DeepTutor: Lifelong Personalized Tutoring. https://deeptutor.info/.
Python ai-agentsai-tutorclawbotcli-tooldeepresearchinteractive-learninglearningmulti-agent-systemsrag
fix(mastery): budget distinct pixels in grouped assessments (#1901)
Bingxi Zhao (Frank)e63bb09e6c40
diff --git a/deeptutor/learning/semantic_assessment.py b/deeptutor/learning/semantic_assessment.pyindex 56af77fba..04255086e 100644--- a/deeptutor/learning/semantic_assessment.py+++ b/deeptutor/learning/semantic_assessment.py@@ -94,6 +94,7 @@ async def _request(config, instruction, payload, sources=None, native_images=Nonfrom deeptutor.learning.lesson_evidence import _json_responsefrom deeptutor.services.llm.config import get_llm_configfrom deeptutor.services.llm.factory import complete_with_config+ from deeptutor.services.llm.image_replay import deduplicate_user_imagestry:messages = [@@ -113,6 +114,9 @@ async def _request(config, instruction, payload, sources=None, native_images=Nonmessages.append(await asyncio.to_thread(image_message, list(unique.values())))if native_images:messages.append({"role": "user", "content": native_images})+ # A group can assess several relationships on the same pages (#1901).+ # Count actual unique pixels; retain each source label via references.+ messages = deduplicate_user_images(messages)image_parts = [partfor message in messages@@ -137,6 +141,8 @@ async def _request(config, instruction, payload, sources=None, native_images=Nonallow_image_fallback=False,)return _json_response(response)+ except AssessmentUnavailable:+ raiseexcept Exception as exc:raise AssessmentUnavailable("Assessment is unavailable; retry. No grade was saved."diff --git a/tests/capabilities/test_mastery_semantic_assessment.py b/tests/capabilities/test_mastery_semantic_assessment.pyindex 603f4f16f..4b90f4143 100644--- a/tests/capabilities/test_mastery_semantic_assessment.py+++ b/tests/capabilities/test_mastery_semantic_assessment.py@@ -464,7 +464,10 @@ async def test_source_assessment_sends_pixels_with_no_silent_fallback(monkeypatcasync def complete(**kwargs):assert kwargs["allow_image_fallback"] is False- assert kwargs["messages"][-1] == image+ assert kwargs["messages"][-1]["role"] == "user"+ assert [+ part for part in kwargs["messages"][-1]["content"] if part["type"] == "image_url"+ ] == image["content"]return '{"verdicts":[{"question_id":"q","result":"correct","confidence":0.99,"feedback":"The visual relationship matches."}]}'monkeypatch.setattr(factory, "complete_with_config", complete)@@ -474,6 +477,77 @@ async def test_source_assessment_sends_pixels_with_no_silent_fallback(monkeypatcassert result["q"].result == "correct"[email protected]("shared", [True, False])+async def test_group_visual_assessment_budgets_unique_pixels_and_retains_source_labels(+ monkeypatch, shared+):+ from copy import deepcopy+ import json++ from deeptutor.learning import semantic_assessment as semantic+ from deeptutor.services.llm import factory++ items = [+ {+ "question_id": f"q{i}",+ "answer": "The membrane separates the two compartments.",+ "_visual_images": [+ part+ for page in range(2)+ for part in (+ {"type": "text", "text": f"Question q{i}, source page {page}"},+ {+ "type": "image_url",+ "image_url": {+ "url": f"data:image/png;base64,page-{page if shared else i * 2 + page}"+ },+ },+ )+ ],+ }+ for i in range(4)+ ]+ before = deepcopy(items)+ requests = []++ async def complete(**kwargs):+ requests.append(kwargs)+ parts = [+ part+ for message in kwargs["messages"]+ if isinstance(message["content"], list)+ for part in message["content"]+ ]+ assert sum(part["type"] == "image_url" for part in parts) == 2+ for i in range(4):+ assert any(f"Question q{i}, source page" in part.get("text", "") for part in parts)+ assert kwargs["allow_image_fallback"] is False+ return json.dumps(+ {+ "verdicts": [+ {+ "question_id": f"q{i}",+ "result": "correct",+ "confidence": 0.99,+ "feedback": "The relationship matches.",+ }+ for i in range(4)+ ]+ }+ )++ monkeypatch.setattr(factory, "complete_with_config", complete)+ if shared:+ verdicts = await semantic.judge_answers(object(), items)+ assert len(verdicts) == 4 and len(requests) == 1+ else:+ with pytest.raises(semantic.AssessmentUnavailable, match="split the group"):+ await semantic.judge_answers(object(), items)+ assert not requests+ assert items == before++@pytest.mark.asyncioasync def test_scoped_group_clarification_resume_and_exactly_once_credit(service, monkeypatch):from deeptutor.learning import lesson_evidence