Fix Bedrock Titan V2 encoding_format parameter support

- Add encoding_format to supported OpenAI parameters list
- Implement encoding_format to embeddingTypes parameter mapping
- Map 'float' to ['float'] and 'base64' to ['binary'] formats
- Handle response with proper fallback: binary > float > embedding field
- Support both float and binary response formats per AWS documentation

Fixes #14685 - UnsupportedParamsError when using encoding_format with Titan V2

Author: timelfrink

litellm/llms/bedrock/embed/amazon_titan_v2_transformation.py

@@ -10,7 +10,7 @@
 """
 
 import types
-from typing import List, Optional
+from typing import List, Optional, Union
 
 from litellm.types.llms.bedrock import (
     AmazonTitanV2EmbeddingRequest,
@@ -30,9 +30,7 @@ class AmazonTitanV2Config:
     normalize: Optional[bool] = None
     dimensions: Optional[int] = None
 
-    def __init__(
-        self, normalize: Optional[bool] = None, dimensions: Optional[int] = None
-    ) -> None:
+    def __init__(self, normalize: Optional[bool] = None, dimensions: Optional[int] = None) -> None:
         locals_ = locals().copy()
         for key, value in locals_.items():
             if key != "self" and value is not None:
@@ -57,32 +55,56 @@ def get_config(cls):
         }
 
     def get_supported_openai_params(self) -> List[str]:
-        return ["dimensions"]
+        return ["dimensions", "encoding_format"]
 
-    def map_openai_params(
-        self, non_default_params: dict, optional_params: dict
-    ) -> dict:
+    def map_openai_params(self, non_default_params: dict, optional_params: dict) -> dict:
         for k, v in non_default_params.items():
             if k == "dimensions":
                 optional_params["dimensions"] = v
+            elif k == "encoding_format":
+                # Map OpenAI encoding_format to AWS embeddingTypes
+                if v == "float":
+                    optional_params["embeddingTypes"] = ["float"]
+                elif v == "base64":
+                    # base64 maps to binary format in AWS
+                    optional_params["embeddingTypes"] = ["binary"]
+                else:
+                    # For any other encoding format, default to float
+                    optional_params["embeddingTypes"] = ["float"]
         return optional_params
 
-    def _transform_request(
-        self, input: str, inference_params: dict
-    ) -> AmazonTitanV2EmbeddingRequest:
+    def _transform_request(self, input: str, inference_params: dict) -> AmazonTitanV2EmbeddingRequest:
         return AmazonTitanV2EmbeddingRequest(inputText=input, **inference_params)  # type: ignore
 
-    def _transform_response(
-        self, response_list: List[dict], model: str
-    ) -> EmbeddingResponse:
+    def _transform_response(self, response_list: List[dict], model: str) -> EmbeddingResponse:
         total_prompt_tokens = 0
 
         transformed_responses: List[Embedding] = []
         for index, response in enumerate(response_list):
             _parsed_response = AmazonTitanV2EmbeddingResponse(**response)  # type: ignore
+
+            # According to AWS docs, embeddingsByType is always present
+            # If binary was requested (encoding_format="base64"), use binary data
+            # Otherwise, use float data from embeddingsByType or fallback to embedding field
+            embedding_data: Union[List[float], List[int]]
+
+            if ("embeddingsByType" in _parsed_response and
+                "binary" in _parsed_response["embeddingsByType"]):
+                # Use binary data if available (for encoding_format="base64")
+                embedding_data = _parsed_response["embeddingsByType"]["binary"]
+            elif ("embeddingsByType" in _parsed_response and
+                  "float" in _parsed_response["embeddingsByType"]):
+                # Use float data from embeddingsByType
+                embedding_data = _parsed_response["embeddingsByType"]["float"]
+            elif "embedding" in _parsed_response:
+                # Fallback to legacy embedding field
+                embedding_data = _parsed_response["embedding"]
+            else:
+                raise ValueError(f"No embedding data found in response: {response}")
+
             transformed_responses.append(
                 Embedding(
-                    embedding=_parsed_response["embedding"],
+                    embedding=embedding_data,
                     index=index,
                     object="embedding",
                 )