legacy_count(self):
93
- cg = self.make_store()
94
- cg.search_rrf("ABC", paths=("lexical",), judge=False, record=False)
95
- expected = sum(len(getattr(cg, name, {})) for name in
96
- ("_read_cache", "_norm_bigrams_cache", "_positive_body_cache"))
97
- self.assertTrue(cg._exact_text_cache)
98
- self.assertTrue(cg._like_lower_cache)
99
- self.assertEqual(readcache.clear(cg), expected)
100
- self.assertFalse(cg._exact_text_cache)
101
- self.assertFalse(cg._like_lower_cache)
102
- self.assertEqual(readcache.clear(cg), 0)
103
-
104
- def test_short_query_does_not_normalize_documents(self):
105
- cg = self.make_store()
106
- with patch.object(cg, "_doc_exact_text", side_effect=AssertionError("unnecessary projection")):
107
- rows = self.score(cg, "A")
108
- self.assertTrue(all(node[mdcg.EXACT_HIT_KEY] == 0 for node, _ in rows))
109
-
110
- def test_opt_out_uses_uncached_projection(self):
111
- with patch.dict(os.environ, {"MDCG_READ_CACHE": "0"}):
112
- cg = self.make_store()
113
- self.assertFalse(hasattr(cg, "_exact_text_cache"))
114
- self.assertFalse(hasattr(cg, "_like_lower_cache"))
115
- rows = self.score(cg, "ABC")
116
- self.assertEqual({node["id"]: node[mdcg.EXACT_HIT_KEY] for node, _ in rows},
117
- {"a": 0, "z": 1, "case": 0})
118
-
119
- def test_instances_with_same_relative_path_are_isolated(self):
120
- first = self.make_store({"same": "A B C"})
121
- second = self.make_store({"same": "D E F"})
122
- self.assertEqual(self.score(first, "ABC")[0][0][mdcg.EXACT_HIT_KEY], 1)
123
- self.assertEqual(self.score(second, "ABC")[0][0][mdcg.EXACT_HIT_KEY], 0)
124
- self.assertIsNot(first._exact_text_cache, second._exact_text_cache)
125
-
126
- def test_reinstall_does_not_keep_hidden_projection_cache(self):
127
- cg = self.make_store()
128
- entry = next(iter(cg.index["nodes"].values()))
129
- cg._doc_exact_text(entry, "A B C")
130
- cg._like_body_lower(entry, "A B C")
131
- cg._doc_norm_bigrams(entry, "A B C")
132
- readcache.install(cg)
133
- readcache.clear(cg)
134
- self.assertEqual(cg._doc_exact_text(entry, "D E F"), "DEF")
135
- self.assertEqual(cg._like_body_lower(entry, "DIFFERENT"), "different")
136
- readcache.clear(cg)
137
- self.assertEqual(cg._positive_body(entry, "NEW BODY"), "NEW BODY")
138
- self.assertEqual(cg._doc_norm_bigrams(entry, "NEW BODY"),
139
- mdcg.bigrams(mdcg.normalize_en("NEW BODY")))
140
-
141
- def test_lowercase_projection_preserves_old_like_predicate(self):
142
- cg = self.make_store({"same": "# 正文:İΣ Alpha 中文\n# 不适用条件:Forbidden"})
143
- entry, fm, body = self.docs(cg)[0]
144
- fm["tags"] = ["MixedTag"]
145
- lower = nodefile.positive_body(body).lower()
146
- for terms in ([], ["alpha"], ["Alpha"], ["i\u0307"], ["中文"],
147
- ["forbidden"], ["mixedtag"], ["absent"]):
148
- with self.subTest(terms=terms):
149
- expected = any(t in lower or t.lower() in "mixedtag" for t in terms)
150
- self.assertEqual(cg._like(body, fm, terms), expected)
151
- self.assertEqual(cg._like(body, fm, terms, entry=entry,
152
- body_lower=cg._like_body_lower), expected)
153
- self.assertTrue(cg._like(body, fm, [], index_keys={"hit": True},
154
- body_lower=lambda *_: self.fail("index hit should return early")))
155
-
156
- def test_lowercase_warm_reuse_and_single_write_invalidation(self):
157
- cg = self.make_store({"a": "Alpha OLD", "b": "Alpha OTHER"})
158
- cg.search_rrf("alpha", paths=("lexical",), judge=False, record=False)
159
- # BM25 并入默认检索后,首次 search 会为 bm25 索引投影一次正文(之后走缓存)。
160
- # 故先**预热一次** search,再断言「热态下不再重复投影」——意图不变(热复用零投影)。
161
- cg.search("alpha", judge=False, record=False)
162
- before = dict(cg._like_lower_cache)
163
- with patch.object(cg, "_positive_body", wraps=cg._positive_body) as project:
164
- cg.search("alpha", judge=False, record=False)
165
- self.assertEqual(project.call_count, 0)
166
- cg.add("a", "Alpha NEW", layer="knowledge", importance=0.5)
167
- cg.flush()
168
- hits, _ = cg.search_rrf("alpha new", paths=("lexical",), judge=False, record=False)
169
- self.assertEqual(project.call_count, 1)
170
- self.assertIn("a", [hit[0]["id"] for hit in hits])
171
- other = cg.index["nodes"]["b"]["path"]
172
- self.assertIs(cg._like_lower_cache[other][1], before[other][1])
173
- changed = cg.index["nodes"]["a"]["path"]
174
- self.assertIn("new", cg._like_lower_cache[changed][1])
175
- self.assertNotIn("old", cg._like_lower_cache[changed][1])
176
-
177
- def test_rebuild_after_external_write_invalidates_both_projections(self):
178
- cg = self.make_store({"same": "A B C Alpha"})
179
- cg.search_rrf("alpha", paths=("lexical",), judge=False, record=False)
180
- entry = cg.index["nodes"]["same"]
181
- path = os.path.join(cg.root, entry["path"])
182
- with open(path, encoding="utf-8") as stream:
183
- original = stream.read()
184
- with open(path, "w", encoding="utf-8") as stream:
185
- stream.write(original.replace("A B C Alpha", "D E F Beta"))
186
- cg.rebuild_index()
187
- hits, _ = cg.search_rrf("beta", paths=("lexical",), judge=False, record=False)
188
- self.assertIn("same", [hit[0]["id"] for hit in hits])
189
- self.assertIn("DEF", cg._doc_exact_text(entry, cg._read(entry)[1]))
190
- self.assertNotIn("ABC", cg._doc_exact_text(entry, cg._read(entry)[1]))
191
-
192
-
193
- if __name__ == "__main__":
194
- unittest.main()
1
+ """Runtime guards for exact-match projection reuse and its invalidation."""
2
+ from __future__ import annotations
3
+
4
+ import os
5
+ import re
6
+ import tempfile
7
+ import unittest
8
+ from unittest.mock import call, patch
9
+
10
+ from . import mdcg, nodefile, readcache
11
+ from .mdcos import MdCGOS
12
+
13
+
14
+ def frozen_exact(q, content, **kw):
15
+ """Pre-optimization predicate, independent of all new projection helpers."""
16
+ query = re.sub(r"\s+", "", q or "")
17
+ return int(len(query) >= 2 and query in re.sub(r"\s+", "", content or ""))
18
+
19
+
20
+ class ExactTextCache(unittest.TestCase):
21
+ def setUp(self):
22
+ clean = {k: v for k, v in os.environ.items() if not k.startswith("MDCG_")}
23
+ clean.update(MDCG_HOTCACHE="0", MDCG_READ_CACHE="1", MDCG_FRESHNESS="0",
24
+ MDCG_UNIFY_QUERY="0", MDCG_EN_ATOMS="0")
25
+ self.env = patch.dict(os.environ, clean, clear=True)
26
+ self.env.start()
27
+ self.addCleanup(self.env.stop)
28
+
29
+ def make_store(self, bodies=None):
30
+ temp = tempfile.TemporaryDirectory(prefix="exact_projection_")
31
+ self.addCleanup(temp.cleanup)
32
+ cg = MdCGOS(temp.name)
33
+ self.addCleanup(cg.close)
34
+ for nid, body in (bodies or {"a": "AB 与 BC 分开", "z": "A B\u2003C", "case": "abc"}).items():
35
+ cg.add(nid, body, layer="knowledge", importance=0.5)
36
+ cg.flush()
37
+ return cg
38
+
39
+ def docs(self, cg):
40
+ return [(entry, *cg._read(entry)) for entry in list(cg.index["nodes"].values())]
41
+
42
+ def score(self, cg, query):
43
+ return cg._score(self.docs(cg), query, mdcg.bigrams(query))
44
+
45
+ def test_raw_and_prepared_predicates_preserve_whitespace_and_case(self):
46
+ cases = [("ABC", "A B\u2003C", 1), ("A\tB C", "ABC", 1),
47
+ ("ABC", "AB 与 BC 分开", 0), ("ABC", "abc", 0),
48
+ ("甲\n乙", "甲\u00a0乙", 1), ("甲乙", "甲\u3000乙", 1),
49
+ ("", "ABC", 0), (None, None, 0), ("A", "A", 0)]
50
+ for query, body, expected in cases:
51
+ with self.subTest(query=query, body=body):
52
+ prepared = (mdcg.compact_exact_text(query), mdcg.compact_exact_text(body))
53
+ self.assertEqual(mdcg.exact_hit_of(query, body), expected)
54
+ self.assertEqual(mdcg.exact_hit_of(query, body, _prepared=prepared), expected)
55
+
56
+ def test_production_results_and_scores_equal_frozen_predicate(self):
57
+ cg = self.make_store()
58
+ for mode in ("legacy", "jaccard"):
59
+ with patch.object(mdcg, "SCORE_MODE", mode):
60
+ for query in ("ABC", "A B C", "甲乙", "unmatched", "", "abc"):
61
+ with self.subTest(mode=mode, query=query):
62
+ cached_scores = self.score(cg, query)
63
+ new_search, _ = cg.search(query, k=3, judge=False, record=False)
64
+ new_rrf, _ = cg.search_rrf(query, k=3, paths=("lexical",), judge=False, record=False)
65
+ with patch.object(mdcg, "exact_hit_of", frozen_exact):
66
+ self.assertEqual(cached_scores, self.score(cg, query))
67
+ self.assertEqual(new_search, cg.search(query, k=3, judge=False, record=False)[0])
68
+ self.assertEqual(new_rrf, cg.search_rrf(query, k=3, paths=("lexical",), judge=False, record=False)[0])
69
+
70
+ def test_warm_scoring_normalizes_only_query_once(self):
71
+ cg = self.make_store()
72
+ self.score(cg, "ABC")
73
+ with patch.object(mdcg, "compact_exact_text", wraps=mdcg.compact_exact_text) as normalize:
74
+ self.score(cg, "A B C")
75
+ self.assertEqual(normalize.call_args_list, [call("A B C")])
76
+
77
+ def test_one_write_invalidates_only_changed_document(self):
78
+ cg = self.make_store()
79
+ self.score(cg, "ABC")
80
+ before = {path: record[1] for path, record in cg._exact_text_cache.items()}
81
+ cg.add("a", "A\tB\nC 新内容", layer="knowledge", importance=0.5)
82
+ cg.flush()
83
+ with patch.object(mdcg, "compact_exact_text", wraps=mdcg.compact_exact_text) as normalize:
84
+ rows = self.score(cg, "ABC")
85
+ self.assertEqual(normalize.call_count, 2) # one query, one changed document
86
+ a = next(node for node, _ in rows if node["id"] == "a")
87
+ self.assertEqual(a[mdcg.EXACT_HIT_KEY], 1)
88
+ for path, record in cg._exact_text_cache.items():
89
+ if path != cg.index["nodes"]["a"]["path"]:
90
+ self.assertIs(record[1], before[path])
91
+
92
+ def test_clear_removes_projection_without_changing_legacy_count(self):
93
+ cg = self.make_store()
94
+ cg.search_rrf("ABC", paths=("lexical",), judge=False, record=False)
95
+ expected = sum(len(getattr(cg, name, {})) for name in
96
+ ("_read_cache", "_norm_bigrams_cache", "_positive_body_cache"))
97
+ self.assertTrue(cg._exact_text_cache)
98
+ self.assertTrue(cg._like_lower_cache)
99
+ self.assertEqual(readcache.clear(cg), expected)
100
+ self.assertFalse(cg._exact_text_cache)
101
+ self.assertFalse(cg._like_lower_cache)
102
+ self.assertEqual(readcache.clear(cg), 0)
103
+
104
+ def test_short_query_does_not_normalize_documents(self):
105
+ cg = self.make_store()
106
+ with patch.object(cg, "_doc_exact_text", side_effect=AssertionError("unnecessary projection")):
107
+ rows = self.score(cg, "A")
108
+ self.assertTrue(all(node[mdcg.EXACT_HIT_KEY] == 0 for node, _ in rows))
109
+
110
+ def test_opt_out_uses_uncached_projection(self):
111
+ with patch.dict(os.environ, {"MDCG_READ_CACHE": "0"}):
112
+ cg = self.make_store()
113
+ self.assertFalse(hasattr(cg, "_exact_text_cache"))
114
+ self.assertFalse(hasattr(cg, "_like_lower_cache"))
115
+ rows = self.score(cg, "ABC")
116
+ self.assertEqual({node["id"]: node[mdcg.EXACT_HIT_KEY] for node, _ in rows},
117
+ {"a": 0, "z": 1, "case": 0})
118
+
119
+ def test_instances_with_same_relative_path_are_isolated(self):
120
+ first = self.make_store({"same": "A B C"})
121
+ second = self.make_store({"same": "D E F"})
122
+ self.assertEqual(self.score(first, "ABC")[0][0][mdcg.EXACT_HIT_KEY], 1)
123
+ self.assertEqual(self.score(second, "ABC")[0][0][mdcg.EXACT_HIT_KEY], 0)
124
+ self.assertIsNot(first._exact_text_cache, second._exact_text_cache)
125
+
126
+ def test_reinstall_does_not_keep_hidden_projection_cache(self):
127
+ cg = self.make_store()
128
+ entry = next(iter(cg.index["nodes"].values()))
129
+ cg._doc_exact_text(entry, "A B C")
130
+ cg._like_body_lower(entry, "A B C")
131
+ cg._doc_norm_bigrams(entry, "A B C")
132
+ readcache.install(cg)
133
+ readcache.clear(cg)
134
+ self.assertEqual(cg._doc_exact_text(entry, "D E F"), "DEF")
135
+ self.assertEqual(cg._like_body_lower(entry, "DIFFERENT"), "different")
136
+ readcache.clear(cg)
137
+ self.assertEqual(cg._positive_body(entry, "NEW BODY"), "NEW BODY")
138
+ self.assertEqual(cg._doc_norm_bigrams(entry, "NEW BODY"),
139
+ mdcg.bigrams(mdcg.normalize_en("NEW BODY")))
140
+
141
+ def test_lowercase_projection_preserves_old_like_predicate(self):
142
+ cg = self.make_store({"same": "# 正文:İΣ Alpha 中文\n# 不适用条件:Forbidden"})
143
+ entry, fm, body = self.docs(cg)[0]
144
+ fm["tags"] = ["MixedTag"]
145
+ lower = nodefile.positive_body(body).lower()
146
+ for terms in ([], ["alpha"], ["Alpha"], ["i\u0307"], ["中文"],
147
+ ["forbidden"], ["mixedtag"], ["absent"]):
148
+ with self.subTest(terms=terms):
149
+ expected = any(t in lower or t.lower() in "mixedtag" for t in terms)
150
+ self.assertEqual(cg._like(body, fm, terms), expected)
151
+ self.assertEqual(cg._like(body, fm, terms, entry=entry,
152
+ body_lower=cg._like_body_lower), expected)
153
+ self.assertTrue(cg._like(body, fm, [], index_keys={"hit": True},
154
+ body_lower=lambda *_: self.fail("index hit should return early")))
155
+
156
+ def test_lowercase_warm_reuse_and_single_write_invalidation(self):
157
+ cg = self.make_store({"a": "Alpha OLD", "b": "Alpha OTHER"})
158
+ cg.search_rrf("alpha", paths=("lexical",), judge=False, record=False)
159
+ # BM25 并入默认检索后,首次 search 会为 bm25 索引投影一次正文(之后走缓存)。
160
+ # 故先**预热一次** search,再断言「热态下不再重复投影」——意图不变(热复用零投影)。
161
+ cg.search("alpha", judge=False, record=False)
162
+ before = dict(cg._like_lower_cache)
163
+ with patch.object(cg, "_positive_body", wraps=cg._positive_body) as project:
164
+ cg.search("alpha", judge=False, record=False)
165
+ self.assertEqual(project.call_count, 0)
166
+ cg.add("a", "Alpha NEW", layer="knowledge", importance=0.5)
167
+ cg.flush()
168
+ hits, _ = cg.search_rrf("alpha new", paths=("lexical",), judge=False, record=False)
169
+ self.assertEqual(project.call_count, 1)
170
+ self.assertIn("a", [hit[0]["id"] for hit in hits])
171
+ other = cg.index["nodes"]["b"]["path"]
172
+ self.assertIs(cg._like_lower_cache[other][1], before[other][1])
173
+ changed = cg.index["nodes"]["a"]["path"]
174
+ self.assertIn("new", cg._like_lower_cache[changed][1])
175
+ self.assertNotIn("old", cg._like_lower_cache[changed][1])
176
+
177
+ def test_rebuild_after_external_write_invalidates_both_projections(self):
178
+ cg = self.make_store({"same": "A B C Alpha"})
179
+ cg.search_rrf("alpha", paths=("lexical",), judge=False, record=False)
180
+ entry = cg.index["nodes"]["same"]
181
+ path = os.path.join(cg.root, entry["path"])
182
+ with open(path, encoding="utf-8") as stream:
183
+ original = stream.read()
184
+ with open(path, "w", encoding="utf-8") as stream:
185
+ stream.write(original.replace("A B C Alpha", "D E F Beta"))
186
+ cg.rebuild_index()
187
+ hits, _ = cg.search_rrf("beta", paths=("lexical",), judge=False, record=False)
188
+ self.assertIn("same", [hit[0]["id"] for hit in hits])
189
+ self.assertIn("DEF", cg._doc_exact_text(entry, cg._read(entry)[1]))
190
+ self.assertNotIn("ABC", cg._doc_exact_text(entry, cg._read(entry)[1]))
191
+
192
+
193
+ if __name__ == "__main__":
194
+ unittest.main()