fix: serialize timeout and max_retries in OpenAIImageGenerator.to_dict

[pcbeingused333]

Related Issues

• No issue. The same defect was fixed for OpenAITextEmbedder and OpenAIDocumentEmbedder in fix: Add missing timeout and max_retries to OpenAITextEmbedder and OpenAIDocumentEmbedder #9421; OpenAIImageGenerator has it too.

Proposed Changes:

OpenAIImageGenerator.__init__ accepts timeout and max_retries, stores them, and _client_kwargs() uses them to build the OpenAI client — but to_dict didn't include them, so they were dropped on serialization and reverted to the OPENAI_TIMEOUT / OPENAI_MAX_RETRIES fallbacks (30s, 5) on load:

original = OpenAIImageGenerator(timeout=120.0, max_retries=10)
restored = OpenAIImageGenerator.from_dict(original.to_dict())

original._client_kwargs()   # {'timeout': 120.0, 'max_retries': 10}
restored._client_kwargs()   # {'timeout': 30.0,  'max_retries': 5}

Silent: a pipeline saved and reloaded runs against different timeout and retry settings than the ones it was configured with, with nothing in the serialized dict to show it.

The two lines add both to default_to_dict, which is what OpenAIChatGenerator, OpenAITextEmbedder and OpenAIDocumentEmbedder already do. response_format stays out on purpose — it's documented as ignored and kept only for backward compatibility.

How did you test it?

Unit tests, in test/components/generators/test_openai_image_generator.py:

• test_to_dict and test_to_dict_with_params updated. Worth noting: test_to_dict_with_params already passed timeout=60, max_retries=10 and then asserted a dict without them, which is what made this look unintentional rather than a deliberate omission.
• test_to_dict_from_dict_roundtrip_preserves_client_settings added: it asserts through _client_kwargs(), so it fails on the observable consequence (the client is built with the wrong settings) rather than only on the shape of the dict.

All three fail without the two-line change, verified by reverting it. Full file: 23 passed, 2 skipped. ruff check and ruff format clean.

Notes for the reviewer

I found this with a small AST script that compares every @component's __init__ parameters against the keys reaching default_to_dict, after seeing the same class of defect in #3808. This is the only instance it found in this repo that survived manual verification (DocumentSplitter's splitting_function is a false positive — it's serialized separately with serialize_callable). It also flagged a few in haystack-core-integrations; I'll raise those there.

This PR was generated with an AI assistant. I have reviewed the changes and run the tests locally.

Checklist

• I have read the contributors guidelines and the code of conduct.
• I have updated the related issue with new insights and changes.
• I have added unit tests and updated the docstrings.
• I've used one of the conventional commit types for my PR title.
• I have documented my code.
• I have added a release note file, following the contributors guidelines.
• I have run pre-commit hooks and fixed any issue.

[vercel]

@pcbeingused333 is attempting to deploy a commit to the deepset Team on Vercel.

A member of the Team first needs to authorize it.

[pcbeingused333]

The red Tests job here is not from this PR: test_convert_chat_completion_chunk_to_streaming_chunk and test_handle_stream_response fail on a clean main too, on all three OSes. openai 3.6.0 added compute_units, cache_write_tokens and image_tokens to the usage models, and those two tests assert an exact usage dict. Pinning openai==3.5.0 makes them pass again with nothing else changed — filed as #12519.

The release-note job was mine (single backticks in reST); fixed in b1c7440f.

test/components/generators/test_openai_image_generator.py is green: 23 passed, 2 skipped.