🐛 Bug Report: streaming choice events use the drained generator instead of the last chunk

Which component is this bug for?

Google Generative AI Instrumentation

📜 Description

Both streaming builders in opentelemetry/instrumentation/google_generativeai/__init__.py pass the generator to emit_choice_events instead of the chunk:

complete_response = "".join(text_parts)

if emit_events:
    emit_choice_events(response, event_logger)
else:
    if last_chunk is not None and getattr(last_chunk, "candidates", None):
        set_response_attributes(span, last_chunk, llm_model)

By that point response has been fully drained by the for item in response loop above. emit_choice_events does for index, candidate in enumerate(response.candidates), and a generator has no candidates. The non-events branch immediately below already uses last_chunk, which is the object that carries them.

This only fires when events are enabled, so use_legacy_attributes=False.

Two things make it worse than a dropped span. _build_from_streaming_response is not decorated @dont_throw, unlike the other handlers in the file, and it has no try/finally around span.end(). So the AttributeError is raised out of the generator on the final next(), into the calling application, and the span is never ended or exported.

👟 Reproduction steps

Instrument with events enabled and stream a response:

GoogleGenerativeAiInstrumentor(use_legacy_attributes=False).instrument()

for chunk in client.models.generate_content_stream(model="gemini-2.0-flash", contents="hi"):
    print(chunk.text)

👍 Expected behavior

Choice events are emitted from the final chunk, the stream is consumed cleanly, and the span is exported.

👎 Actual Behavior with Screenshots

AttributeError: 'generator' object has no attribute 'candidates'

Raised into the caller on the last iteration. Spans exported: 0. With use_legacy_attributes=True the same call streams cleanly and exports one span.

🤖 Python Version

3.12

📃 Provide any additional context for the Bug.

An empty stream needs a guard too, since last_chunk stays None there.

Separately, the missing @dont_throw and span.end() in a finally on these builders is worth addressing, but I have left that out to keep the change focused.

👀 Have you spent some time to check if this bug has been raised before?

• I checked and didn't find similar issue

🏢 Have you read the Contributing Guidelines?

• I have read the Contributing Guidelines

Are you willing to submit PR?

Yes I am willing to submit a PR!

Start in opentelemetry/instrumentation/google_generativeai/init.py and inspect both streaming builders, the response-draining loop, and emit_choice_events. Reproduce the issue with GoogleGenerativeAiInstrumentor(use_legacy_attributes=False) and generate_content_stream; done means choice events use the final chunk, empty streams are guarded, and the stream completes with its span exported.