Document recommended local embedding routing

README.md

@@ -157,6 +157,22 @@ pip install -e .
 openspace-mcp --help   # verify installation
 ```
 
+> [!TIP]
+> **Recommended split routing for OpenAI-compatible gateways**
+>
+> If your main model runs through an OpenAI-compatible provider or local relay (for example `gpt-5.4` via `http://127.0.0.1:8080/v1`), the recommended default is:
+> - keep the main LLM on that provider via `OPENSPACE_LLM_*`
+> - keep skill-router embeddings local via `OPENSPACE_SKILL_EMBEDDING_BACKEND=local`
+> - use `OPENSPACE_SKILL_EMBEDDING_MODEL=BAAI/bge-small-en-v1.5`
+>
+> Why this is the default recommendation:
+> - lower latency for routing and prefilter
+> - no dependence on a remote `/v1/embeddings` endpoint
+> - no extra token spend for embedding generation
+> - stronger main LLM still handles final reasoning and selection
+>
+> Architecture notes and flow diagram: [`docs/current-routing-flow.md`](docs/current-routing-flow.md)
+
 > [!TIP]
 > **Slow clone?** The `assets/` folder (~50 MB of images) makes the default clone large. Use this lightweight alternative to skip it:
 > ```bash

README_CN.md

@@ -157,6 +157,22 @@ pip install -e .
 openspace-mcp --help   # 验证安装
 ```
 
+> [!TIP]
+> **OpenAI 兼容网关下，默认推荐双路由方案 A**
+>
+> 如果你的主模型走的是 OpenAI 兼容 provider 或本地 relay（例如 `gpt-5.4` 走 `http://127.0.0.1:8080/v1`），当前最推荐的默认配置是：
+> - 主 LLM 继续走 `OPENSPACE_LLM_*`
+> - skill router 的 embedding 走本地：`OPENSPACE_SKILL_EMBEDDING_BACKEND=local`
+> - 本地 embedding 模型使用：`OPENSPACE_SKILL_EMBEDDING_MODEL=BAAI/bge-small-en-v1.5`
+>
+> 这样选的原因：
+> - 路由和预筛选延迟更低
+> - 不依赖远程 `/v1/embeddings`
+> - embedding 不额外消耗 provider token
+> - 更强的主 LLM 仍然负责最终推理和选择
+>
+> 架构说明和流程图见：[`docs/current-routing-flow.md`](docs/current-routing-flow.md)
+
 > [!TIP]
 > **Clone 太慢？** `assets/` 目录包含约 50 MB 的图片文件，导致仓库较大。使用以下轻量方式跳过它：
 > ```bash