Add read-only production runtime audit

apps/ai-service/rag/agent.py

@@ -28,10 +28,12 @@ from .clinical import ConditionRelation, MedicationCandidateAssessment
 from .models import EvidenceDecision, RetrievalResult
 from .policy import looks_non_human
 from .service import RetrievalService
+from .sections import SectionResolver
 from .text import normalize_name
 from .understanding import QueryFrame, QueryUnderstander
 
 logger = logging.getLogger(__name__)
+_SECTION_RESOLVER = SectionResolver()
 
 TUONG_TAC = "tuong_tac_thuoc"
 HISTORY_TURNS = 6
@@ -142,7 +144,12 @@ class RagAgent:
             return []
         return [_display_name(drug_id) for drug_id in self._autocomplete.complete(prefix, k)]
 
-    def handle(self, turn: str, conversation_id: str | None = None) -> AgentReply:
+    def handle(
+        self,
+        turn: str,
+        conversation_id: str | None = None,
+        response_mode: str = "ai",
+    ) -> AgentReply:
         # F-08: one budget per turn, threaded through every LLM call this
         # turn makes (understand, then whatever `_route` reaches).
         t0 = time.monotonic()
@@ -154,7 +161,7 @@ class RagAgent:
             turn, tuple(history), budget=budget, prior_frame=prior_frame
         )
         t2 = time.monotonic()
-        reply = self._route(turn, frame, budget)
+        reply = self._route(turn, frame, budget, response_mode=response_mode)
         reply = self._enforce_clarify_circuit_breaker(conversation_id, reply)
         t3 = time.monotonic()
         if conversation_id is not None:
@@ -242,7 +249,13 @@ class RagAgent:
                 return []
         return self._history.get(conversation_id, [])
 
-    def _route(self, turn: str, frame: QueryFrame, budget: RequestBudget) -> AgentReply:
+    def _route(
+        self,
+        turn: str,
+        frame: QueryFrame,
+        budget: RequestBudget,
+        response_mode: str = "ai",
+    ) -> AgentReply:
         tt = frame.turn_type
         section_overview = _is_section_overview(turn, frame)
         if section_overview and not frame.section_overview:
@@ -253,10 +266,10 @@ class RagAgent:
         # an out-of-scope request look recoverable.
         if looks_non_human(turn):
             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.",
+                "abstain", "out_of_scope_non_human",
+                answer="Dược thư Quốc gia Việt Nam trong hệ thống này chỉ bao "
+                       "phủ thuốc dùng cho người. Hệ thống không tra cứu liều "
+                       "dùng hoặc hướng dẫn điều trị cho động vật.",
                 turn_type=tt)
 
         # Dosing is a small state machine, not an unconstrained model opinion.
@@ -407,6 +420,24 @@ class RagAgent:
                 turn_type=tt,
             )
 
+        if response_mode == "monograph" and (
+            tt == "drug_overview"
+            or (
+                tt == "drug_attribute"
+                and frame.attribute is None
+                and not frame.needs_clarify
+            )
+            or _is_bare_monograph_request(turn)
+        ) and frame.drugs:
+            return AgentReply(
+                "clarify", "select_drug_sections",
+                clarification=(
+                    "Đã nhận diện chuyên luận thuốc. Anh/chị chọn các mục cần "
+                    "xem; nếu không chọn mục nào, hệ thống sẽ hiển thị toàn bộ."
+                ),
+                drugs=frame.drugs, turn_type=tt,
+            )
+
         if tt == "drug_attribute" and frame.drugs and frame.attribute is None:
             return AgentReply(
                 "clarify", "missing_attribute",
@@ -713,6 +744,28 @@ def _is_section_overview(turn: str, frame: QueryFrame) -> bool:
     return frame.section_overview or any(cue in text for cue in overview_cues)
 
 
+def _is_bare_monograph_request(turn: str) -> bool:
+    """True for a plain drug name in explicit monograph-browse mode.
+
+    A persisted conversation can contribute a stale attribute to a new bare
+    drug turn (for example the prior question was about contraindications).
+    The UI mode is an explicit current-turn instruction, so a plain name must
+    open the picker rather than inherit that old section.  Any actual section
+    phrase or clinical-question cue keeps the normal AI route.
+    """
+    text = normalize_name(turn)
+    if not text or len(text) > 100 or _SECTION_RESOLVER.resolve_all(turn):
+        return False
+    clinical_cues = (
+        " dung ", " dieu tri ", " tuong tac ", " tac dung ", " lieu ",
+        " benh ", " thai ", " cho con bu ", " tre em ", " nguoi lon ",
+        " suy than ", " suy gan ", " di ung ", " bao nhieu ", " la gi ",
+        " co the ", " duoc khong ",
+    )
+    padded = f" {text} "
+    return not any(cue in padded for cue in clinical_cues)
+
+
 _POPULATION_LABELS = {
     "tre_em": "trẻ em",
     "tre_so_sinh": "trẻ sơ sinh",