From 09d95675b5bd9bdcddf2824b8bea9c557a050a7f Mon Sep 17 00:00:00 2001
From: damingishere-coder <damingishere@gmail.com>
Date: Sat, 5 Sep 2026 13:12:02 +0800
Subject: [PATCH] =?UTF-8?q?fix:=20=E5=85=81=E8=AE=B8=E6=98=B5=E7=A7=B0?=
 =?UTF-8?q?=E6=95=B0=E5=AD=97=E7=9A=84=20OCR=20=E7=A9=BA=E6=A0=BC?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 app/image/fact_verification.py        | 38 ++++++++++++++++++++-------
 tests/test_image_fact_verification.py | 26 ++++++++++++++++++
 2 files changed, 55 insertions(+), 9 deletions(-)

diff --git a/app/image/fact_verification.py b/app/image/fact_verification.py
index e9b8422..61105d1 100644
--- a/app/image/fact_verification.py
+++ b/app/image/fact_verification.py
@@ -111,6 +111,31 @@ def _numeric_fact_is_allowed(candidate: str, allowed: set[str]) -> bool:
     return False
 
 
+def _unverified_numeric_facts(
+    ocr_text: str,
+    allowed_numbers: set[str],
+    evidence_lines: list[str],
+) -> tuple[str, ...]:
+    """逐行判断数字，允许 OCR 仅在真实字母数字昵称中插入空格。"""
+
+    allowed_line_compacts = {
+        compact
+        for line in evidence_lines
+        if len(compact := _compact_text(line)) >= 4
+    }
+    unknown: set[str] = set()
+    for line in ocr_text.splitlines():
+        line_compact = _compact_text(line)
+        exact_evidence_line = line_compact in allowed_line_compacts
+        for candidate in _numeric_facts(line):
+            if _numeric_fact_is_allowed(candidate, allowed_numbers):
+                continue
+            if exact_evidence_line:
+                continue
+            unknown.add(candidate)
+    return tuple(sorted(unknown, key=lambda item: (len(item), item)))
+
+
 def strip_unverified_prompt_numeric_units(prompt_file: Path) -> tuple[str, tuple[str, ...]]:
     """去掉摘要擅自附加、但原始证据没有的数字单位。
 
@@ -275,15 +300,10 @@ def review_image_facts(
     allowed_numbers = _numeric_facts(numeric_evidence)
     # 分镜序号属于版式，不是聊天事实。
     allowed_numbers.update(str(number) for number in range(0, 11))
-    unknown_numeric = tuple(
-        sorted(
-            (
-                item
-                for item in _numeric_facts(text)
-                if not _numeric_fact_is_allowed(item, allowed_numbers)
-            ),
-            key=lambda item: (len(item), item),
-        )
+    unknown_numeric = _unverified_numeric_facts(
+        text,
+        allowed_numbers,
+        evidence_lines,
     )
 
     normalized_lines = [line.strip() for line in text.splitlines() if line.strip()]
diff --git a/tests/test_image_fact_verification.py b/tests/test_image_fact_verification.py
index a271d17..39138c7 100644
--- a/tests/test_image_fact_verification.py
+++ b/tests/test_image_fact_verification.py
@@ -84,6 +84,32 @@ def test_allows_ocr_fragments_of_known_numbers_and_ignores_name_garble(tmp_path)
     assert review.unknown_text == ()
 
 
+def test_allows_ocr_space_inside_evidenced_alphanumeric_sender_name(tmp_path):
+    prompt, image = _evidence(tmp_path)
+    messages_path = tmp_path / "messages.json"
+    messages = json.loads(messages_path.read_text(encoding="utf-8"))
+    messages.append({"sender_name": "Gaosong925", "content": "推荐 Grok App"})
+    messages_path.write_text(json.dumps(messages, ensure_ascii=False), encoding="utf-8")
+
+    review = review_image_facts(prompt, image, ocr_text="Gaosong 925")
+
+    assert review.ok
+    assert review.unknown_numeric == ()
+
+
+def test_rejects_changed_number_in_alphanumeric_sender_name(tmp_path):
+    prompt, image = _evidence(tmp_path)
+    messages_path = tmp_path / "messages.json"
+    messages = json.loads(messages_path.read_text(encoding="utf-8"))
+    messages.append({"sender_name": "Gaosong925", "content": "推荐 Grok App"})
+    messages_path.write_text(json.dumps(messages, ensure_ascii=False), encoding="utf-8")
+
+    review = review_image_facts(prompt, image, ocr_text="Gaosong 926")
+
+    assert not review.ok
+    assert review.unknown_numeric == ("926",)
+
+
 def test_numeric_units_do_not_join_across_ocr_lines(tmp_path):
     prompt, image = _evidence(tmp_path)