Wire up query history: localStorage session persistence + sidebar UI

apps/ai-service/rag/agent.py

@@ -28,6 +28,7 @@ from .clinical import ConditionRelation, MedicationCandidateAssessment
 from .models import EvidenceDecision, RetrievalResult
 from .policy import looks_non_human
 from .service import RetrievalService
+from .text import normalize_name
 from .understanding import QueryFrame, QueryUnderstander
 
 logger = logging.getLogger(__name__)
@@ -53,6 +54,27 @@ MAX_LLM_CALLS_PER_TURN = 8
 # turns on the same conversation, force a hard stop instead of asking again.
 MAX_CONSECUTIVE_CLARIFY = 4
 
+# Feature-List #18/#19: "out_of_scope" used to cover two different things
+# with one identical message that named neither — a question about data
+# the formulary genuinely never contains for ANY drug (price, vendor,
+# brand availability), and a question with nothing to do with drugs at
+# all. Found live 2026-08-14 asking both "Paracetamol giá bao nhiêu?" and
+# "Thời tiết Hà Nội hôm nay?" and getting the same vague "nằm ngoài phần
+# chuyên luận... có thể thuộc phụ lục chưa được đưa vào" — which reads as
+# "maybe added later" for data that will never be in a drug formulary.
+# Deliberately a keyword heuristic, same trade-off as policy.py's
+# looks_non_human: cheap, auditable, no model call on the abstain path.
+_OUT_OF_SCOPE_DATA_PHRASES = (
+    "gia bao nhieu", "gia ca", "gia tien", "bao nhieu tien",
+    "mua o dau", "ban o dau", "nha thuoc nao ban",
+    "thuong hieu nao", "hang san xuat", "nha san xuat",
+)
+
+
+def _looks_like_missing_data_question(query: str) -> bool:
+    normalized = normalize_name(query)
+    return any(phrase in normalized for phrase in _OUT_OF_SCOPE_DATA_PHRASES)
+
 
 class ConversationStore(Protocol):
     """Durable, cross-worker alternative to the in-process history dict —
@@ -404,11 +426,23 @@ class RagAgent:
                 turn_type=tt)
 
         if tt == "out_of_scope":
+            if _looks_like_missing_data_question(turn):
+                return AgentReply(
+                    "abstain", "out_of_scope",
+                    answer="Dược thư Quốc gia Việt Nam KHÔNG chứa giá bán, nơi bán "
+                           "hay thương hiệu/nhà sản xuất cụ thể của thuốc — đây "
+                           "không phải nội dung sách này tra cứu (sách chỉ có chỉ "
+                           "định, liều dùng, chống chỉ định, tương tác thuốc và các "
+                           "mục chuyên môn khác). Vui lòng tra các nguồn khác cho "
+                           "thông tin này.",
+                    turn_type=tt)
             return AgentReply(
                 "abstain", "out_of_scope",
-                answer="Nội dung này nằm ngoài phần chuyên luận thuốc của Dược thư "
-                       "(có thể thuộc phần hướng dẫn chung/phụ lục chưa được đưa vào). "
-                       "Tôi chưa có dữ liệu để trả lời chính xác.",
+                answer="Câu hỏi này nằm ngoài phạm vi hỗ trợ của hệ thống. Hệ "
+                       "thống chỉ tra cứu thông tin thuốc theo Dược thư Quốc gia "
+                       "Việt Nam 2018 — chỉ định, liều dùng, chống chỉ định, tác "
+                       "dụng phụ, tương tác thuốc và các mục chuyên môn khác của "
+                       "một thuốc cụ thể.",
                 turn_type=tt)
 
         if not frame.drugs:

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(

apps/ai-service/rag/ports.py

@@ -76,3 +76,12 @@ class SectionRetriever(Protocol):
 
 class ParentStore(Protocol):
     def get(self, parent_id: str) -> ParentDocument | None: ...
+
+
+class SectionListRetriever(Protocol):
+    """The per-drug section checklist Feature-List #4 needs, cheap because
+    `drug_id`/`section_key` are already-indexed Qdrant payload fields — no
+    new indexing. Separate from `SectionRetriever` (interface segregation):
+    this returns labels only, never chunk text."""
+
+    def list_sections(self, drug_id: str) -> list[tuple[str, str]]: ...

apps/ai-service/rag/sections.py

@@ -147,12 +147,27 @@ SECTION_PHRASES: dict[str, tuple[str, ...]] = {
 }
 
 
-# Book order of monograph sections (Hướng dẫn sử dụng, printed page 39). Used to
-# present a whole-drug overview when the query names the drug but no attribute —
-# typing "PARACETAMOL" should return the monograph, never a "specify an
-# attribute" dead-end.
+# Book order of monograph sections, verified 2026-08-14 against the actual
+# PDF (physical page 38 = printed page 39, "HƯỚNG DẪN SỬ DỤNG DƯỢC THƯ QUỐC
+# GIA VIỆT NAM"), not assumed from an earlier reading of this constant. The
+# guide numbers 19 items; item 1, "Tên chuyên luận thuốc", is the monograph's
+# own title/heading, not a content section with a `section_key` — items 2-19
+# are exactly these 18 keys, in exactly this order. Confirms this tuple was
+# already complete and correctly ordered for the book's own stated template.
+#
+# `ten_thuong_mai` (trade name) is real, present in the corpus (492/684
+# drugs) but is NOT one of the guide's 19 numbered items — the book's own
+# template never promises it, so there is no book-verified position to place
+# it at. Inserted right after `ten_chung_quoc_te` (generic/INN name) as the
+# most natural adjacency (same convention `understanding.py`'s `SECTION_KEYS`
+# already uses) — a judgment call, not a sourced fact, unlike the 18 above.
+#
+# Used to present a whole-drug overview when the query names the drug but no
+# attribute — typing "PARACETAMOL" should return the monograph, never a
+# "specify an attribute" dead-end.
 SECTION_ORDER: tuple[str, ...] = (
     "ten_chung_quoc_te",
+    "ten_thuong_mai",
     "ma_atc",
     "loai_thuoc",
     "dang_thuoc_va_ham_luong",
@@ -208,3 +223,37 @@ class SectionResolver:
             if f" {normalized_phrase} " in padded:
                 return SectionMatch(section_key, phrase)
         return None
+
+    def resolve_all(self, query: str) -> tuple[SectionMatch, ...]:
+        """Every distinct section a question genuinely names, not just the
+        first. Same longest-first order as `resolve()`, and the same
+        span-claiming rule: once a phrase's occurrence is accepted, any
+        shorter phrase whose only occurrence falls inside that already-
+        claimed span is a substring of it, not a second section — e.g.
+        "chỉ định" inside "chống chỉ định của X" must NOT count as a second,
+        separate mention of `chi_dinh`. A phrase counts only when it has an
+        occurrence that does not overlap any span already claimed by a
+        longer, earlier-accepted phrase. Returns `()` for no match and
+        exactly one item when the question names only one section — this is
+        a superset of `resolve()`, not a replacement for it.
+        """
+        normalized_query = normalize_name(query)
+        if not normalized_query:
+            return ()
+        padded = f" {normalized_query} "
+        claimed: list[tuple[int, int]] = []
+        seen_sections: set[str] = set()
+        matches: list[SectionMatch] = []
+        for normalized_phrase, section_key, phrase in self._index:
+            needle = f" {normalized_phrase} "
+            idx = padded.find(needle)
+            if idx == -1:
+                continue
+            span = (idx, idx + len(needle))
+            if any(span[0] < c_end and c_start < span[1] for c_start, c_end in claimed):
+                continue
+            claimed.append(span)
+            if section_key not in seen_sections:
+                seen_sections.add(section_key)
+                matches.append(SectionMatch(section_key, phrase))
+        return tuple(matches)

apps/ai-service/rag/understanding.py

@@ -118,6 +118,31 @@ SECTION_KEY_HINTS: dict[str, str] = {
     "thong_tin_quy_che": "thông tin quy chế/pháp lý",
 }
 
+# Short, chip-sized Vietnamese section names — SECTION_KEY_HINTS above is
+# full-sentence disambiguation text for the model, well over
+# _QUICK_REPLY_MAX_CHARS, so it cannot double as a quick_reply label.
+_SECTION_SHORT_LABEL: dict[str, str] = {
+    "ten_chung_quoc_te": "Tên chung quốc tế",
+    "ten_thuong_mai": "Tên thương mại",
+    "ma_atc": "Mã ATC",
+    "loai_thuoc": "Loại thuốc",
+    "dang_thuoc_va_ham_luong": "Dạng thuốc và hàm lượng",
+    "duoc_ly_va_co_che_tac_dung": "Dược lý, cơ chế tác dụng",
+    "chi_dinh": "Chỉ định",
+    "chong_chi_dinh": "Chống chỉ định",
+    "than_trong": "Thận trọng",
+    "thoi_ky_mang_thai": "Thời kỳ mang thai",
+    "thoi_ky_cho_con_bu": "Thời kỳ cho con bú",
+    "tac_dung_khong_mong_muon": "Tác dụng không mong muốn",
+    "huong_dan_xu_tri_adr": "Xử trí ADR",
+    "lieu_luong_va_cach_dung": "Liều lượng và cách dùng",
+    "tuong_tac_thuoc": "Tương tác thuốc",
+    "qua_lieu_va_xu_tri": "Quá liều và xử trí",
+    "do_on_dinh_va_bao_quan": "Độ ổn định, bảo quản",
+    "tuong_ky": "Tương kỵ",
+    "thong_tin_quy_che": "Thông tin quy chế",
+}
+
 # What kind of turn this is — the router branches on it. Deliberately explicit so a
 # symptom lookup is never silently treated as a failed drug lookup, and a two-drug
 # interaction never collapses to an "ambiguous drug" abstain.
@@ -294,7 +319,7 @@ _ALLOWED_ROUTES = {
     "uong", "tiem_tinh_mach", "tiem_bap", "tiem_duoi_da",
     "dat_truc_trang", "boi_ngoai_da", "nho_mat", "nho_mui", "khac",
 }
-_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
 
 _SYSTEM = """Bạn là bộ HIỂU CÂU HỎI cho một chatbot tra cứu Dược thư Quốc gia Việt Nam.
@@ -577,6 +602,7 @@ class LlmQueryUnderstander:
             resolved_section_key=(section_match.section_key if section_match else None),
             resolved_section_phrase=(section_match.phrase if section_match else None),
         )
+        frame = _apply_multi_section_clarify(frame, turn)
         return _merge_with_prior_frame(frame, prior_frame)
 
     @staticmethod
@@ -856,6 +882,51 @@ def _apply_named_drug_cues(
     return frame
 
 
+def _apply_multi_section_clarify(frame: QueryFrame, turn: str) -> QueryFrame:
+    """A turn naming two-or-more distinct monograph sections for one drug
+    ("chỉ định và chống chỉ định của X") cannot be served by narrowing
+    `attribute` to a single value: both the model's own JSON output and
+    `_apply_named_drug_cues` above only ever produce one `attribute`, so
+    retrieval silently fetches evidence for whichever section wins while the
+    question handed to generation still promises both. The completeness check
+    then correctly reports a real, quote-backed gap against the one section
+    actually retrieved, generation retries with the same mismatched scope,
+    and the turn fails closed as a confusing `incomplete_answer` abstain —
+    found live 2026-08-14, reproduced 2/2 on "Chỉ định và chống chỉ định của
+    Aspirin là gì?". Answering every named section at once is a larger,
+    separate change (multi-attribute retrieval); asking which one first is
+    the safe interim behavior, using the generic `needs_clarify`+
+    `clarify_reason` clarify path `RagAgent` already has (checked ahead of
+    the narrower `attribute is None` -> `missing_attribute` branch, and the
+    only one of the two that forwards `quick_replies`).
+
+    Scoped to turns already read as a single named drug's own attribute(s)
+    — `_apply_named_drug_cues` runs first, so by this point `turn_type` is
+    already "drug_attribute"/"drug_overview" for exactly the cases this
+    bug affects; interaction/condition/dosing turns are untouched.
+    """
+    if frame.turn_type not in ("drug_attribute", "drug_overview") or not frame.drugs:
+        return frame
+    matches = _SECTION_RESOLVER.resolve_all(turn)
+    distinct_sections = tuple(dict.fromkeys(match.section_key for match in matches))
+    if len(distinct_sections) < 2:
+        return frame
+    options = tuple(
+        _SECTION_SHORT_LABEL.get(key, key) for key in distinct_sections
+    )[:_QUICK_REPLY_MAX_ITEMS]
+    return replace(
+        frame,
+        turn_type="drug_attribute",
+        attribute=None,
+        needs_clarify=True,
+        clarify_reason=(
+            "Câu hỏi nêu nhiều mục cùng lúc — anh/chị muốn xem mục nào trước?"
+        ),
+        quick_replies=options,
+        system_error=None,
+    )
+
+
 _KNOWN_FACT_LABELS: tuple[tuple[str, str], ...] = (
     ("population", "Đối tượng"),
     ("age_text", "Tuổi"),