fix(vlm): handle content_filter finish reason in API responses#3051
Conversation
|
✅ DCO Check Passed Thanks @Br1an67, all your commits are properly signed off. 🎉 |
Merge ProtectionsYour pull request matches the following merge protections and will not be merged until they are valid. 🟢 Enforce conventional commitWonderful, this rule succeeded.Make sure that we follow https://www.conventionalcommits.org/en/v1.0.0/
|
|
Related Documentation Checked 20 published document(s) in 1 knowledge base(s). No updates required. How did I do? Any feedback? |
dolfim-ibm
left a comment
dolfim-ibm
left a comment
There was a problem hiding this comment.
Overall the idea of the PR looks ok, but I'm not sure about partial success (see prev comment)
Codecov Report✅ All modified and coverable lines are covered by tests. 📢 Thoughts on this report? Let us know! |
Br1an67
force-pushed
the
fix/issue-2988-vlm-content-filter
branch
3 times, most recently
from
f314776 to
40e7ff0
Compare
|
Hi — just a gentle bump on this. Happy to make any changes if needed! |
|
@Br1an67 only missing thing is your DCO remediation commit. It is mandated by the Linux Foundation that all contributors must sign off their commits. Could you please complete this? Then we can merge. |
Add `CONTENT_FILTERED` to `VlmStopReason` enum and detect `finish_reason == "content_filter"` in `api_image_request()`. This prevents silent drops when an API provider (e.g. Azure OpenAI) filters content, and logs a warning for downstream consumers. Signed-off-by: Br1an67 <932039080@qq.com>
Add unit tests for the api_image_request function covering all finish reason types: content_filter, length, and stop/end_of_sequence. Signed-off-by: Br1an67 <932039080@qq.com>
- Convert Optional[X] to X | None for type annotations (UP045) - Replace percent format with f-strings in tests (UP031) - Prefix unused unpacked variables with underscore (RUF059) This fixes the code-checks / lint (3.12) CI failure. Signed-off-by: Br1an67 <932039080@qq.com>
Signed-off-by: Br1an67 <932039080@qq.com>
Br1an67
force-pushed
the
fix/issue-2988-vlm-content-filter
branch
from
bc90e86 to
aa85b9a
Compare
3 participants
Issue resolved by this Pull Request:
Resolves #2988
Add
CONTENT_FILTERED = "content_filter"to theVlmStopReasonenum and detectfinish_reason == "content_filter"inapi_image_request(). This prevents silent drops when an API provider (e.g. Azure OpenAI) filters content due to safety policies.Changes:
docling/datamodel/base_models.py: AddCONTENT_FILTEREDenum value toVlmStopReasondocling/utils/api_image_request.py: Detectcontent_filterfinish reason, log a warning, and returnVlmStopReason.CONTENT_FILTEREDdocling/pipeline/extraction_vlm_pipeline.py: TreatCONTENT_FILTEREDas partial success (same asLENGTHandSTOP_SEQUENCE)Checklist: