fix: litellm embedding dimension adapts

openviking/models/embedder/litellm_embedders.py

@@ -85,6 +85,19 @@ def __init__(
             )
         self._dimension = dimension
 
+    def _truncate_vector(self, vector: List[float]) -> List[float]:
+        """Truncate vector to target dimension if needed.
+
+        Args:
+            vector: Input vector from API
+
+        Returns:
+            Truncated vector if dimension is set and smaller than input, otherwise original vector
+        """
+        if self.dimension is not None and len(vector) > self.dimension:
+            return vector[: self.dimension]
+        return vector
+
     def _build_kwargs(self, is_query: bool = False) -> Dict[str, Any]:
         """Build kwargs dict for litellm.embedding() call."""
         kwargs: Dict[str, Any] = {"model": self.model_name}
@@ -95,8 +108,9 @@ def _build_kwargs(self, is_query: bool = False) -> Dict[str, Any]:
             kwargs["api_base"] = self.api_base
         if self.extra_headers:
             kwargs["extra_headers"] = self.extra_headers
-        if self.dimension:
-            kwargs["dimensions"] = self.dimension
+        # Don't pass dimensions parameter to API - some models don't support it
+        # (e.g., Qwen3-Embedding-4B doesn't support matryoshka representation)
+        # Instead, we'll truncate the result vector if needed
 
         # Non-symmetric embedding support
         active_param = None
@@ -171,6 +185,8 @@ def _call() -> EmbedResult:
             response = litellm.embedding(**kwargs)
             self._update_telemetry_token_usage(response)
             vector = response.data[0]["embedding"]
+            # Truncate vector if needed
+            vector = self._truncate_vector(vector)
             return EmbedResult(dense_vector=vector)
 
         try:
@@ -221,7 +237,11 @@ def _call() -> List[EmbedResult]:
             kwargs["input"] = texts
             response = litellm.embedding(**kwargs)
             self._update_telemetry_token_usage(response)
-            return [EmbedResult(dense_vector=item["embedding"]) for item in response.data]
+            # Truncate vectors if needed
+            return [
+                EmbedResult(dense_vector=self._truncate_vector(item["embedding"]))
+                for item in response.data
+            ]
 
         try:
             return self._run_with_retry(