Wire up query history: localStorage session persistence + sidebar UI

apps/ai-service/rag/answer.py

@@ -30,7 +30,7 @@ from .routing import QueryRoutingService
 # be different prompts (or a different judge) to be independent evidence —
 # see this function's own reasoning above.
 
-_QUICK_REPLY_MAX_ITEMS = 18  # one per monograph section (see rag/sections.py SECTION_ORDER)
+_QUICK_REPLY_MAX_ITEMS = 19  # one per monograph section (see rag/sections.py SECTION_ORDER)
 _QUICK_REPLY_MAX_CHARS = 40
 logger = logging.getLogger(__name__)
 
@@ -368,6 +368,39 @@ def _presentation_for_sources(source_ids: tuple[str, ...]) -> tuple[str, str]:
     return "Trả lời", "fact_list"
 
 
+# Feature-List #14: a condition/symptom -> drug reverse lookup
+# (`list_mode=True`, `RagAgent._condition_to_drug`) states which drugs the
+# formulary's `chi_dinh` sections name for this indication — a factual
+# lookup, never a treatment ranking (see
+# `[[feedback_no_recommendation_gate]]`) — but rendered as a bare drug list
+# it reads exactly like a recommendation. Fixed, non-generated block, same
+# guarantee as `DISCLAIMER` above: the model never sees or writes this text,
+# so it cannot reword, shorten, or drop it. Always the first block; `kind`
+# matches `_plan_answer`'s forced `needs_warning=True` for `list_mode`, so
+# the UI renders it with the same visual separation as any other warning
+# block, not a plain bullet.
+_LIST_MODE_NOTICE_BLOCK = AnswerBlock(
+    title="Đọc cho đúng",
+    kind="warning",
+    claims=(
+        AnswerClaim(
+            text=(
+                "Đây là kết quả TRA CỨU các thuốc mà Dược thư Quốc gia Việt "
+                "Nam ghi nhận chỉ định cho triệu chứng/bệnh này, KHÔNG PHẢI "
+                "khuyến nghị điều trị hay chỉ định lâm sàng cho một người "
+                "bệnh cụ thể."
+            ),
+        ),
+    ),
+)
+
+
+def _with_list_mode_notice(
+    blocks: tuple[AnswerBlock, ...], list_mode: bool
+) -> tuple[AnswerBlock, ...]:
+    return (_LIST_MODE_NOTICE_BLOCK, *blocks) if list_mode else blocks
+
+
 def _build_blocks(
     claims: tuple[tuple[str, tuple[int, ...]], ...],
     indexed: list[tuple[str, Citation]],
@@ -437,7 +470,10 @@ def _plan_answer(query: str, result: RetrievalResult, list_mode: bool) -> Answer
         layout=layout,
         reasoning_mode="synthesis" if multi_source else "direct_lookup",
         show_heading=multi_source or verbosity == "detailed",
-        needs_warning=any(section in _WARNING_SECTIONS for section in sections),
+        # `list_mode` always carries `_LIST_MODE_NOTICE_BLOCK` (see below) —
+        # this is what gives that block its actual warning presentation in
+        # the UI, not just a plain untitled bullet.
+        needs_warning=list_mode or any(section in _WARNING_SECTIONS for section in sections),
     )
 
 
@@ -677,7 +713,7 @@ class GroundedAnswerService:
                     (text, (index,))
                     for index, text in enumerate(evidence_texts, start=1)
                 )
-                blocks = _build_blocks(claims, indexed)
+                blocks = _with_list_mode_notice(_build_blocks(claims, indexed), list_mode)
                 return GroundedAnswer(
                     result,
                     "\n".join(text for text, _ in claims),
@@ -724,7 +760,7 @@ class GroundedAnswerService:
             if 1 <= index <= len(indexed)
         ))
         citations = tuple(indexed[index - 1][1] for index in cited_indices)
-        blocks = _build_blocks(outcome.claims, indexed)
+        blocks = _with_list_mode_notice(_build_blocks(outcome.claims, indexed), list_mode)
         clean_answer = "\n".join(text for text, _ in outcome.claims)
         self._metrics.increment(metric_names.GENERATION_SERVED)
         return GroundedAnswer(