OpenConceptLab/ocl_issues#2388 | encoder_model param for reranking

[snyaggarwal]

Linked Issue

Closes OpenConceptLab/ocl_issues#2388

[paynejd]

Review — OpenConceptLab/ocl_issues#2388

1. Narrow the exception handler (core/concepts/views.py)

The except Exception as e catches everything including SystemExit, OOM, and DB errors. Narrow to the specific failures that model loading/inference can raise:

except (ValueError, RuntimeError, OSError) as e:
    return Response({'detail': str(e)}, status=status.HTTP_400_BAD_REQUEST)

2. Change Closes to Ref in PR body

Both this PR and oclmap#8 say "Closes OpenConceptLab/ocl_issues#2388". Whichever merges first will auto-close the issue before the other ships. Suggest changing this PR body to "Ref OpenConceptLab/ocl_issues#2388" and letting oclmap#8 (the final piece) carry the Closes.

3. Consider API-level permission check for encoder_model

The issue specifies only core team users should access custom reranker config. The oclmap PR gates the UI on isCoreUser, but the $rerank/ endpoint accepts encoder_model from anyone — a non-core user could call the API directly with a custom model. Is this intentionally open at the API level? If not, add a permission check.

4. No input validation on encoder_model

An arbitrary string can be saved to the encoder_model field. The clean() method resets empty values to the default, but garbage strings persist in the DB and will cause errors on every subsequent rerank attempt. Consider a basic format check (e.g. must contain /, strip whitespace) or validate that the model can be loaded before saving.

5. Include the encoder model used in the rerank response

For traceability and debugging, include the model name in the rerank response so the frontend can log which model produced the scores for a given row:

return Response({'results': results, 'encoder_model': reranker.model_name or self.default_model})

This supports the frontend logging which model was used per row — important when users are experimenting with different reranker models.

---

Related follow-up tickets filed:

• OpenConceptLab/ocl_issues#2463 — Upgrade sentence-transformers from 3.3.1 to 5.4+
• OpenConceptLab/ocl_issues#2464 — Add Qwen3-Reranker models (blocked by #2463)

[snyaggarwal]

Addressed 1, 3 and 4.

Include the encoder model used in the rerank response

Mapper logs which model was used.

[paynejd]

Follow-up Review

Items 1, 3, and 4 from the original review are properly addressed — nice work, Sunny. Two small fixes below, then I think this is good to merge.

PR body: Closes → Ref

The body still says Closes OpenConceptLab/ocl_issues#2388 — since oclmap#8 also targets the same issue, whichever merges first will auto-close it. Suggest changing to Ref and letting the final piece carry the Closes.

CORE_USER_GROUP constant

The diff imports it from core.users.constants but doesn't show it being added — just confirming it already exists there?

Model name validation

Not a blocker, but worth noting: a typo'd model name (e.g. "not-a-real-model") will persist in the DB and cause 400s on every rerank call. A simple format check (must contain /) would catch obvious garbage. Fine to defer to a follow-up if preferred.

CodeQL alert on line 1032

The str(e) in the (ValueError, RuntimeError, OSError) handler is still flagged by GitHub Advanced Security. Probably fine for these exception types, but FYI if the team wants a clean scan.

[paynejd]

One more — bug in the generic exception handler.

[snyaggarwal]

CORE_USER_GROUP constant
The diff imports it from core.users.constants but doesn't show it being added — just confirming it already exists there?

CORE_USER_GROUP does exists in core.user.constants

Model name validation

Dont want to do strict / check, not sure if there might be model names without scopes like all-MiniLM-L6-v2, leaving it as is for now.