diff --git a/ingest_progress.txt b/ingest_progress.txt
deleted file mode 100644
index 7f31f0a..0000000
--- a/ingest_progress.txt
+++ /dev/null
@@ -1,56 +0,0 @@
-初始化...
-待处理: 46 个文件
-跳过超大: 6 个
-  SKIP 人工智能协会内容创作通知.md (1915KB)
-  SKIP Git团队协作指南.md (94KB)
-  SKIP Git团队协作指南.sync-conflict-20260615-122209-R4RFYM7.md (94KB)
-  SKIP OpenClaw介绍.md (64KB)
-  SKIP Git内部原理详解.md (59KB)
-  SKIP Git团队协作指南大纲.md (58KB)
-[1/46]   17 chunks | 博客/AI时代的Token战争-能源算力与我们每个人的选择.md (0.6s)
-[2/46]   39 chunks | 博客/AI时代的VibeCoding与程序员进化.md (1.2s)
-[3/46]   59 chunks | 博客/CLI在AI时代的浴火重生.md (1.7s)
-[4/46]   68 chunks | 博客/DeepSeek-V4全面解析.md (1.1s)
-[5/46]   16 chunks | 博客/大模型赋能架构设计.md (0.5s)
-[6/46]   27 chunks | 博客/大语言模型算子详解.md (1.0s)
-[7/46]    8 chunks | 博客/学习Agent，越学越像在重新理解操作系统.md (0.3s)
-[8/46]   69 chunks | 博客/深入解析 Claude Code：Vibe Coding 时代的 AI 编程利器.md (1.5s)
-[9/46]   12 chunks | 博客/自建Git服务-在NAS上部署Gitea.md (0.4s)
-[10/46]   38 chunks | 博客/DeepSeek-V4博客大纲.md (0.4s)
-[11/46]  110 chunks | 博客/uv工具推荐博客大纲.md (1.0s)
-[12/46]   48 chunks | 博客/AI助你轻松上手LaTeX论文写作.md (1.1s)
-[13/46]   15 chunks | 博客/SEO分析报告.md (0.4s)
-[14/46]   18 chunks | 博客/博客爬取报告.md (0.3s)
-[15/46]   15 chunks | 博客/大数据技术栈.md (0.3s)
-[16/46]    7 chunks | 博客/摘要检查报告.md (0.3s)
-[17/46]   27 chunks | 博客/文章列表.md (0.5s)
-[18/46]   19 chunks | 博客/标签分类审查报告.md (0.7s)
-[19/46]   13 chunks | 博客/LinearRegression线性回归.md (0.6s)
-[20/46]   53 chunks | 博客/从全连接层到卷积.md (1.2s)
-[21/46]  114 chunks | 博客/深度学习完全指南.md (1.6s)
-[22/46]   43 chunks | 博客/视觉语言模型技术综述.md (1.4s)
-[23/46]   52 chunks | 博客/Docker安装与入门指南.md (1.4s)
-[24/46]  167 chunks | 博客/Docker部署完全指南.md (3.6s)
-[25/46]  115 chunks | 博客/Git团队协作指南（精简版）.md (1.3s)
-[26/46]   10 chunks | 博客/OpenClaw安装教程.md (0.1s)
-[27/46]  247 chunks | 博客/uv工具推荐博客.md (2.8s)
-[28/46]    6 chunks | Club/CLAUDE.md (0.1s)
-[29/46]    7 chunks | Club/README.md (0.1s)
-[30/46]   11 chunks | Club/仓库说明.md (0.4s)
-[31/46]    9 chunks | Club/导航页.md (0.3s)
-[32/46]    3 chunks | Club/说明.md (0.1s)
-[33/46]    4 chunks | Club/说明.md (0.2s)
-[34/46]    4 chunks | Club/MetaRL核心团队实施方案.md (0.3s)
-[35/46]    3 chunks | Club/说明.md (0.1s)
-[36/46]   20 chunks | Club/定量考核材料清单.md (0.7s)
-[37/46]   14 chunks | Club/年度工作总结起草指南.md (0.6s)
-[38/46]    3 chunks | Club/说明.md (0.2s)
-[39/46]    3 chunks | Club/说明.md (0.1s)
-[40/46]    7 chunks | Club/人工智能协会近期工作任务清单.md (0.3s)
-[41/46]    3 chunks | Club/说明.md (0.2s)
-[42/46]    7 chunks | halo/README.md (0.3s)
-[43/46]    6 chunks | halo/README.zh-CN.md (0.3s)
-[44/46]   21 chunks | halo/usage-guide.md (0.4s)
-[45/46]    9 chunks | 顶层/CLAUDE.md (0.4s)
-[46/46]   57 chunks | 顶层/temp-agent-os-v2.md (1.6s)
-[DONE] 完成: 46 文件, 1623 chunks (45s)
diff --git a/ingest_stdout.txt b/ingest_stdout.txt
deleted file mode 100644
index 4b23c91..0000000
--- a/ingest_stdout.txt
+++ /dev/null
@@ -1,57 +0,0 @@
-��ʼ��...
-
Loading weights:   0%|          | 0/71 [00:00<?, ?it/s]
Loading weights: 100%|��������������������| 71/71 [00:00<00:00, 12790.81it/s]
-������: 46 ���ļ�
-��������: 6 ��
-  SKIP �˹�����Э�����ݴ���֪ͨ.md (1915KB)
-  SKIP Git�Ŷ�Э��ָ��.md (94KB)
-  SKIP Git�Ŷ�Э��ָ��.sync-conflict-20260615-122209-R4RFYM7.md (94KB)
-  SKIP OpenClaw����.md (64KB)
-  SKIP Git�ڲ�ԭ�����.md (59KB)
-  SKIP Git�Ŷ�Э��ָ�ϴ��.md (58KB)
-[1/46]   17 chunks | ����/AIʱ����Tokenս��-��Դ����������ÿ���˵�ѡ��.md (0.6s)
-[2/46]   39 chunks | ����/AIʱ����VibeCoding�����Ա����.md (1.2s)
-[3/46]   59 chunks | ����/CLI��AIʱ����ԡ������.md (1.7s)
-[4/46]   68 chunks | ����/DeepSeek-V4ȫ�����.md (1.1s)
-[5/46]   16 chunks | ����/��ģ�͸��ܼܹ����.md (0.5s)
-[6/46]   27 chunks | ����/������ģ���������.md (1.0s)
-[7/46]    8 chunks | ����/ѧϰAgent��ԽѧԽ���������������ϵͳ.md (0.3s)
-[8/46]   69 chunks | ����/������� Claude Code��Vibe Coding ʱ���� AI �������.md (1.5s)
-[9/46]   12 chunks | ����/�Խ�Git����-��NAS�ϲ���Gitea.md (0.4s)
-[10/46]   38 chunks | ����/DeepSeek-V4���ʹ��.md (0.4s)
-[11/46]  110 chunks | ����/uv�����Ƽ����ʹ��.md (1.0s)
-[12/46]   48 chunks | ����/AI������������LaTeX�����.md (1.1s)
-[13/46]   15 chunks | ����/SEO��������.md (0.4s)
-[14/46]   18 chunks | ����/������ȡ����.md (0.3s)
-[15/46]   15 chunks | ����/�����ݼ���ջ.md (0.3s)
-[16/46]    7 chunks | ����/ժҪ��鱨��.md (0.3s)
-[17/46]   27 chunks | ����/�����б�.md (0.5s)
-[18/46]   19 chunks | ����/��ǩ������鱨��.md (0.7s)
-[19/46]   13 chunks | ����/LinearRegression���Իع�.md (0.6s)
-[20/46]   53 chunks | ����/��ȫ���Ӳ㵽����.md (1.2s)
-[21/46]  114 chunks | ����/���ѧϰ��ȫָ��.md (1.6s)
-[22/46]   43 chunks | ����/�Ӿ�����ģ�ͼ�������.md (1.4s)
-[23/46]   52 chunks | ����/Docker��װ������ָ��.md (1.4s)
-[24/46]  167 chunks | ����/Docker������ȫָ��.md (3.6s)
-[25/46]  115 chunks | ����/Git�Ŷ�Э��ָ�ϣ�����棩.md (1.3s)
-[26/46]   10 chunks | ����/OpenClaw��װ�̳�.md (0.1s)
-[27/46]  247 chunks | ����/uv�����Ƽ�����.md (2.8s)
-[28/46]    6 chunks | Club/CLAUDE.md (0.1s)
-[29/46]    7 chunks | Club/README.md (0.1s)
-[30/46]   11 chunks | Club/�ֿ�˵��.md (0.4s)
-[31/46]    9 chunks | Club/����ҳ.md (0.3s)
-[32/46]    3 chunks | Club/˵��.md (0.1s)
-[33/46]    4 chunks | Club/˵��.md (0.2s)
-[34/46]    4 chunks | Club/MetaRL�����Ŷ�ʵʩ����.md (0.3s)
-[35/46]    3 chunks | Club/˵��.md (0.1s)
-[36/46]   20 chunks | Club/�������˲����嵥.md (0.7s)
-[37/46]   14 chunks | Club/��ȹ����ܽ����ָ��.md (0.6s)
-[38/46]    3 chunks | Club/˵��.md (0.2s)
-[39/46]    3 chunks | Club/˵��.md (0.1s)
-[40/46]    7 chunks | Club/�˹�����Э����ڹ��������嵥.md (0.3s)
-[41/46]    3 chunks | Club/˵��.md (0.2s)
-[42/46]    7 chunks | halo/README.md (0.3s)
-[43/46]    6 chunks | halo/README.zh-CN.md (0.3s)
-[44/46]   21 chunks | halo/usage-guide.md (0.4s)
-[45/46]    9 chunks | ����/CLAUDE.md (0.4s)
-[46/46]   57 chunks | ����/temp-agent-os-v2.md (1.6s)
-[DONE] ���: 46 �ļ�, 1623 chunks (45s)
diff --git a/scripts/ingest_obsidian.py b/scripts/ingest_obsidian.py
index 720fd06..f1c7130 100644
--- a/scripts/ingest_obsidian.py
+++ b/scripts/ingest_obsidian.py
@@ -8,7 +8,7 @@ from src.core.db import VectorDB
 from src.core.embedder import create_embedder
 from src.core.ingest import DocumentIngestor
 
-MAX_SIZE = 50_000  # 跳过 >50KB 的文件
+MAX_SIZE = 200_000  # GPU 嵌入,无需跳过
 
 progress_file = Path(__file__).parent.parent / "ingest_progress.txt"
 def log(msg):
diff --git a/src/core/embedder.py b/src/core/embedder.py
index 1132a62..d999c75 100644
--- a/src/core/embedder.py
+++ b/src/core/embedder.py
@@ -52,21 +52,33 @@ class Embedder(Protocol):
 
 # -- 本地模型 --
 class LocalEmbedder:
-    """sentence-transformers 本地模型嵌入器."""
+    """sentence-transformers 本地模型嵌入器. 自动检测 GPU."""
 
     def __init__(self, config: EmbedConfig):
         from sentence_transformers import SentenceTransformer
+
+        # 自动检测 GPU
+        try:
+            import torch
+            device = "cuda" if torch.cuda.is_available() else "cpu"
+            if device == "cuda":
+                logger.info("检测到 GPU: %s", torch.cuda.get_device_name(0))
+        except ImportError:
+            device = "cpu"
+
         self._config = config
         try:
             self._model = SentenceTransformer(
-                config.local_model, local_files_only=True
+                config.local_model, local_files_only=True, device=device
             )
         except Exception:
             logger.info("模型未缓存, 通过镜像下载 %s", config.local_model)
             old_endpoint = os.environ.get("HF_ENDPOINT")
             os.environ["HF_ENDPOINT"] = _HF_MIRROR
             try:
-                self._model = SentenceTransformer(config.local_model)
+                self._model = SentenceTransformer(
+                    config.local_model, device=device
+                )
             finally:
                 if old_endpoint is not None:
                     os.environ["HF_ENDPOINT"] = old_endpoint
diff --git a/src/core/ingest.py b/src/core/ingest.py
index a9e7d58..d439915 100644
--- a/src/core/ingest.py
+++ b/src/core/ingest.py
@@ -155,7 +155,9 @@ class MarkdownSplitter:
             part = text[start:break_point].strip()
             if part:
                 parts.append(part)
-            start = break_point - self.overlap if self.overlap > 0 else break_point
+            # 确保 start 始终前进（避免分隔符距 start 小于 overlap 时 start 回退导致死循环）
+            next_start = break_point - self.overlap if self.overlap > 0 else break_point
+            start = max(start + 1, next_start)
         return parts