Fix ai-service Dockerfile: bake in drug_entities.json, override its path

apps/ai-service/tests/test_citation_and_intro.py

@@ -16,6 +16,7 @@ from rag.models import (
     SourceRef,
     SubjectScope,
 )
+from rag.ports import AnswerGenerationUnavailable
 from rag.prompt import build_request
 
 
@@ -41,22 +42,31 @@ class _Routing:
 
 
 class _Generator:
-    def __init__(self, payload: dict, entailment_payload: dict | None = None) -> None:
+    def __init__(
+        self, payload: dict, entailment_payload: dict | None = None,
+        sufficiency_payload: dict | None = None,
+    ) -> None:
         self._payload = payload
         self._entailment_payload = entailment_payload or {
             "entailed": True,
             "unsupported": [],
         }
+        self._sufficiency_payload = sufficiency_payload
 
     def generate(self, system: str, user: str, schema: dict) -> str:  # noqa: ARG002
-        # `_generate` also runs a post-generation entailment check; tell the
-        # two request shapes apart by schema so callers here only need to
-        # fake the main answer, not both.
-        payload = (
-            self._entailment_payload
-            if "entailed" in schema.get("properties", {})
-            else self._payload
-        )
+        # `_generate` also runs a post-generation entailment check, and
+        # `_check_sufficiency` runs its own separate call before that — tell
+        # the three request shapes apart by schema so a test asserting on one
+        # doesn't have to also shape a payload for the others.
+        props = schema.get("properties", {})
+        if "entailed" in props:
+            payload = self._entailment_payload
+        elif "sufficient" in props and self._sufficiency_payload is not None:
+            payload = self._sufficiency_payload
+        else:
+            payload = self._payload
+        if isinstance(payload, BaseException):
+            raise payload
         return json.dumps(payload, ensure_ascii=False)
 
 
@@ -122,6 +132,42 @@ def test_underspecified_dose_asks_instead_of_dumping():
     assert "tuổi" in g.clarification
     assert g.answer == g.clarification
     assert g.generated is False
+    assert g.quick_replies == ()
+
+
+def test_underspecified_dose_carries_quick_replies_when_the_model_offers_them():
+    result = _answerable(_evidence(0, 100), _evidence(1, 200))
+    gen = _Generator(
+        {"sufficient": False,
+         "clarifying_question": "Người lớn hay trẻ em?",
+         "quick_replies": ["Người lớn", "Trẻ em"]}
+    )
+    service = GroundedAnswerService(_Routing(result), gen)
+
+    g = service.answer("paracetamol", SubjectScope.HUMAN, QueryIntent.FACT_LOOKUP)
+
+    assert g.quick_replies == ("Người lớn", "Trẻ em")
+
+
+def test_sufficiency_check_outage_fails_open_to_generation_not_abstain():
+    """F-10: `_check_sufficiency` documents (and this pins) a deliberate
+    fail-OPEN on provider outage — unlike every other failure mode in this
+    service, a sufficiency-check outage does not abstain, it just skips the
+    clarify heuristic and lets grounding/entailment (tested elsewhere) be
+    the real safety net on whatever gets generated next."""
+    result = _answerable(_evidence(0, 100), _evidence(1, 200))
+    gen = _Generator(
+        {"answer": "Đoạn bằng chứng 0 [1].", "evidence_sufficient": True,
+         "clarifying_question": None},
+        sufficiency_payload=AnswerGenerationUnavailable("Bedrock unreachable"),
+    )
+    service = GroundedAnswerService(_Routing(result), gen)
+
+    g = service.answer("liều người lớn", SubjectScope.HUMAN, QueryIntent.FACT_LOOKUP)
+
+    assert g.clarification is None
+    assert g.result.decision == EvidenceDecision.ANSWERABLE
+    assert g.generated is True
 
 
 def test_sufficient_query_is_not_turned_into_a_clarification():
@@ -144,3 +190,52 @@ def test_bare_name_builds_an_intro_prompt():
     normal = build_request("Liều?", ("đoạn A",), intro=False)
     assert "CÂU HỎI:" in normal.user
     assert "GIỚI THIỆU" not in normal.user
+
+
+def test_list_mode_prompt_instructs_enumerating_every_drug_not_ranking():
+    """2026-08-07 finding: without an explicit instruction, the model picked
+    one drug out of 8 real symptom_to_drug matches and silently dropped the
+    rest — verified live. `list_mode` closes that."""
+    request = build_request("thuốc gì trị sốt", ("chỉ định A", "chỉ định B"), list_mode=True)
+    assert "LIỆT KÊ TẤT CẢ" in request.user
+    assert "không xếp hạng" in request.user or "KHÔNG" in request.user
+    assert "CÂU HỎI:" in request.user
+
+
+def test_list_mode_skips_the_sufficiency_clarify():
+    """If the sufficiency call were NOT actually skipped, it would read
+    `sufficiency_payload` (`sufficient=False`) and clarify. `list_mode=True`
+    must never call it at all, so only the real answer payload is ever read."""
+    result = _answerable(_evidence(0, 100), _evidence(1, 200))
+    gen = _Generator(
+        {"answer": "Đoạn bằng chứng 0 [1]. Đoạn bằng chứng 1 [2].",
+         "evidence_sufficient": True, "clarifying_question": None},
+        sufficiency_payload={
+            "sufficient": False, "clarifying_question": "Loại nào?", "quick_replies": [],
+        },
+    )
+    service = GroundedAnswerService(_Routing(result), gen)
+
+    g = service.answer_from_result("thuốc gì trị sốt", result, list_mode=True)
+
+    assert g.clarification is None
+    assert g.answer is not None
+    assert g.generated is True
+
+
+def test_without_list_mode_the_same_evidence_does_ask_for_clarification():
+    """Control for the test above: the same sufficiency payload, without
+    `list_mode`, must actually clarify — proving the previous test's "not
+    skipped" branch is reachable and would have failed loudly."""
+    result = _answerable(_evidence(0, 100), _evidence(1, 200))
+    gen = _Generator(
+        {"answer": "unused", "evidence_sufficient": True, "clarifying_question": None},
+        sufficiency_payload={
+            "sufficient": False, "clarifying_question": "Loại nào?", "quick_replies": [],
+        },
+    )
+    service = GroundedAnswerService(_Routing(result), gen)
+
+    g = service.answer_from_result("thuốc gì trị sốt", result, list_mode=False)
+
+    assert g.clarification == "Loại nào?"