fix: use --runner pooling instead of --task embed for vllm serve

README.md

@@ -86,13 +86,7 @@ Start the embedding server once, then route from any process without reloading t
 ```bash
 # Terminal 1: Start vLLM embedding server (runs once, stays alive)
 uv pip install vllm
-
-# vLLM >= 0.8
-vllm serve Qwen/Qwen3-0.6B --task embed --port 8000
-
-# vLLM < 0.8 (use this if the above fails)
-python -m vllm.entrypoints.openai.api_server \
-    --model Qwen/Qwen3-0.6B --task embed --port 8000
+vllm serve Qwen/Qwen3-0.6B --runner pooling --port 8000
 ```
 
 ```python