diff --git a/digital-avatar-app/backend/embeddings.py b/digital-avatar-app/backend/embeddings.py
index 0e20019..259e061 100644
--- a/digital-avatar-app/backend/embeddings.py
+++ b/digital-avatar-app/backend/embeddings.py
@@ -18,6 +18,14 @@ EMBED_DIM = 256
 MODEL = os.getenv("EMBEDDING_MODEL", "mock-hash-embed-v1")
 
 
+def _embedding_endpoint(api_url):
+    """Accept either an OpenAI-compatible base URL or its full endpoint."""
+    api_url = (api_url or "").strip().rstrip("/")
+    if not api_url or api_url.endswith("/embeddings"):
+        return api_url
+    return f"{api_url}/embeddings"
+
+
 def _tokenize(text):
     text = (text or "").lower()
     # 英文/数字按词，CJK 逐字（中文无空格，需拆到字级才能命中子词）
@@ -47,7 +55,7 @@ def embed(texts):
     """返回 list[list[float]]，与输入顺序一致。"""
     if not texts:
         return []
-    api_url = os.getenv("EMBEDDING_API_URL")
+    api_url = _embedding_endpoint(os.getenv("EMBEDDING_API_URL"))
     if api_url:
         api_key = os.getenv("EMBEDDING_API_KEY", "")
         model = os.getenv("EMBEDDING_MODEL", "text-embedding-3-small")
diff --git a/digital-avatar-app/backend/tests/test_embeddings.py b/digital-avatar-app/backend/tests/test_embeddings.py
index eafb1d8..55098f5 100644
--- a/digital-avatar-app/backend/tests/test_embeddings.py
+++ b/digital-avatar-app/backend/tests/test_embeddings.py
@@ -48,9 +48,11 @@ class RemoteEmbeddingTests(unittest.TestCase):
     def test_large_input_is_split_into_provider_safe_batches(self):
         texts = [f"chunk-{index}" for index in range(14)]
         batch_sizes = []
+        requested_urls = []
 
         def fake_urlopen(request, timeout):
             self.assertEqual(timeout, 30)
+            requested_urls.append(request.full_url)
             payload = json.loads(request.data.decode("utf-8"))
             batch_sizes.append(len(payload["input"]))
             return FakeResponse({
@@ -61,7 +63,7 @@ class RemoteEmbeddingTests(unittest.TestCase):
             })
 
         with patch.dict(os.environ, {
-            "EMBEDDING_API_URL": "https://embedding.example/v1/embeddings",
+            "EMBEDDING_API_URL": "https://embedding.example/v1",
             "EMBEDDING_API_KEY": "test-key",
             "EMBEDDING_MODEL": "text-embedding-v4",
             "EMBEDDING_BATCH_SIZE": "10",
@@ -69,8 +71,18 @@ class RemoteEmbeddingTests(unittest.TestCase):
             result = embeddings.embed(texts)
 
         self.assertEqual(batch_sizes, [10, 4])
+        self.assertEqual(requested_urls, [
+            "https://embedding.example/v1/embeddings",
+            "https://embedding.example/v1/embeddings",
+        ])
         self.assertEqual(result, [[float(index)] for index in range(14)])
 
+    def test_full_embedding_endpoint_is_not_modified(self):
+        self.assertEqual(
+            embeddings._embedding_endpoint("https://embedding.example/v1/embeddings/"),
+            "https://embedding.example/v1/embeddings",
+        )
+
 
 if __name__ == "__main__":
     unittest.main()
diff --git a/digital-avatar-app/docs/H5_PRODUCTION_DEPLOYMENT.md b/digital-avatar-app/docs/H5_PRODUCTION_DEPLOYMENT.md
index 3139594..cce121a 100644
--- a/digital-avatar-app/docs/H5_PRODUCTION_DEPLOYMENT.md
+++ b/digital-avatar-app/docs/H5_PRODUCTION_DEPLOYMENT.md
@@ -47,10 +47,16 @@ HUIHUI_PAYMENT_TIMEOUT_SECONDS=30
 DATABASE_URL=sqlite:////data/avatar.db
 UPLOAD_DIR=/data/uploads
 CHAT_MODEL_CONFIG_URL=http://<huihuisquare-api>/api/ai-models/runtime/digital-avatar
+EMBEDDING_API_URL=https://dashscope.aliyuncs.com/compatible-mode/v1
+EMBEDDING_API_KEY=<production-embedding-api-key>
+EMBEDDING_MODEL=text-embedding-v3
+EMBEDDING_BATCH_SIZE=10
 ```
 
 如生产 AI 配置中心不可用，还应提供当前项目支持的 `OPENAI_API_KEY`、`OPENAI_BASE_URL`、`CHAT_MODEL` 等兜底配置。`/data` 必须挂载持久卷，数据库与知识库文件不可存放在容器临时层。
 
+`EMBEDDING_API_URL` 同时支持 OpenAI 兼容基础地址（如上面的 `/v1`）和完整的 `/v1/embeddings` 地址，后端会统一请求 `/embeddings`。发布后必须在后端容器内执行一次最小向量探针，确认返回向量数量和维度，而不能只检查 `/api/health`。
+
 积分充值使用会会支付体系的 `payment-v3/payment/pay`，渠道值为 `WECHAT` / `ALIPAY`，端内支付场景为 `APP`，微信内 H5 使用 `JSAPI`。`HUIHUI_PAYMENT_CALLBACK_SECRET` 只用于为每笔订单生成 HMAC 回调签名，不会发送到前端或直接出现在回调地址中。支付回调确认状态成功且金额与套餐价格完全一致后才增加积分，重复回调不会重复到账。
 
 ## 3. 构建与发布
@@ -74,6 +80,7 @@ docker compose build --pull avatar-backend avatar-frontend
 docker compose up -d avatar-backend avatar-frontend
 docker compose ps
 curl -fsS http://127.0.0.1:8099/api/health
+docker compose exec avatar-backend python -c 'import embeddings; v=embeddings.embed(["部署向量探针"]); print(len(v), len(v[0]))'
 ```
 
 生产编排应把示例中的测试端口改为内网暴露，由统一 HTTPS 网关接入。后端暂时使用 SQLite，必须保持单实例写入；若扩展为多后端实例，应先迁移到 PostgreSQL，并把延迟接管任务改为共享队列。