fix(embedding): 自动检测改为探测 OpenAI embedding 最大可用维度

Author: exynos967

astrbot/core/provider/provider.py

@@ -305,6 +305,10 @@ def get_dim(self) -> int:
         """获取向量的维度"""
         ...
 
+    async def detect_dim(self) -> int:
+        """探测模型原生向量维度（默认实现）"""
+        return len(await self.get_embedding("astrbot"))
+
     async def test(self) -> None:
         await self.get_embedding("astrbot")
 

astrbot/core/provider/sources/gemini_embedding_source.py

@@ -78,6 +78,19 @@ async def get_embeddings(self, text: list[str]) -> list[list[float]]:
         except APIError as e:
             raise Exception(f"Gemini Embedding API批量请求失败: {e.message}")
 
+    async def detect_dim(self) -> int:
+        """探测模型原生向量维度（不传 output_dimensionality）"""
+        try:
+            result = await self.client.models.embed_content(
+                model=self.model,
+                contents="echo",
+            )
+            assert result.embeddings is not None
+            assert result.embeddings[0].values is not None
+            return len(result.embeddings[0].values)
+        except APIError as e:
+            raise Exception(f"Gemini Embedding 维度探测失败: {e.message}")
+
     def get_dim(self) -> int:
         """获取向量的维度"""
         return int(self.provider_config.get("embedding_dimensions", 768))

astrbot/core/provider/sources/openai_embedding_source.py

@@ -52,6 +52,64 @@ async def get_embeddings(self, text: list[str]) -> list[list[float]]:
         )
         return [item.embedding for item in embeddings.data]
 
+    async def detect_dim(self) -> int:
+        """探测模型可用的最大向量维度"""
+
+        async def _request_dim(dimensions: int | None) -> int:
+            kwargs = {
+                "input": "echo",
+                "model": self.model,
+            }
+            if dimensions is not None:
+                kwargs["dimensions"] = dimensions
+            embedding = await self.client.embeddings.create(**kwargs)
+            return len(embedding.data[0].embedding)
+
+        # 1) 默认调用，获取当前默认维度
+        base_dim = await _request_dim(None)
+
+        # 2) 先判断 dimensions 参数是否可调
+        probe_dim = base_dim + 1
+        try:
+            probe_result = await _request_dim(probe_dim)
+            if probe_result != probe_dim:
+                return base_dim
+        except Exception:
+            return base_dim
+
+        # 3) 可调时探测上界：指数扩张 + 二分
+        max_cap = 32768
+        low = probe_dim
+        high = max(base_dim * 2, probe_dim + 1)
+        if high > max_cap:
+            high = max_cap
+
+        while high < max_cap:
+            try:
+                result_dim = await _request_dim(high)
+                if result_dim != high:
+                    break
+                low = high
+                high = min(high * 2, max_cap)
+            except Exception:
+                break
+
+        left = low + 1
+        right = high - 1
+        while left <= right:
+            mid = (left + right) // 2
+            try:
+                result_dim = await _request_dim(mid)
+                if result_dim == mid:
+                    low = mid
+                    left = mid + 1
+                else:
+                    right = mid - 1
+            except Exception:
+                right = mid - 1
+
+        return low
+
     def get_dim(self) -> int:
         """获取向量的维度"""
         return int(self.provider_config.get("embedding_dimensions", 1024))

astrbot/dashboard/routes/config.py

@@ -754,6 +754,16 @@ async def get_embedding_dim(self):
             if not provider_type:
                 return Response().error("provider_config 缺少 type 字段").__dict__
 
+            # 首次添加某类提供商时，provider_cls_map 可能尚未注册该适配器
+            if provider_type not in provider_cls_map:
+                try:
+                    self.core_lifecycle.provider_manager.dynamic_import_provider(
+                        provider_type,
+                    )
+                except ImportError as e:
+                    logger.error(traceback.format_exc())
+                    return Response().error(f"动态导入提供商适配器失败: {e!s}").__dict__
+
             # 获取对应的 provider 类
             if provider_type not in provider_cls_map:
                 return (
@@ -779,9 +789,8 @@ async def get_embedding_dim(self):
             if inspect.iscoroutinefunction(init_fn):
                 await init_fn()
 
-            # 获取嵌入向量维度
-            vec = await inst.get_embedding("echo")
-            dim = len(vec)
+            # 探测嵌入向量维度（优先使用 provider 的原生探测逻辑）
+            dim = await inst.detect_dim()
 
             logger.info(
                 f"检测到 {provider_config.get('id', 'unknown')} 的嵌入向量维度为 {dim}",