Description

With encoding_format set to base64, OpenAIEmbeddingClient.get_embeddings decodes each payload with struct.unpack sized as len(raw) // 4. When the decoded byte length is not a multiple of 4 (a truncated or malformed response), the buffer is longer than the format expects and the call fails with a bare struct.error: unpack requires a buffer of X bytes, with no wrapping and no indication of which input's embedding broke.

Observed Behavior

A response item whose base64 payload decodes to, say, 5 bytes raises struct.error straight out of get_embeddings.

Expected Behavior

The client should reject the payload with a descriptive error naming the input index and the decoded length, so callers can tell a malformed response apart from a client bug.

Reproduction Steps

1. Request embeddings with options={"encoding_format": "base64"}.
2. Have the endpoint (or a proxy) return a payload whose decoded length is not divisible by 4.
3. get_embeddings raises struct.error.

Minimal Reproduction

raw = base64.b64encode(struct.pack("<1f", 0.5) + b"\\x00").decode()
# served as the embedding string for one item

Error Messages and Stack Traces

struct.error: unpack requires a buffer of 4 bytes

Package Versions

agent_framework_openai from main.

Python Version

3.11