[Bug] Embedding 队列错误导致 auto-capture 返回 0 memories #811
Description
问题描述
在使用 OpenViking 作为 OpenClaw 的上下文引擎时，遇到以下问题：
-
auto-capture 捕获消息但提取 0 记忆
- 日志显示：
- 捕获成功，但提取阶段失败
-
Embedding 队列错误
- [queue] (healthy)
+----------------+---------+-------------+-----------+--------+-------+
| Queue | Pending | In Progress | Processed | Errors | Total |
+----------------+---------+-------------+-----------+--------+-------+
| Embedding | 0 | 0 | 0 | 0 | 0 |
| Semantic | 0 | 0 | 0 | 0 | 0 |
| Semantic-Nodes | 0 | 0 | 0 | 0 | 0 |
| TOTAL | 0 | 0 | 0 | 0 | 0 |
+----------------+---------+-------------+-----------+--------+-------+
- [queue] (healthy)
[vikingdb] (healthy)
+------------+-------------+--------------+--------+
| Collection | Index Count | Vector Count | Status |
+------------+-------------+--------------+--------+
| context | 1 | 117 | OK |
| TOTAL | 1 | 117 | |
+------------+-------------+--------------+--------+
[vlm] (healthy)
No token usage data available.
[lock] (healthy)
No active locks.
[retrieval] (healthy)
No retrieval queries recorded.
[system] (healthy) 显示：，Embedding 队列有 12 个错误
- 导致新对话无法向量化，无法存储为记忆
- 检索质量下降
- 平均相似度分从 0.30 降至 0.10
- 零结果率从 50% 升至 57%
环境配置
- OpenViking 版本: 2026.2.6-3 (通过 OpenClaw 插件安装)
- Embedding 模型: BGE-large-zh-v1.5 (通过 Ollama)
- VLM 模型: qwen2.5:3b
- 配置:
排查步骤
- ✅ Ollama 服务正常，BGE 模型能生成 embedding
- ✅ 配置已修复 (之前是 )
- ✅ VLM 能正常响应提取请求
- ❌ Embedding 队列持续错误，导致新记忆无法创建
预期行为
- auto-capture 应该能提取并存储新记忆
- Embedding 队列应该正常工作
- 检索相似度分应该合理（>0.5）
实际行为
- 只有旧记忆能被检索
- 新对话无法添加到记忆库
- 系统处于"能读不能写"状态
临时解决方案
暂时切换到 lossless-claw 上下文压缩引擎，等待此问题修复。
日志片段
[queue] (healthy)
+----------------+---------+-------------+-----------+--------+-------+
| Queue | Pending | In Progress | Processed | Errors | Total |
+----------------+---------+-------------+-----------+--------+-------+
| Embedding | 0 | 0 | 0 | 0 | 0 |
| Semantic | 0 | 0 | 0 | 0 | 0 |
| Semantic-Nodes | 0 | 0 | 0 | 0 | 0 |
| TOTAL | 0 | 0 | 0 | 0 | 0 |
+----------------+---------+-------------+-----------+--------+-------+
[vikingdb] (healthy)
+------------+-------------+--------------+--------+
| Collection | Index Count | Vector Count | Status |
+------------+-------------+--------------+--------+
| context | 1 | 117 | OK |
| TOTAL | 1 | 117 | |
+------------+-------------+--------------+--------+
[vlm] (healthy)
No token usage data available.
[lock] (healthy)
No active locks.
[retrieval] (healthy)
No retrieval queries recorded.
[system] (healthy) 输出：
请帮忙排查 Embedding 队列错误的原因，谢谢！