Checkpoint frontend UI/UX overhaul and ingestion embed benchmark work

apps/ai-service/tests/test_agent.py

@@ -0,0 +1,212 @@
+"""`rag/agent.py` — the new orchestrator wired live in F-03.
+
+Codex's 2026-08-06 review (F-03/F-10) flagged that this module had zero test
+coverage despite being built to replace the live front end. This is the
+first coverage: routing branches with fakes, not an exhaustive golden set
+(that is F-10's job — a production-path regression suite against real
+Qdrant/Bedrock fixtures).
+"""
+from __future__ import annotations
+
+from rag.agent import RagAgent
+from rag.answer import GroundedAnswerService
+from rag.models import Evidence, EvidenceDecision, RetrievalResult, SourceRef
+from rag.understanding import QueryFrame
+
+SOURCE = SourceRef(physical_page=100, precision="region", printed_page=100)
+
+
+def _evidence(text: str) -> Evidence:
+    return Evidence(
+        evidence_id="e0", matched_doc_id="e0", kind="prose", text=text, score=1.0,
+        source_refs=(SOURCE,), hydrated_from_parent=False, requires_visual_check=False,
+    )
+
+
+class _FixedUnderstander:
+    def __init__(self, frame: QueryFrame) -> None:
+        self._frame = frame
+
+    def understand(self, turn, history=()):
+        return self._frame
+
+
+class _FixedRetrieval:
+    """Stands in for `RetrievalService.retrieve_framed` — a canned result per
+    drug_id regardless of section/query, so these tests assert routing, not
+    retrieval (that's `test_retrieval_service.py`'s job)."""
+
+    def __init__(self, results: dict[str, RetrievalResult]) -> None:
+        self._results = results
+
+    def retrieve_framed(self, drug_id, section_key, query, is_overview=False):
+        return self._results.get(
+            drug_id, RetrievalResult(EvidenceDecision.ABSTAIN, "not_configured")
+        )
+
+
+def _agent(frame: QueryFrame, results: dict[str, RetrievalResult] | None = None) -> RagAgent:
+    # `routing=None`: `answer_from_result` (the only method this path calls)
+    # never touches it — see `rag/answer.py`.
+    answers = GroundedAnswerService(routing=None)
+    return RagAgent(_FixedUnderstander(frame), _FixedRetrieval(results or {}), answers)
+
+
+def test_smalltalk_does_not_touch_retrieval():
+    agent = _agent(QueryFrame(turn_type="smalltalk"))
+    reply = agent.handle("chào bạn")
+    assert reply.decision == "answerable"
+    assert reply.turn_type == "smalltalk"
+    assert "Dược thư" in reply.answer
+
+
+def test_out_of_scope_turn_type_abstains():
+    agent = _agent(QueryFrame(turn_type="out_of_scope"))
+    reply = agent.handle("cách tiêm truyền tĩnh mạch")
+    assert reply.decision == "abstain"
+    assert reply.reason == "out_of_scope"
+
+
+def test_veterinary_phrase_abstains_even_if_the_model_missed_it():
+    # The model returned an ordinary-looking frame; the keyword backstop
+    # (`rag.policy.looks_non_human`, the same one F-02 wired server-side)
+    # still catches it before any retrieval happens.
+    agent = _agent(QueryFrame(turn_type="drug_attribute", drugs=("metformin",)))
+    reply = agent.handle("liều metformin cho chó bao nhiêu")
+    assert reply.decision == "abstain"
+    assert reply.reason == "out_of_scope"
+
+
+def test_unknown_drug_name_is_reported_not_substituted():
+    agent = _agent(QueryFrame(turn_type="drug_attribute", unknown_drugs=("aspirinol",)))
+    reply = agent.handle("liều aspirinol")
+    assert reply.decision == "abstain"
+    assert reply.reason == "drug_not_in_formulary"
+    assert "aspirinol" in reply.answer
+
+
+def test_no_drug_named_asks_which_one():
+    agent = _agent(QueryFrame(turn_type="drug_attribute"))
+    reply = agent.handle("liều dùng bao nhiêu")
+    assert reply.decision == "clarify"
+    assert reply.reason == "no_drug"
+
+
+def test_needs_clarify_frame_is_surfaced_directly():
+    agent = _agent(QueryFrame(
+        turn_type="dosing_calc", drugs=("paracetamol",),
+        needs_clarify=True, clarify_reason="Bé mấy tuổi, cân nặng bao nhiêu kg?",
+    ))
+    reply = agent.handle("liều paracetamol cho trẻ em")
+    assert reply.decision == "clarify"
+    assert reply.clarification == "Bé mấy tuổi, cân nặng bao nhiêu kg?"
+
+
+def test_single_drug_attribute_retrieves_and_answers():
+    result = RetrievalResult(
+        EvidenceDecision.ANSWERABLE, "grounded_evidence_available",
+        (_evidence("Liều 500 mg mỗi ngày."),), resolved_drug_id="metformin",
+    )
+    agent = _agent(
+        QueryFrame(turn_type="drug_attribute", drugs=("metformin",),
+                   attribute="lieu_luong_va_cach_dung"),
+        {"metformin": result},
+    )
+    reply = agent.handle("liều metformin")
+    assert reply.decision == "answerable"
+    assert reply.drugs == ("metformin",)
+    assert "500 mg" in reply.answer
+
+
+def test_interaction_combines_both_drugs_evidence():
+    warfarin = RetrievalResult(
+        EvidenceDecision.ANSWERABLE, "grounded_evidence_available",
+        (_evidence("Tương tác với aspirin làm tăng nguy cơ chảy máu."),),
+    )
+    aspirin = RetrievalResult(
+        EvidenceDecision.ANSWERABLE, "grounded_evidence_available",
+        (_evidence("Tương tác với warfarin làm tăng nguy cơ chảy máu."),),
+    )
+    agent = _agent(
+        QueryFrame(turn_type="interaction", drugs=("warfarin", "aspirin")),
+        {"warfarin": warfarin, "aspirin": aspirin},
+    )
+    reply = agent.handle("warfarin với aspirin có dùng chung được không")
+    assert reply.decision == "answerable"
+    assert reply.turn_type == "interaction"
+    assert "chảy máu" in reply.answer
+
+
+def test_interaction_with_no_evidence_abstains_and_does_not_imply_safety():
+    agent = _agent(
+        QueryFrame(turn_type="interaction", drugs=("drug_a", "drug_b")), {},
+    )
+    reply = agent.handle("drug_a và drug_b dùng chung được không")
+    assert reply.decision == "abstain"
+    assert reply.reason == "no_interaction_evidence"
+    assert "KHÔNG có nghĩa là an toàn" in reply.answer
+
+
+def test_symptom_to_drug_without_a_drug_name_asks_honestly_not_wired_yet():
+    agent = _agent(QueryFrame(turn_type="symptom_to_drug", indication="sốt cao"))
+    reply = agent.handle("sốt cao uống thuốc gì")
+    assert reply.decision == "clarify"
+    assert reply.reason == "reverse_lookup_not_ready"
+
+
+def test_history_is_passed_to_the_understander_on_the_next_turn():
+    received_history: list[tuple[str, ...]] = []
+
+    class _RecordingUnderstander:
+        def understand(self, turn, history=()):
+            received_history.append(tuple(history))
+            return QueryFrame(turn_type="smalltalk")
+
+    answers = GroundedAnswerService(routing=None)
+    agent = RagAgent(_RecordingUnderstander(), _FixedRetrieval({}), answers)
+
+    agent.handle("chào bạn", conversation_id="c1")
+    agent.handle("còn liều thì sao?", conversation_id="c1")
+
+    assert received_history[0] == ()
+    assert any("chào bạn" in line for line in received_history[1])
+
+
+def test_history_is_isolated_per_conversation_id():
+    received_history: list[tuple[str, ...]] = []
+
+    class _RecordingUnderstander:
+        def understand(self, turn, history=()):
+            received_history.append(tuple(history))
+            return QueryFrame(turn_type="smalltalk")
+
+    answers = GroundedAnswerService(routing=None)
+    agent = RagAgent(_RecordingUnderstander(), _FixedRetrieval({}), answers)
+
+    agent.handle("xin chào", conversation_id="a")
+    agent.handle("liều dùng bao nhiêu", conversation_id="b")
+
+    # The second call, on a different conversation id, must not see "a"'s turn.
+    assert received_history[1] == ()
+
+
+def test_autocomplete_delegates_to_the_configured_source():
+    class _Source:
+        def complete(self, prefix, k):
+            return ["metformin_id", "metoprolol_id"]
+
+    answers = GroundedAnswerService(routing=None)
+    agent = RagAgent(
+        _FixedUnderstander(QueryFrame(turn_type="smalltalk")),
+        _FixedRetrieval({}), answers, autocomplete=_Source(),
+    )
+    assert agent.complete("met") == ["Metformin Id", "Metoprolol Id"]
+
+
+def test_autocomplete_with_no_source_configured_returns_empty():
+    answers = GroundedAnswerService(routing=None)
+    agent = RagAgent(
+        _FixedUnderstander(QueryFrame(turn_type="smalltalk")),
+        _FixedRetrieval({}), answers,
+    )
+    assert agent.complete("met") == []