fix(data-process): 合并文档结构短切片

backend/app/api/v1/endpoints/data_process.py

@@ -256,7 +256,7 @@ def _merge_short_chunks(
     min_token_count: int,
     chunk_size: int,
 ) -> list[TextChunk]:
-    """在同一章节内合并短块，同时严格遵守切片大小上限。"""
+    """按原文顺序合并相邻短块，同时严格遵守切片大小上限。"""
 
     merged: list[TextChunk] = []
     index = 0
@@ -308,7 +308,7 @@ def _chunk_source_text(
     config: dict[str, Any],
     preprocess_options: set[str],
 ) -> list[tuple[TextChunk, tuple[str, ...]]]:
-    """按可选文档结构分段后切片，章节之间绝不共享 overlap。"""
+    """按可选文档结构分段后切片，结构边界之间不共享 overlap。"""
 
     method = str(_value(config, "chunk_method", "chunkMethod", "structure"))
     detect_structure = (
@@ -368,14 +368,19 @@ def _chunk_source_text(
         section_text = text[start:end]
         local_chunks = chunk_unstructured(section_text, **common)
         shifted = [_shift_chunk(chunk, start, text) for chunk in local_chunks]
-        if merge_short:
-            shifted = _merge_short_chunks(
-                shifted,
-                text,
-                min_token_count=configured_minimum,
-                chunk_size=chunk_size,
-            )
         result.extend((chunk, heading_path) for chunk in shifted)
+
+    if merge_short and result:
+        # 结构分段只负责提供标题路径和隔离 overlap，不应让目录项或短小节
+        # 突破 min_chunk_size 约束。合并后保留首个原始块的标题路径。
+        heading_paths = {chunk.start: heading_path for chunk, heading_path in result}
+        merged = _merge_short_chunks(
+            [chunk for chunk, _ in result],
+            text,
+            min_token_count=configured_minimum,
+            chunk_size=chunk_size,
+        )
+        result = [(chunk, heading_paths.get(chunk.start, ())) for chunk in merged]
     return result
 
 

backend/tests/test_data_process_api.py

@@ -1402,7 +1402,6 @@ def test_every_unstructured_preprocess_option_changes_preview_behavior() -> None
     )
     assert without_merge[0]["token_count"] < 5
     assert with_merge[0]["token_count"] >= 5
-
     mojibake = "这是无法可靠读取的内容，锟斤拷锟斤拷锟斤拷，需要预先过滤。"
     assert len(_preview_task(mojibake, options=[])) == 1
     assert _preview_task(mojibake, options=["filter_low_quality"]) == []
@@ -1451,6 +1450,25 @@ def test_every_unstructured_preprocess_option_changes_preview_behavior() -> None
     assert "[EMAIL]" in masked["edited_content"]
 
 
+def test_merge_short_content_applies_across_adjacent_structure_sections() -> None:
+    content = "\n".join(f"{index}. 小节{index}\n内容{index}。" for index in range(1, 9))
+    items = _preview_task(
+        content,
+        options=["merge_short_content"],
+        config={
+            "chunk_method": "structure",
+            "chunk_size": 40,
+            "chunk_overlap": 0,
+            "min_chunk_size": 20,
+        },
+    )
+
+    assert len(items) == 2
+    assert all(20 <= item["token_count"] <= 40 for item in items)
+    assert items[0]["source_start_line"] == 1
+    assert items[0]["source_end_line"] == 8
+
+
 def test_stored_binary_document_text_is_not_reparsed_as_binary() -> None:
     for file_format in ("pdf", "docx", "pptx"):
         items = _preview_task(