OpenAI Python SDK embeddings.create() injects encoding_format="base64" by default, breaking providers that reject unknown parameters

The OpenAI Python SDK's embeddings.create() method automatically sets encoding_format="base64" when the parameter is not explicitly provided, then decodes the base64 response client-side for performance. This is an internal SDK optimization, not documented as a required API parameter. Providers that implement the OpenAI-compatible /v1/embeddings endpoint but do not support the encoding_format parameter (e.g., SambaNova) will reject the request with a 400 error: "Unknown parameter: 'encoding_format'". Verified with openai SDK 2.36.0, May 2026. The behavior is in the AsyncEmbeddings.create() source: if not is_given(encoding_format): params["encoding_format"] = "base64".

When wrapping an OpenAI-compatible provider that does not support encoding_format, override the embedding method to bypass client.embeddings.create(). Instead, call client.post("/embeddings", body=..., cast_to=CreateEmbeddingResponse) directly with only the parameters the provider supports (typically input, model, and optionally dimensions). This avoids the SDK's automatic encoding_format injection.