feat: google-genai support image url pdf types

[GeLi2001]

resolves #2194

---

Note

Adds proper handling of Part.inline_data and Part.file_data (images/files) with image semconv, Phoenix-friendly input serialization, plus new tests, example, and test dependency updates.

• Instrumentation (Google GenAI):
  • Handle Part.inline_data and Part.file_data:
    • Images: emit message.content.image.image_url with data URLs or file URIs using OpenInference image semantics.
    • Non-images (e.g., PDF): record descriptive [File: mime, size/uri] in message.content.
  • Add Phoenix-friendly input serialization replacing binary data with data URLs/descriptions; integrate into _as_input_attributes.
  • Preserve image attributes during part flattening; extract tools as high-priority; downgrade unknown part handling to debug.
• Examples:
  • New examples/generate_content_with_images.py demonstrating inline/URI images, PDFs, and async usage.
• Tests:
  • Add coverage for inline_data/file_data handling and ensure existing text/function parts remain unchanged.
• Dependencies:
  • Update test extras: opentelemetry-sdk>=1.38.0, opentelemetry-exporter-otlp>=1.38.0.

^{Written by Cursor Bugbot for commit fd65407. This will update automatically on new commits. Configure here.}

[nate-mar]

will

 yield (
                    f"{MessageContentAttributes.MESSAGE_CONTENT_IMAGE}.{ImageAttributes.IMAGE_URL}",
                    data_url,
                )

get ignored or inadvertently converted into a singular message_content in _flatten_parts? Seems like we only check for certain attributes, otherwise we flatten into a single message. Don't we want to preserve this image_url attribute?

[GeLi2001]

yess thanks for bringing this up, I have reviewed and updated the pr so that the attributes are exported and rendered correctly on phoenix as shown in latest screenshot above in description

[cursor]

Bug: Image Overwrite Due to Missing Indexing

Multiple image parts within a single message overwrite each other in the trace attributes. This happens because the generated OpenInference attributes for images lack indexing, causing all images to use the same key and only the last image's data to be preserved.

 

[cursor]

Bug: Incorrect Indentation Causes Debug Log Misplacement

The else clause is incorrectly indented, associating it with the inner for loop instead of the if/elif chain. This causes the debug log "Non-text part encountered" to print after processing all attributes for each part, even when they're handled correctly. This else clause is also redundant, as unsupported part types are now handled in _get_attributes_from_part.