[pcbeingused333]

Continues #4561.

api_image_request_streaming read the answer from delta.content only, and a server error inside the stream ended the call as if it had succeeded.

• Fallback. The non-streaming path (_extract_generated_text) falls back to reasoning_content and then to the text in the tool-call arguments when content is empty, because some OpenAI-compatible servers route the whole answer there. The streaming path now does the same: it collects delta.reasoning_content and the delta.tool_calls[].function.arguments fragments (joined per call index), and uses them when no content was streamed. content still wins, and the stoppers still look at content only.
• Error chunks. An SSE chunk of the form data: {"error": {...}} with no choices was skipped like any unexpected shape and the stream ended as END_OF_SEQUENCE with whatever text had arrived. It now ends the call the way an HTTP error does: an empty result with error set and INFERENCE_ERROR, so the page is reported as failed instead of as a truncated success.

Checklist:

• Documentation has been updated, if necessary. (not needed: no user-facing option changes)
• Examples have been added, if necessary. (not needed)
• Tests have been added, if necessary.

Review evidence: four tests in tests/test_api_image_request.py (TestApiImageRequestStreamingFallbacks): reasoning-only stream, tool-call-argument stream split over two chunks, content taking precedence over reasoning_content, and an error chunk after partial text. On main three of them fail (the precedence one already passed); with the change tests/test_api_image_request.py and tests/test_api_usage_propagation.py pass (28 tests) and ruff check is clean. mypy, the full suite and requests against a real server were not run.

This change was prepared with AI assistance (Claude Code).

🤖 Generated with Claude Code

[github-actions]

✅ DCO Check Passed

Thanks @pcbeingused333, all your commits are properly signed off. 🎉

[mergify]

Merge Protections

🟢 Merge protection satisfied — ready to merge.

Show 1 satisfied protection

🟢 Enforce conventional commit

Make sure that we follow https://www.conventionalcommits.org/en/v1.0.0/

• title ~= ^(fix|feat|docs|style|refactor|perf|test|build|ci|chore|revert)(?:\(.+\))?(!)?:

[github-actions]

AI triage (advisory)

Checked commit b4c090f3e4. The analysis read the diff only. It did not run code from this pull request.

The streaming API request path now uses reasoning_content or tool-call arguments as the answer when no content arrives. An error chunk in the stream (data: {"error": ...}) now returns a failed result with INFERENCE_ERROR.

Topic labels added: bug, vlm-pipeline, error-handling.

CI safety: no concern found. A maintainer must still approve CI.

1 related pull request(s) in the same code

PR	State	Signal	Assessment
#4561	merged	referenced as #4561	related: The merged PR added _failed_streaming_request and error handling for HTTP and connection failures. This PR builds on that helper and adds the reasoning_content/tool-call fallback and in-stream error chunks. These are different changes.

_{The ai:* labels are advisory. They do not approve CI and they do not add tests:full. A new push runs this check again.}

[github-actions]

AI first review (advisory): no blocker found

This review covers the full diff at b4c090f3e4. It read the code and did not run it.

I checked that streaming fallbacks for reasoning content and tool-call arguments align with the non-streaming path, and that SSE error chunks are correctly caught to report failures instead of truncations. All added test cases are complete and valid.

_{A maintainer makes the merge decision. This review never approves the pull request.}

[codecov]

Codecov Report

❌ Patch coverage is 86.95652% with 3 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
docling/utils/api_image_request.py	86.95%	0 Missing and 3 partials ⚠️

📢 Thoughts on this report? Let us know!

[PeterStaar-IBM]

@cau-git I think we need to think about this one ...

[PeterStaar-IBM]

I am not sure we want the reasoning or tool-call in the text. It would be cleaner to have dedicated fields for those in the return object and assign them.

[PeterStaar-IBM]

@cau-git ^^ what is your opinion