Wire up query history: localStorage session persistence + sidebar UI

apps/ai-service/tests/test_understanding.py

@@ -110,6 +110,58 @@ def test_golden_named_drug_section_overrides_a_misclassified_relation_frame():
     assert frame.needs_clarify is False
 
 
+def test_two_sections_named_at_once_clarifies_instead_of_silently_narrowing():
+    """Regression for `abstain/incomplete_answer` reproduced live 2026-08-14
+    on "Chỉ định và chống chỉ định của Aspirin là gì?": the turn used to
+    silently collapse to whichever one section `_apply_named_drug_cues`
+    picked (chống chỉ định, being the longer phrase), so retrieval only
+    fetched that section's evidence while the question handed to generation
+    still promised both — a real, quote-backed completeness gap the
+    generator could never close. Must now clarify instead."""
+    understander = LlmQueryUnderstander(_FixedLlm({
+        "turn_type": "drug_attribute",
+        "drugs": ["paracetamol_acetaminophen"],
+        "unknown_drugs": [],
+        "attribute": "chong_chi_dinh",
+        "population": None,
+        "weight_kg": None,
+        "age_text": None,
+        "indication": None,
+        "needs_clarify": False,
+        "clarify_reason": None,
+    }), CATALOG, RESOLVER)
+
+    frame = understander.understand(
+        "Chỉ định và chống chỉ định của Paracetamol là gì?"
+    )
+
+    assert frame.turn_type == "drug_attribute"
+    assert frame.drugs == ("paracetamol_acetaminophen",)
+    assert frame.attribute is None
+    assert frame.needs_clarify is True
+    assert set(frame.quick_replies) == {"Chỉ định", "Chống chỉ định"}
+
+
+def test_single_section_named_is_unaffected_by_the_multi_section_clarify():
+    understander = LlmQueryUnderstander(_FixedLlm({
+        "turn_type": "drug_attribute",
+        "drugs": ["paracetamol_acetaminophen"],
+        "unknown_drugs": [],
+        "attribute": "chong_chi_dinh",
+        "population": None,
+        "weight_kg": None,
+        "age_text": None,
+        "indication": None,
+        "needs_clarify": False,
+        "clarify_reason": None,
+    }), CATALOG, RESOLVER)
+
+    frame = understander.understand("Chống chỉ định của Paracetamol là gì?")
+
+    assert frame.attribute == "chong_chi_dinh"
+    assert frame.needs_clarify is False
+
+
 def test_exact_candidate_does_not_repeat_the_catalog_wide_fuzzy_scan():
     resolver = _FakeResolver({"metformin": "metformin"})
     understander = LlmQueryUnderstander(_FixedLlm({
@@ -244,11 +296,11 @@ def test_quick_replies_are_parsed_when_the_model_offers_them():
 
 
 def test_quick_replies_are_dynamic_but_bounded_before_becoming_ui_chips():
-    # 19 distinct valid entries (after "  Người   lớn  " / "người lớn" dedup
-    # and the non-string 12 are dropped) so the 18-item cap — one per
+    # 20 distinct valid entries (after "  Người   lớn  " / "người lớn" dedup
+    # and the non-string 12 are dropped) so the 19-item cap — one per
     # monograph section, see rag/sections.py SECTION_ORDER — still trims
     # the last one, not just the old 4-item cap.
-    extra = [f"Lựa chọn {i}" for i in range(6, 20)]
+    extra = [f"Lựa chọn {i}" for i in range(6, 21)]
     understander = LlmQueryUnderstander(_FixedLlm({
         "turn_type": "dosing_calc", "drugs": ["paracetamol_acetaminophen"],
         "unknown_drugs": [], "attribute": None, "population": None,
@@ -263,11 +315,11 @@ def test_quick_replies_are_dynamic_but_bounded_before_becoming_ui_chips():
 
     frame = understander.understand("liều paracetamol")
 
-    assert len(frame.quick_replies) == 18
+    assert len(frame.quick_replies) == 19
     assert frame.quick_replies[:5] == (
         "Người lớn", "Trẻ em", "Phụ nữ có thai", "Người cao tuổi", "Lựa chọn thứ năm"
     )
-    assert "Lựa chọn 19" not in frame.quick_replies
+    assert "Lựa chọn 20" not in frame.quick_replies
 
 
 def test_string_false_does_not_turn_into_a_clarification():