hermes 外部记忆openviking服务端embedding模型使用llama.cpp
环境
Win11，WSL2，Ubuntu24.04
openviking-server --version
openviking-server 0.4.10
ov --help
OpenViking v0.4.11.dev0
一、相关文档
在线服务
二、备份数据
停止服务
（Ctrl+C 或）pkill -f openviking-server
备份数据目录
cp -r ~/.openviking/data ~/.openviking/data_backup_volc_embedding
三、切换embedding模型
新版本Seed-1.6-embedding图文多模态向量化模型全新上线，主要面向图文多模向量检索的使用场景，支持图片输入及中、英双语文本输入，最长 8K 上下文长度。
输入：文本、图片、视频
（openviking不支持其他厂商嵌入模型的多模态能力）
bge-small-zh-v1.5-f16（huggingface）
FlagEmbedding 能够将任意文本映射为低维稠密向量，可用于检索、分类、聚类、语义搜索等任务，同时也可配合大语言模型应用于向量数据库场景中。
openviking-server init操作
uadmin@PC26:~$ openviking-server init
╭─────────────────────────────────────────────────────────╮
│ OpenViking Setup │
│ Context database for AI agents — data in, context out │
╰─────────────────────────────────────────────────────────╯
Data will be stored under /home/uadmin/.openviking/data unless you edit ov.conf later.
Existing config found: /home/uadmin/.openviking/ov.conf
Current configuration
Embedding: ollama · qwen3-embedding-cpu:0.6b (1024d)
VLM: litellm · ollama/qwen3.6:35b
Query planner: litellm · ollama/guoxuter/ov_intent_analysis_sft:v7_q8
Server: 127.0.0.1:1933 · auth dev (no auth)
What would you like to do? ↑/↓ move · enter select
❯ Start over (full setup, current config backed up as .bak)
Update VLM (now: litellm · ollama/qwen3.6:35b)
Update embedding (now: ollama · qwen3-embedding-cpu:0.6b (1024d))
Update server & auth (now: 127.0.0.1:1933 · auth dev (no auth))
Cancel
选择Update embedding
---------------------------------------------------------------------------------------
❯ Update embedding
Current embedding: ollama · qwen3-embedding-cpu:0.6b (1024d)
Embedding setup: ↑/↓ move · enter select
Cloud API (VolcEngine, BytePlus, OpenAI)
❯ Local via Ollama (no API key, runs on this machine)
Lightweight CPU embedding (llama.cpp, ~24 MB, no Ollama needed)
选择Lightweight CPU embedding
---------------------------------------------------------------------------------------
❯ Lightweight CPU embedding
Checking llama-cpp-python... not installed
llama-cpp-python is required for local CPU embedding.
Install now? (pip install "openviking[local-embed]") [Y/n]:
输入Y
---------------------------------------------------------------------------------------
Install now? (pip install "openviking[local-embed]") [Y/n]: Y
/home/uadmin/.local/share/uv/tools/openviking/bin/python: No module named pip
Native build failed, retrying with generic CPU flags...
/home/uadmin/.local/share/uv/tools/openviking/bin/python: No module named pip
Installation failed.
Try manually: pip install "openviking[local-embed]"
Continue anyway? (config will be generated) [y/N]:
输入N
---------------------------------------------------------------------------------------
Continue anyway? (config will be generated) [y/N]: N
Setup cancelled.安装系统编译依赖（llama-cpp-python 编译需要）
sudo apt update && sudo apt install -y build-essential cmake
用 uv 原生方式安装本地嵌入依赖
直接用 uv 给 OpenViking 工具环境添加 local-embed 额外依赖，uv 会自动处理安装，不需要 pip：
uv tool install --reinstall "openviking[local-embed]"
说明：--reinstall 会保留你原有配置，仅补充安装额外依赖包（包括 llama-cpp-python）。
另一种安装方法
手动给 openviking 的 uv 环境安装依赖：
uv pip install --python ~/.local/share/uv/tools/openviking/bin/python "openviking[local-embed]"
重新执行openviking-server init
❯ Lightweight CPU embedding
Checking llama-cpp-python... installed
Embedding model: ↑/↓ move · enter select
❯ BGE-small-zh v1.5 (f16) (512d, ~24 MB)
选择BGE-small-zh v1.5 (f16)
---------------------------------------------------------------------------------------
❯ BGE-small-zh v1.5 (f16)
Model 'bge-small-zh-v1.5-f16' not downloaded yet. Download now? (~24 MB) [Y/n]:
输入Y
---------------------------------------------------------------------------------------
Model 'bge-small-zh-v1.5-f16' not downloaded yet. Download now? (~24 MB) [Y/n]: Y
Downloading...
---------------------------------------------------------------------------------------
Model 'bge-small-zh-v1.5-f16' not downloaded yet. Download now? (~24 MB) [Y/n]: Y
Download failed: HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /CompendiumLabs/bge-small-zh-v1.5-gguf/resolve/main/bge-small-zh-v1.5-f16.gguf?download=true (Caused by NewConnectionError("HTTPSConnection(host='huggingface.co', port=443): Failed to establish a new connection: [Errno 101] Network is unreachable"))
Model will be auto-downloaded on first server start.
Embedding dimension changes from 1024 to 512.
Existing vector indexes become unusable — data must be re-ingested.
Continue? [y/N]:
网络报错，先输入N
---------------------------------------------------------------------------------------
Continue? [y/N]: N
Setup cancelled.重新执行openviking-server init
打开网络代理
Model 'bge-small-zh-v1.5-f16' not downloaded yet. Download now? (~24 MB) [Y/n]: Y
Downloading...
---------------------------------------------------------------------------------------
Model 'bge-small-zh-v1.5-f16' not downloaded yet. Download now? (~24 MB) [Y/n]: Y
OK Model downloaded to /home/uadmin/.cache/openviking/models/bge-small-zh-v1.5-f16.gguf
Embedding dimension changes from 1024 to 512.
Existing vector indexes become unusable — data must be re-ingested.
Continue? [y/N]:
输入y
---------------------------------------------------------------------------------------
Continue? [y/N]: y
Change: ollama · qwen3-embedding-cpu:0.6b (1024d)
→ local · bge-small-zh-v1.5-f16 (512d)
Save configuration? [Y/n]:
输入Y
---------------------------------------------------------------------------------------
Save configuration? [Y/n]: Y
Existing config backed up to /home/uadmin/.openviking/ov.conf.bak.5
OK Configuration updated
Validate the setup now? (runs `openviking-server doctor`) [Y/n]:
输入Y
---------------------------------------------------------------------------------------
Validate the setup now? (runs `openviking-server doctor`) [Y/n]: Y
OpenViking Doctor
Config: PASS /home/uadmin/.openviking/ov.conf
Python: PASS 3.11.15 (>= 3.10 required)
Native Engine: PASS variant=x86_avx2
AGFS: PASS AGFS SDK 0.1.7
Embedding: PASS local/bge-small-zh-v1.5-f16 (/home/uadmin/.cache/openviking/models/bge-small-zh-v1.5-f16.gguf)
VLM: PASS litellm/ollama/qwen3.6:35b
Ollama: PASS running at localhost:11434
VikingBot: PASS VikingBot aligned with dev OpenViking auth
Disk: PASS 947.5 GB free in /home/uadmin/.openviking/data
All checks passed.
Start the server now? [y/N]:
输入N，手动启动
---------------------------------------------------------------------------------------
Start the server now? [y/N]: N配置文件ov.conf
~/.openviking/ov.conf
"embedding": {
"dense": {
"provider": "local",
"model": "bge-small-zh-v1.5-f16",
"dimension": 512
}
},更换嵌入模型后，向量数据需要处理。
如果是少量测试数据，可以清理了。如果想保留数据可以重建向量数据。
四、清理向量数据
停止服务
（Ctrl+C 或）pkill -f openviking-server
备份数据目录
cp -r ~/.openviking/data ~/.openviking/data_backup_volc_embedding
删除目录
rm -rf ~/.openviking/data/*
#mv ~/.openviking/data ~/.openviking/data1
重启服务会生成data目录
openviking-server
五、处理向量数据
启动服务
openviking-server
日志报错
openviking.storage.errors.EmbeddingRebuildRequiredError: Existing collection embedding dimension (1024) does not match current configuration (512). Vectors are incompatible; rebuild is required.
删除向量数据库目录
rm -rf ~/.openviking/data/vectordb
再次启动服务
openviking-server
重建向量数据
看根目录有什么
ov ls viking:///
uadmin@UD26:~$ ov ls viking:///
cmd: ov ls viking:/// -l 256 -n 256
1. dir · 2026-07-27 06:28
viking://resources
本目录是 Hermes Agent
框架的核心文档集合，专注于指导开发者构建具备长期记忆能力和复杂推理能力的智能 Agent...
2. dir · 2026-07-27 06:52
viking://user
This directory serves as the central hub for the `user` component within a sophisticated AI
assistant framework. The primary focus is on managing persistent user identity, ensuring secu...# 1. 重建全局资源库的向量（手动导入的文档、素材、公共资源）
ov reindex viking://resources/ --mode vectors_only --wait true
# 2. 重建默认用户的所有对话记忆与个人数据的向量
ov reindex viking://user/default/ --mode vectors_only --wait true
uadmin@UD26:~$ ov reindex viking://resources/ --mode vectors_only --wait true
status completed
uri viking://resources
object_type resource
mode vectors_only
scanned_records 12
rebuilt_records 17
deleted_records 0
would_delete_records 0
unsupported_records 0
failed_records 0
duration_ms 6528
warnings []
uadmin@UD26:~$ ov reindex viking://user/default/ --mode vectors_only --wait true
status completed
uri viking://user/default
object_type user_namespace
mode vectors_only
scanned_records 34
rebuilt_records 36
deleted_records 0
would_delete_records 0
unsupported_records 0
failed_records 0
duration_ms 11923
warnings []测试
ov find "测试标记"
uadmin@UD26:~$ ov find "测试标记"
cmd: ov find --uri= -n 10 "测试标记"
9 results
Ranked by relevance · limit 10 final results
1. memory · Level 2 · score 0.565
viking://user/default/memories/preferences/2026/07/24/凌晨开发_openviking-20260724.md
测试标记为 openviking-20260724，开发者凌晨仍在开发中。
2. memory · Level 2 · score 0.467
viking://user/default/memories/experiences/viking_search验证开发环境记忆.md
## Situation - 用户要求使用viking_search工具搜索开发环境相关内容，验证新存储记忆的语义检索有效性
## Approach - 调用viking_search工具，传入开发环境相关的搜索关键词 -...
3. memory · Level 1 · score 0.459
viking://user/default/memories/cases/.overview.md
No abstract available.
...ov find --help
重新生成语义摘要和重建向量索引
一般重建向量数据即可，不用重建语义摘要。
# 1. 全局资源库：重新生成语义摘要 + 重建向量索引
ov reindex viking://resources/ --mode semantic_and_vectors --wait true
# 2. 默认用户记忆库：重新生成语义摘要 + 重建向量索引
ov reindex viking://user/default/ --mode semantic_and_vectors --wait true
uadmin@PC26:~$ ov reindex viking://resources/ --mode semantic_and_vectors --wait true
status completed
uri viking://resources
object_type resource
mode semantic_and_vectors
scanned_records 12
rebuilt_records 17
deleted_records 0
would_delete_records 0
unsupported_records 0
failed_records 0
duration_ms 48866
warnings []
# 这句命令更容易报错
uadmin@PC26:~$ ov reindex viking://user/default/ --mode semantic_and_vectors
ov reindex viking://user/default/ --mode semantic_and_vectors
╭─ Request Timeout ────────────────────────────────────────────────────╮
│ OpenViking did not respond before the configured timeout expired. │
╰──────────────────────────────────────────────────────────────────────╯
Next:
ov config Increase the request timeout
ov config show Show the active configopenviking-server日志报错
[07/27/26 06:51:58] ERROR Task was destroyed but it is pending! base_events.py:1785
task: <Task pending name='Task-1659' coro=<LoggingWorker._worker_loop() done, defined at
/home/uadmin/.local/share/uv/tools/openviking/lib/python3.11/site-packages/litellm/litellm_core_utils/logging_worker.py:109> wait_for=<Future cancelled>>问题分析，可能是Ollama压力稍大，处理不过来
适当调整配置
Ollma环境变量
# 允许同时处理 2 个推理请求，VLM 与嵌入任务并行执行，避免完全串行排队
OLLAMA_NUM_PARALLEL=2
# 允许 2 个模型同时常驻显存，避免 VLM 和嵌入模型来回加载卸载
OLLAMA_MAX_LOADED_MODELS=2
# 开启 FlashAttention 硬件加速，降低显存占用、提升推理速度
OLLAMA_FLASH_ATTENTION=1
ollama serve（重启Ollama）
启动日志能看见这个几个变量
ov.conf配置
vlm模型从"num_ctx": 16384，改到8192（重启服务）（后来又改回去了）
再次执行
如果openviking-server日志里有报错，ov命令执行成功了，可以接受。
测试ov find
ov find "测试标记" -u viking://user/default
更多推荐
4
0- 0
已为社区贡献3条内容
所有评论(0)