union_kb_ingest 1.0.5 → 1.0.6

package/normalizer.py

@@ -156,8 +156,12 @@ def _normalize_with_llm(block: ParsedBlock, status: str) -> List[KnowledgeItem]:
                     f"preview={_preview('；'.join(coverage_issues[:3]))}"
                 )
                 if attempt >= LLM_MAX_RETRIES:
-                    _abort_llm("source fact coverage failed after 10 attempts", block)
-                coverage_retry_feedback = _coverage_retry_prompt(coverage_issues)
+                    print(
+                        "WARNING: source fact coverage failed after "
+                        f"{LLM_MAX_RETRIES} attempts; releasing draft for manual review"
+                    )
+                    return _items_with_coverage_warning(items, block, coverage_issues)
+                coverage_retry_feedback = _coverage_retry_prompt(block, coverage_issues, items)
                 time.sleep(min(2 ** (attempt - 1), 30))
                 continue
             return items
@@ -195,14 +199,31 @@ def _compact_retry_prompt(base_prompt: str) -> str:
     )
 
 
-def _coverage_retry_prompt(missing_facts: List[str]) -> str:
+def _coverage_retry_prompt(
+    block: ParsedBlock,
+    missing_facts: List[str],
+    items: List[KnowledgeItem],
+) -> str:
     """构造事实覆盖不足时的重试提示。"""
     lines = "\n".join(f"- {fact}" for fact in missing_facts[:8])
+    source_excerpt = _preview(block.content)[:4000] or "无"
+    context_excerpt = _preview(block.context)[:2000] or "无"
+    current_core = _preview(
+        "\n\n".join(_core_sections_for_coverage(item.body) for item in items)
+    )[:3000] or "无"
     return (
         "重试补充要求：上一次输出遗漏了以下来源正文事实。"
         "请重新生成 JSON，并把这些事实写入 ## 1. 核心内容、## 2. 适用边界 或 ## 3. 使用要求。"
         "短定义句、简称句、阈值句和规则句应优先保留原句或等价完整表述，不要只概括关键词。\n"
-        f"{lines}"
+        f"{lines}\n\n"
+        f"来源文档：{block.source_doc}\n"
+        f"来源章节：{block.source_section or '全文'}\n\n"
+        "当前来源原文片段如下，请结合上下文判断缺失事实应补入哪个正文小节：\n"
+        f"{source_excerpt}\n\n"
+        "辅助上下文如下。辅助上下文只用于理解位置和术语，不要把其中独有事实写入正文：\n"
+        f"{context_excerpt}\n\n"
+        "上一轮生成的核心正文如下，请在此基础上补齐遗漏事实并保持 JSON 根结构不变：\n"
+        f"{current_core}"
     )
 
 
@@ -719,6 +740,42 @@ def _normalize_heuristically(block: ParsedBlock, status: str) -> KnowledgeItem:
     )
 
 
+def _items_with_coverage_warning(
+    items: List[KnowledgeItem],
+    block: ParsedBlock,
+    missing_facts: List[str],
+) -> List[KnowledgeItem]:
+    """把覆盖不足的 LLM 结果降级为需人工复核的草稿。"""
+    for item in items:
+        item.status = "draft"
+        item.review_status = "coverage_warning"
+        if "需人工复核" not in item.tags:
+            item.tags.append("需人工复核")
+        item.body = _body_with_coverage_warning(item.body, block, missing_facts)
+    return items
+
+
+def _body_with_coverage_warning(
+    body: str,
+    block: ParsedBlock,
+    missing_facts: List[str],
+) -> str:
+    """在正文核心章节标注覆盖不足的来源事实。"""
+    fact_lines = "\n".join(f"> - {fact}" for fact in missing_facts[:20])
+    note = (
+        "\n\n"
+        "> WARNING: LLM 多次重试后仍未通过来源事实覆盖校验，本条目已降为草稿，需人工复核。\n"
+        f"> 来源文档：{block.source_doc}\n"
+        f"> 来源章节：{block.source_section or '全文'}\n"
+        "> 待人工补齐或确认的来源事实：\n"
+        f"{fact_lines}"
+    )
+    pattern = r"(^##\s+1\.\s+核心内容\s*$)"
+    if re.search(pattern, body, flags=re.M):
+        return re.sub(pattern, r"\1" + note, body, count=1, flags=re.M)
+    return f"{body.rstrip()}\n\n## 人工复核提示{note}"
+
+
 def _source_fact_coverage_issues(block: ParsedBlock, items: List[KnowledgeItem]) -> List[str]:
     """检查来源事实是否进入核心正文。"""
     facts = _source_fact_sentences(block.content)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "union_kb_ingest",
-  "version": "1.0.5",
+  "version": "1.0.6",
   "description": "Offline knowledge-base ingest helper for PDF, Word, Markdown and TXT documents.",
   "bin": {
     "union_kb_ingest": "bin/union_kb_ingest"

package/schemas.py

@@ -104,6 +104,7 @@ class KnowledgeItem:
             "applicable_roles": self.applicable_roles,
             "tags": self.tags,
             "status": self.status,
+            "review_status": self.review_status,
         }