Clarify split Ragas model endpoints

@@ -5,11 +5,6 @@ WEKNORA_KB_NAME=ragas-eval-pilot
 
 # Ragas generation and judge models. These are evaluation-side models, not the
 # model configuration used by the WeKnora backend.
-OPENAI_API_KEY=replace-me
-OPENAI_BASE_URL=https://api.openai.com/v1
-
-# Optional split deployment. Use these when LLM and embedding are served by
-# different OpenAI-compatible services, such as vLLM + Infinity.
 RAGAS_LLM_API_KEY=replace-me
 RAGAS_LLM_BASE_URL=http://localhost:8000/v1
 RAGAS_EMBEDDING_API_KEY=replace-me

@@ -34,8 +34,9 @@ cp .env.example .env
 - `WEKNORA_API_KEY` 是 WeKnora API Key
 - `WEKNORA_KB_ID` 是目标知识库 ID；如果还没有，先运行 `python scripts/00_create_kb.py`
 - `WEKNORA_KB_NAME` 是创建知识库时使用的名称
-- `OPENAI_API_KEY`、`OPENAI_BASE_URL`、`RAGAS_*_MODEL` 是评估侧模型配置
-- 如果 LLM 和 embedding 分开部署，使用 `RAGAS_LLM_BASE_URL` 指向 vLLM 的 `/v1`，使用 `RAGAS_EMBEDDING_BASE_URL` 指向 Infinity 的 `/v1`
+- `RAGAS_LLM_BASE_URL` 指向 vLLM 的 OpenAI-compatible `/v1`
+- `RAGAS_EMBEDDING_BASE_URL` 指向 Infinity embedding 的 OpenAI-compatible `/v1`
+- `RAGAS_*_MODEL` 是评估侧模型名称
 
 ## 首轮 Pilot
 

@@ -41,10 +41,6 @@ qa:
 
 ragas:
   provider: "openai-compatible"
-  # Backward-compatible defaults. If the split LLM/embedding values below are
-  # empty, these values are used for both clients.
-  api_key: "${OPENAI_API_KEY}"
-  base_url: "${OPENAI_BASE_URL}"
   # vLLM OpenAI-compatible endpoint, for example http://localhost:8000/v1.
   llm_api_key: "${RAGAS_LLM_API_KEY}"
   llm_base_url: "${RAGAS_LLM_BASE_URL}"

@@ -22,10 +22,10 @@ def run_ragas_eval(
     from ragas.run_config import RunConfig
 
     ragas_config = config["ragas"]
-    llm_api_key = _first_non_empty(ragas_config, "llm_api_key", "api_key")
-    llm_base_url = _first_non_empty(ragas_config, "llm_base_url", "base_url")
-    embedding_api_key = _first_non_empty(ragas_config, "embedding_api_key", "api_key")
-    embedding_base_url = _first_non_empty(ragas_config, "embedding_base_url", "base_url")
+    llm_api_key = _required_ragas_value(ragas_config, "llm_api_key")
+    llm_base_url = _required_ragas_value(ragas_config, "llm_base_url")
+    embedding_api_key = _required_ragas_value(ragas_config, "embedding_api_key")
+    embedding_base_url = _required_ragas_value(ragas_config, "embedding_base_url")
     judge_model = str(require_config(config, "ragas.judge_model"))
     embedding_model = str(require_config(config, "ragas.embedding_model"))
     temperature = float(ragas_config.get("temperature", 0))
@@ -127,12 +127,11 @@ def _metric_map() -> dict[str, Any]:
         }
 
 
-def _first_non_empty(config: dict[str, Any], *keys: str) -> str:
-    for key in keys:
-        value = config.get(key)
-        if value not in {None, ""}:
-            return str(value)
-    raise ValueError(f"Missing required Ragas config value. Checked: {', '.join(keys)}")
+def _required_ragas_value(config: dict[str, Any], key: str) -> str:
+    value = config.get(key)
+    if value in {None, ""}:
+        raise ValueError(f"Missing required Ragas config value: ragas.{key}")
+    return str(value)
 
 
 def _wrap_langchain_models(llm: Any, embeddings: Any) -> tuple[Any, Any]: