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

apps/ai-service/rag/understanding.py

@@ -203,6 +203,18 @@ class QueryFrame:
     # ordinary clarifying question in the API response and trace — this lets
     # `RagAgent._route` surface the real cause instead.
     system_error: str | None = None
+    # The part of the turn the Dược thư cannot answer AT ALL, as the model
+    # named it — a property the book does not record (giá, nơi bán, bảo hiểm)
+    # or a comparative judgement it never makes ("hãng nào tốt nhất"). Set
+    # means refuse, and `_parse` forces `turn_type` to "out_of_scope" on it.
+    #
+    # Why a field and not another phrase list: `attribute` is validated against
+    # SECTION_KEYS and anything unrecognised collapses to None, which made
+    # "user asked for a section but did not say which" and "user asked for
+    # something the book has no section for" indistinguishable — both became
+    # `attribute=None` and both clarified. Found live 2026-08-24: "Paracetamol
+    # giá bao nhiêu?" answered "Bạn muốn hỏi liều cho người lớn hay trẻ em?".
+    unsupported_request: str | None = None
     raw: dict = field(default_factory=dict, compare=False)
 
 
@@ -213,6 +225,19 @@ FRAME_SCHEMA = {
     "drugs": ["drug_id exactly as it appears in the provided catalog list"],
     "unknown_drugs": ["a drug name the user mentioned that is NOT in the catalog"],
     "attribute": "one of the section keys provided, or null",
+    "unsupported_request": (
+        "Null in the ordinary case. Set it ONLY when the turn asks for "
+        "something the Dược thư does not contain at all, and name that thing "
+        "briefly in Vietnamese. Two kinds qualify: (a) a commercial or "
+        "administrative property the book never records — giá/giá tiền, nơi "
+        "bán/mua ở đâu, bảo hiểm chi trả, hạn dùng của một hộp cụ thể; (b) a "
+        "comparative or evaluative judgement the book never makes — 'hãng nào "
+        "tốt nhất', 'thuốc nào hay hơn', 'nên chọn loại nào'. "
+        "IMPORTANT — do NOT set it for trade names: the monograph HAS a "
+        "'Tên thương mại' section, so 'Paracetamol của hãng nào', 'biệt dược "
+        "của X' are ordinary in-scope lookups (attribute=ten_thuong_mai). "
+        "Only RANKING trade names is unsupported, not listing them."
+    ),
     "population": "tre_em | tre_so_sinh | nguoi_lon | nguoi_cao_tuoi | phu_nu_co_thai | phu_nu_cho_con_bu | suy_than | suy_gan | null",
     "weight_kg": (
         "number if a body weight is given, else null. Vietnamese casual speech "
@@ -412,6 +437,10 @@ Quy tắc bắt buộc:
   thực sự muốn hỏi điều gì (vd "Anh/chị muốn hỏi đường dùng nào ạ?"), không tự
   suy đoán lại giá trị cũ.
 - Chào hỏi/vu vơ -> "smalltalk". Ngoài phạm vi chuyên luận thuốc -> "out_of_scope".
+- Nếu câu hỏi nhắm vào thứ Dược thư không ghi (giá tiền, nơi mua, bảo hiểm) hoặc
+  đòi xếp hạng hơn kém ("hãng nào tốt nhất", "thuốc nào hay hơn"), đặt
+  "unsupported_request" nêu ngắn gọn thứ đó. Tên biệt dược CÓ trong sách (mục
+  "Tên thương mại"), nên hỏi biệt dược là hợp lệ — không đặt unsupported_request.
 - Khi needs_clarify=true, kèm "quick_replies": 2-4 phương án NGẮN cho câu hỏi lại
   đó, CHỈ khi nó thực sự có vài lựa chọn rời rạc tự nhiên (vd đối tượng: "Người
   lớn"/"Trẻ em"). Để mảng rỗng nếu cần một giá trị cụ thể không có lựa chọn ngắn
@@ -658,6 +687,24 @@ class LlmQueryUnderstander:
         if turn_type not in TURN_TYPES:
             turn_type = "drug_attribute" if drugs else "out_of_scope"
         needs_clarify = data.get("needs_clarify") is True
+        # The scope gate, applied deterministically rather than trusted to the
+        # model's own `turn_type`. When the turn asks for something the book
+        # does not contain, refusing is the only correct outcome — Feature-List
+        # F3 requires 100% of out-of-scope turns to be refused, and the failure
+        # this fixes was precisely the model saying `drug_attribute` while
+        # leaving `attribute` null, which downstream read as "which section did
+        # you mean?" and asked the user a question the book cannot answer.
+        #
+        # Deliberately unconditional: it fires even when a valid `attribute`
+        # was also parsed. A turn mixing an answerable section with an
+        # unanswerable property ("giá bao nhiêu và liều người lớn?") is refused
+        # whole rather than half-answered. Over-refusing is the safe direction
+        # for a safety threshold; the 90-case suite is the guard against
+        # over-refusing in practice.
+        unsupported_request = _clean_str(data.get("unsupported_request"))
+        if unsupported_request:
+            turn_type = "out_of_scope"
+            needs_clarify = False
         clarify_reason = _clean_str(data.get("clarify_reason"))
         quick_replies = (
             _clean_quick_replies(data.get("quick_replies"))
@@ -700,6 +747,7 @@ class LlmQueryUnderstander:
             needs_clarify=needs_clarify,
             clarify_reason=clarify_reason,
             quick_replies=quick_replies,
+            unsupported_request=unsupported_request,
             raw=data if isinstance(data, dict) else {},
         )
 
@@ -750,6 +798,17 @@ def _apply_reverse_relation_cues(frame: QueryFrame, turn: str) -> QueryFrame:
     )
 
 
+# Shared with `_apply_condition_candidate_cue` below: a turn naming these
+# signals is describing one patient's own combined profile ("BN X kèm Y"),
+# not asking the model to pick between unrelated conditions.
+_PATIENT_CONTEXT_CUES = (
+    " bn ", " benh nhan ", " nguoi benh ", " kem ", " di ung ",
+    " dang dung ", " mang thai ", " cho con bu ", " tuoi ", " kg ",
+    " ckd ", " suy than ", " suy gan ", " child pugh ", " egfr ",
+    " creatinin ", " ast ", " alt ",
+)
+
+
 def _apply_condition_candidate_cue(
     frame: QueryFrame, turn: str, normalizer: ConditionNormalizer
 ) -> QueryFrame:
@@ -774,6 +833,23 @@ def _apply_condition_candidate_cue(
     )
     if not any(cue in text for cue in candidate_cues):
         return frame
+    if condition.ambiguous and any(cue in text for cue in _PATIENT_CONTEXT_CUES):
+        # Found live 2026-08-20 (eval case P08): "BN tăng huyết áp kèm xơ gan
+        # Child-Pugh B dùng thuốc nào cần lưu ý?" reliably clarified instead
+        # of answering, 4/4 reproductions. The raw understanding call reads
+        # a comorbidity ("kèm xơ gan...") as a FORK in what the question
+        # means ("thuốc nào cần lưu ý" vs "thuốc nào gây tăng huyết áp") and
+        # marks the condition ambiguous with its own clarify_question — but
+        # this turn already told us which drug lane it wants (a candidate
+        # cue matched, e.g. "thuốc nào cần"), so the fork the model raised
+        # is not genuine: `_apply_general_condition_scope` already treats
+        # this same cue set as "this is one patient's profile, not a choice
+        # between diseases", and `frame.patient_context` (hepatic/renal/etc,
+        # parsed separately and left untouched here) is exactly what lets
+        # `_condition_to_drug`'s `assess_patient_candidates` answer safely
+        # instead — clearing the stale ambiguity is what lets a turn reach
+        # that path instead of dead-ending in a clarify loop.
+        condition = replace(condition, ambiguous=False, clarify_question=None)
     return replace(
         frame,
         turn_type="condition_to_drug",
@@ -799,13 +875,7 @@ def _apply_general_condition_scope(frame: QueryFrame, turn: str) -> QueryFrame:
     if frame.turn_type not in {"condition_to_drug", "symptom_to_drug"}:
         return frame
     text = f" {normalize_name(turn)} "
-    patient_cues = (
-        " bn ", " benh nhan ", " nguoi benh ", " kem ", " di ung ",
-        " dang dung ", " mang thai ", " cho con bu ", " tuoi ", " kg ",
-        " ckd ", " suy than ", " suy gan ", " child pugh ", " egfr ",
-        " creatinin ", " ast ", " alt ",
-    )
-    if any(cue in text for cue in patient_cues):
+    if any(cue in text for cue in _PATIENT_CONTEXT_CUES):
         return frame
     primary = (
         frame.condition.normalized_condition