fix(embedding): decode base64 embedding responses before vector processing (#6807)

* fix(embedding): decode base64 embedding responses before vector processing

When a model's extra body config includes `encoding_format: "base64"`, the
embedding API returns a base64-encoded IEEE 754 little-endian float32 array
instead of a `number[]`. The previous code passed this raw string directly
to `formatVectors`, which called `.reduce()` on it and threw:
  TypeError: a.reduce is not a function

Add `decodeEmbedding()` that detects base64 strings and decodes them to
`number[]` via `Buffer → Float32Array → Array.from()`, then use it in
`getVectorsByText` before calling `formatVectors`.

Fixes #6769

* perf: test

---------

Co-authored-by: octo-patch <octo-patch@github.com>
Co-authored-by: archer <545436317@qq.com>

packages/service/core/ai/embedding/index.ts

@@ -91,7 +91,9 @@ export async function getVectorsByText({ model, input, type, headers }: GetVecto
                 return tokens.reduce((sum, item) => sum + item, 0);
               })(),
               Promise.all(
-                res.data.map((item) => formatVectors(item.embedding, model.normalization))
+                res.data.map((item) =>
+                  formatVectors(decodeEmbedding(item.embedding), model.normalization)
+                )
               )
             ]);
 
@@ -121,6 +123,16 @@ export async function getVectorsByText({ model, input, type, headers }: GetVecto
   }
 }
 
+export function decodeEmbedding(embedding: number[] | string): number[] {
+  if (typeof embedding === 'string') {
+    // base64-encoded IEEE 754 little-endian float32 array
+    const buf = Buffer.from(embedding, 'base64');
+    const floats = new Float32Array(buf.buffer, buf.byteOffset, buf.byteLength / 4);
+    return Array.from(floats);
+  }
+  return embedding;
+}
+
 export function formatVectors(vector: number[], normalization = false) {
   // normalization processing
   function normalizationVector(vector: number[]) {