Checkpoint frontend UI/UX overhaul and ingestion embed benchmark work

apps/ai-service/tests/test_citation_and_intro.py

@@ -41,11 +41,23 @@ class _Routing:
 
 
 class _Generator:
-    def __init__(self, payload: dict) -> None:
+    def __init__(self, payload: dict, entailment_payload: dict | None = None) -> None:
         self._payload = payload
+        self._entailment_payload = entailment_payload or {
+            "entailed": True,
+            "unsupported": [],
+        }
 
     def generate(self, system: str, user: str, schema: dict) -> str:  # noqa: ARG002
-        return json.dumps(self._payload, ensure_ascii=False)
+        # `_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
+        )
+        return json.dumps(payload, ensure_ascii=False)
 
 
 def _answerable(*evidence: Evidence, is_overview: bool = False) -> RetrievalResult:
@@ -72,17 +84,56 @@ def test_only_cited_sources_are_returned():
     assert grounded.citations[0].printed_page_start == 200
 
 
-def test_answer_citing_nothing_falls_back_to_all_citations():
+def test_answer_citing_nothing_is_rejected_not_dressed_up_with_borrowed_citations():
     result = _answerable(_evidence(0, 100), _evidence(1, 200))
     service = GroundedAnswerService(
         _Routing(result),
-        # no [n] marker at all: rather than show zero provenance, show all.
+        # no [n] marker at all: grounding.verify rejects this outright (an
+        # uncited claim, per F-01). A generator is configured, so the
+        # rejection abstains — it must not attach every retrieved citation
+        # to dress an uncited generation up as sourced (the old behavior),
+        # and it must not silently degrade to a raw extractive quote either
+        # (owner correction, 2026-08-06: no fallback to the retired
+        # offline-extractive shape when a real generator is configured).
         _Generator({"answer": "Không có trích dẫn.", "evidence_sufficient": True}),
     )
 
     grounded = service.answer("q", SubjectScope.HUMAN, QueryIntent.FACT_LOOKUP)
 
-    assert len(grounded.citations) == 2
+    assert grounded.generated is False
+    assert grounded.answer is None
+    assert grounded.citations == ()
+    assert grounded.result.decision == EvidenceDecision.ABSTAIN
+
+
+def test_underspecified_dose_asks_instead_of_dumping():
+    """The reasoning step: a dose question spanning bands with no age/weight is
+    turned into a clarification, not the whole section."""
+    result = _answerable(_evidence(0, 100), _evidence(1, 200))
+    gen = _Generator(
+        {"sufficient": False,
+         "clarifying_question": "Bé mấy tuổi, cân nặng bao nhiêu kg?"}
+    )
+    service = GroundedAnswerService(_Routing(result), gen)
+
+    g = service.answer("paracetamol cho trẻ em", SubjectScope.HUMAN, QueryIntent.FACT_LOOKUP)
+
+    assert g.clarification is not None
+    assert "tuổi" in g.clarification
+    assert g.answer == g.clarification
+    assert g.generated is False
+
+
+def test_sufficient_query_is_not_turned_into_a_clarification():
+    result = _answerable(_evidence(0, 100), _evidence(1, 200))
+    gen = _Generator({"sufficient": True, "clarifying_question": None})
+    service = GroundedAnswerService(_Routing(result), gen)
+
+    g = service.answer("liều người lớn", SubjectScope.HUMAN, QueryIntent.FACT_LOOKUP)
+
+    # sufficiency passes; generation then runs (its payload lacks answer keys, so
+    # it falls back to the source text) — the point is no clarification fired.
+    assert g.clarification is None
 
 
 def test_bare_name_builds_an_intro_prompt():