union_kb_ingest 1.0.0 → 1.0.1

package/config/config.yaml

@@ -1,8 +1,8 @@
 llm:
-  enabled: false
+  enabled: true
   timeout_seconds: 120
   max_tokens: 4096
   temperature: 0.1
-  api_key: "your-api-key"          # API Key
-  model: "your-model"              # 模型名称
-  base_url: "https://your-model-endpoint"  # 完整模型调用地址
+  api_key: "15f066c4509845038027ea5746524af5.w4CLSC6ODiKVC1wK"          # GLM API Key
+  model: "glm-4.7-flash"                 # 模型名称，可选: glm-4, glm-4-plus, glm-4-flash 等
+  base_url: "https://open.bigmodel.cn/api/paas/v4/"  # 模型服务地址

package/normalizer.py

@@ -53,13 +53,22 @@ def _normalize_with_llm(block: ParsedBlock, status: str) -> List[KnowledgeItem]:
             response = client.chat.completions.create(
                 model=config.model,
                 messages=[
-                    {"role": "system", "content": "你是严谨的运维知识库整理助手，只能依据输入原文生成结构化知识条目。"},
+                    {
+                        "role": "system",
+                        "content": (
+                            "你是严谨的运维知识库整理助手，只能依据输入原文生成结构化知识条目。"
+                            "你必须只返回一个 JSON object，根节点必须只有 items 字段，"
+                            "且 items 必须是数组。不要返回 Markdown、解释文字或其他根字段。"
+                        ),
+                    },
                     {"role": "user", "content": prompt},
                 ],
+                stream=False,
                 max_tokens=config.max_tokens,
                 temperature=config.temperature,
+                do_sample=False,
                 response_format={"type": "json_object"},
-                thinking={"type": "disabled"},
+                thinking={"type": "disabled", "clear_thinking": True},
             )
         except Exception as exc:
             elapsed = time.monotonic() - started_at
@@ -94,7 +103,10 @@ def _normalize_with_llm(block: ParsedBlock, status: str) -> List[KnowledgeItem]:
 
         raw_items = parsed.get("items") if isinstance(parsed, dict) else parsed
         if not isinstance(raw_items, list):
-            print("llm parse failed: JSON does not contain an items list")
+            print(
+                "llm parse failed: JSON does not contain an items list "
+                f"top_level={_json_shape(parsed)} preview={_preview(content)}"
+            )
             if attempt >= LLM_MAX_RETRIES:
                 _abort_llm("JSON does not contain an items list after 10 attempts", block)
             time.sleep(min(2 ** (attempt - 1), 30))
@@ -116,6 +128,13 @@ def _normalize_with_llm(block: ParsedBlock, status: str) -> List[KnowledgeItem]:
 
 
 def _get_zhipu_client_class():
+    try:
+        from zai import ZaiClient
+
+        return ZaiClient
+    except (ImportError, AttributeError):
+        pass
+
     try:
         from zai import ZhipuAiClient
 
@@ -177,10 +196,15 @@ def _first_choice(response):
 
 
 def _response_debug(response) -> str:
-    text = repr(response)
-    if len(text) > 300:
-        text = text[:300] + "..."
-    return _preview(text, limit=300)
+    return _preview(repr(response))
+
+
+def _json_shape(value) -> str:
+    if isinstance(value, dict):
+        return f"object keys={list(value.keys())[:10]}"
+    if isinstance(value, list):
+        return f"array len={len(value)}"
+    return type(value).__name__
 
 
 def _abort_llm(message: str, block: ParsedBlock) -> None:
@@ -205,6 +229,7 @@ def _build_prompt(block: ParsedBlock, status: str) -> str:
 6. 涉及表格、阈值、比较符、单位、持续时间、笔数、适用对象时必须保留原始逻辑。
 7. 输出严格 JSON 对象，不要 Markdown 代码围栏，不要解释文字，不要在 JSON 前后添加任何内容。
 8. status 固定为 "{status}"。
+9. JSON 根节点必须严格为一个对象：{{"items": [...]}}。禁止返回单个 item 对象、禁止返回纯数组、禁止返回 result/data/records/knowledge_items 等其他根字段。
 
 doc_type 只能取：
 {", ".join(sorted(DOC_TYPES))}
@@ -281,8 +306,8 @@ def _json_candidates(text: str) -> List[str]:
     return candidates
 
 
-def _preview(text: str, limit: int = 500) -> str:
-    return re.sub(r"\s+", " ", text).strip()[:limit]
+def _preview(text: str) -> str:
+    return re.sub(r"\s+", " ", text).strip()
 
 
 def _normalize_heuristically(block: ParsedBlock, status: str) -> KnowledgeItem:

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "union_kb_ingest",
-  "version": "1.0.0",
+  "version": "1.0.1",
   "description": "Offline knowledge-base ingest helper for PDF, Word, Markdown and TXT documents.",
   "bin": {
     "union_kb_ingest": "bin/union_kb_ingest"

package/requirements.txt

@@ -3,6 +3,9 @@ pyyaml>=6.0.1
 zhipuai>=2.1.0
 sniffio>=1.3.0
 
+# Pin numpy to avoid incompatible x86-64-v2 wheels on older offline machines.
+numpy==1.26.4
+
 # Docling slim plus only file-format backends used by this offline tool.
 # Do not install `docling` or `docling-slim[standard]` here: those pull OCR,
 # layout/table ML models, torch/onnxruntime, and may try to download artifacts.