Log the 2026-08-24 session: F3 fix live, audit filled, corpus re-ingest scoped

2026-08-24 15:07:30 +07:00
parent f3eaab0948
commit 33b16c885b
7 changed files with 575 additions and 11 deletions
@@ -163,6 +163,96 @@ def test_single_section_named_is_unaffected_by_the_multi_section_clarify():
    assert frame.needs_clarify is False


+def test_patient_comorbidity_question_answers_instead_of_forking_on_ambiguity():
+    """Regression for eval case P08, reproduced live 4/4 times 2026-08-20:
+    "BN tăng huyết áp kèm xơ gan Child-Pugh B dùng thuốc nào cần lưu ý?"
+    reliably clarified instead of running the patient-safety assessment path
+    `patient_ckd`/`patient_multi`/etc. already use successfully. The payload
+    below is the real raw understanding output captured for that turn — the
+    model read "kèm xơ gan..." as a fork in what the QUESTION means (caution
+    vs causation) and marked the condition ambiguous with its own
+    clarify_question, discarding the richly-parsed `patient_context` it
+    produced in the very same call."""
+    understander = LlmQueryUnderstander(_FixedLlm({
+        "turn_type": "condition_relation",
+        "drugs": [], "unknown_drugs": [], "attribute": None,
+        "population": None, "weight_kg": None, "age_text": None,
+        "indication": None,
+        "condition": {
+            "original_text": "tăng huyết áp kèm xơ gan Child-Pugh B",
+            "normalized_condition": "tăng huyết áp",
+            "subtype": "Child-Pugh B",
+            "qualifiers": ["xơ gan Child-Pugh B"],
+            "ambiguous": True,
+            "clarify_question": (
+                "Anh/chị muốn biết thuốc nào cần lưu ý khi tăng huyết áp kèm "
+                "xơ gan Child-Pugh B, hay thuốc nào gây tăng huyết áp ở bệnh "
+                "nhân này?"
+            ),
+        },
+        "condition_relation": "adverse_effect",
+        "patient_context": {
+            "comorbidities": ["xơ gan Child-Pugh B"],
+            "hepatic": {"description": "xơ gan Child-Pugh B", "child_pugh": "Child-Pugh B"},
+        },
+        "context_action": "new",
+        "needs_clarify": True,
+        "clarify_reason": (
+            "Anh/chị muốn hỏi thuốc nào GÂY tăng huyết áp ở bệnh nhân xơ gan "
+            "Child-Pugh B, hay thuốc nào DÙNG ĐỂ điều trị tăng huyết áp "
+            "nhưng cần lưu ý ở bệnh nhân này?"
+        ),
+        "quick_replies": ["Thuốc nào gây tăng huyết áp?", "Thuốc nào dùng cần lưu ý?"],
+    }), CATALOG, RESOLVER)
+
+    frame = understander.understand(
+        "BN tăng huyết áp kèm xơ gan Child-Pugh B dùng thuốc nào cần lưu ý?"
+    )
+
+    assert frame.turn_type == "condition_to_drug"
+    assert frame.needs_clarify is False
+    assert frame.condition is not None
+    assert frame.condition.ambiguous is False
+    assert frame.condition.clarify_question is None
+    # The hepatic data must survive untouched -- this is what lets
+    # `_condition_to_drug`'s `assess_patient_candidates` answer safely
+    # instead of listing indications blind to the Child-Pugh B impairment.
+    assert frame.patient_context.hepatic.child_pugh == "Child-Pugh B"
+
+
+def test_bare_broad_disease_still_clarifies_despite_the_candidate_cue():
+    """The P08 fix must not swallow a genuine "which disease" fork: a bare
+    broad condition with NO patient-comorbidity cue (no "BN", "kèm", organ
+    impairment, ...) still has to ask for a subtype, same as before."""
+    understander = LlmQueryUnderstander(_FixedLlm({
+        "turn_type": "condition_relation",
+        "drugs": [], "unknown_drugs": [], "attribute": None,
+        "population": None, "weight_kg": None, "age_text": None,
+        "indication": None,
+        "condition": {
+            "original_text": "viêm gan",
+            "normalized_condition": "viêm gan",
+            "ambiguous": True,
+            "clarify_question": "Anh/chị muốn hỏi viêm gan B hay viêm gan C?",
+        },
+        "condition_relation": "indication",
+        "needs_clarify": True,
+        "clarify_reason": "Anh/chị muốn hỏi viêm gan B hay viêm gan C?",
+    }), CATALOG, RESOLVER)
+
+    frame = understander.understand("Viêm gan dùng thuốc gì?")
+
+    # `_apply_broad_condition_cue` (unaffected by this fix, and applied
+    # AFTER it) is what actually decides this case: it re-derives `condition`
+    # straight from the raw turn text via its own deterministic broad-disease
+    # regex, independent of whatever `_apply_condition_candidate_cue` did —
+    # `frame.needs_clarify` itself is reset to False either way, same as the
+    # `ambiguous_hepatitis`/`ambiguous_cancer`/`ambiguous_infection` eval
+    # cases; `agent.py` clarifies off `frame.condition.ambiguous`, not this.
+    assert frame.condition is not None
+    assert frame.condition.ambiguous is True
+
+
 def test_multi_section_clarify_does_not_inherit_a_stale_prior_attribute():
    prior = QueryFrame(
        turn_type="drug_attribute",