fix(llm): reject nonconforming structured payloads

CHANGELOG.md

@@ -1,4 +1,5 @@
 ## [Unreleased]
+- LLM structured-output 계약은 알 수 없는 필드와 암묵적 타입 변환뿐 아니라 범위를 벗어나거나 유한하지 않은 project graph confidence도 provider 경계에서 거부합니다. 잘못된 모델 값을 0 또는 1로 보정해 신뢰 근거처럼 저장하지 않습니다.
 - 긴 이메일·첨부 본문을 의미 단위 청크로 임베딩한 뒤 기존 email/attachment 벡터 계약으로 평균화하고, 청크 요청·벡터 누적을 제한된 창으로 처리합니다. OpenAI `text-embedding-3-*`에는 저장 차원(`1536`)을 직접 요청하도록 보강했습니다. 합성 메일 fixture 5건(70청크)과 provider 요청 계약으로 1,536차원 벡터 경로를 검증했으며, 실행 시 선택한 임베딩 제공자에 본문·파싱된 첨부 텍스트를 전송할 수 있습니다. 회사 기밀 데이터는 fixture·commit·PR·log에 포함하지 않습니다.
 - EmailDetail 테스트가 지원하지 않는 스레드 병합/분리 버튼을 `textContent`뿐 아니라 `aria-label`과 `title` 접근 가능 이름으로도 검출하도록 바꿔, 아이콘 전용 버튼 회귀를 놓치지 않습니다.
 

backend/services/llm_service.py

@@ -9,7 +9,7 @@
 from core.exceptions import LLMServiceError
 from services.circuit_breaker import provider_circuit_breaker
 from services.retry import retry_transient
-from pydantic import BaseModel, Field
+from pydantic import BaseModel, ConfigDict, Field
 from services.llm_provider_urls import build_llm_provider_http_client
 
 logger = logging.getLogger(__name__)
@@ -26,6 +26,8 @@
 class ExtractionResult(BaseModel):
     """Result model for email extraction."""
 
+    model_config = ConfigDict(extra="forbid", strict=True)
+
     summary: str
     action_items: list[str]
     provenance: str | None = None

backend/services/project_graph/llm_extractor.py

@@ -32,7 +32,7 @@
 from typing import Iterable
 
 from openai import AsyncOpenAI
-from pydantic import BaseModel
+from pydantic import BaseModel, ConfigDict, Field
 
 from services.llm_provider_urls import build_llm_provider_http_client
 
@@ -79,25 +79,37 @@
 
 
 class ExtractedObjectPayload(BaseModel):
+    """One provider-produced project object before provenance validation."""
+
+    model_config = ConfigDict(extra="forbid", strict=True)
+
     object_type: str
     title: str
     summary: str
     source_segment_uids: list[str]
-    confidence: float
+    confidence: float = Field(ge=0.0, le=1.0, allow_inf_nan=False)
     # Stable within-response handle the model assigns so relations can reference
     # objects before their persisted uid exists. Optional for backward
     # compatibility; when omitted the object's position supplies the handle.
     local_key: str = ""
 
 
 class ExtractedRelationPayload(BaseModel):
+    """One provider-produced relation before endpoint validation."""
+
+    model_config = ConfigDict(extra="forbid", strict=True)
+
     source_local_key: str
     target_local_key: str
     relation_type: str
-    confidence: float
+    confidence: float = Field(ge=0.0, le=1.0, allow_inf_nan=False)
 
 
 class ExtractionPayload(BaseModel):
+    """Strict provider response envelope for project graph extraction."""
+
+    model_config = ConfigDict(extra="forbid", strict=True)
+
     objects: list[ExtractedObjectPayload]
     relations: list[ExtractedRelationPayload] = []
 
@@ -211,7 +223,6 @@ def _validated_objects(
         if not title or not summary:
             continue
         primary = segments_by_uid[cited[0]]
-        confidence = min(max(candidate.confidence, 0.0), 1.0)
         local_key = candidate.local_key.strip() or f"__object_{position}"
         objects.append(
             (
@@ -222,7 +233,7 @@ def _validated_objects(
                     title=title,
                     summary=summary,
                     source_segment_uids=cited,
-                    confidence=confidence,
+                    confidence=candidate.confidence,
                     extractor_name=LLM_EXTRACTOR_NAME,
                     extractor_version=LLM_EXTRACTOR_VERSION,
                     attributes={
@@ -300,7 +311,7 @@ def _relation_edges(
                 source_uid=source_object.uid,
                 target_uid=target_object.uid,
                 edge_type=relation_type,
-                confidence=min(max(relation.confidence, 0.0), 1.0),
+                confidence=relation.confidence,
                 source_segment_uids=_relation_citation(source_object, target_object),
             )
         )

backend/services/rag_service.py

@@ -13,7 +13,7 @@
 import logging
 
 from openai import AsyncOpenAI
-from pydantic import BaseModel
+from pydantic import BaseModel, ConfigDict
 
 from core.config import settings
 from services.llm_provider_urls import build_llm_provider_http_client
@@ -25,6 +25,10 @@
 
 
 class GroundedAnswerPayload(BaseModel):
+    """Strict provider response before citation validation and attribution."""
+
+    model_config = ConfigDict(extra="forbid", strict=True)
+
     answer: str
     cited_email_ids: list[int]
 

backend/tests/test_llm_structured_output_contract.py

@@ -0,0 +1,110 @@
+"""Structured payload and OpenAI JSON Schema wire contracts."""
+
+import pytest
+from pydantic import BaseModel, ValidationError
+
+from services.llm_service import ExtractionResult
+from services.project_graph.llm_extractor import ExtractionPayload
+from services.rag_service import GroundedAnswerPayload
+
+
+@pytest.mark.parametrize(
+    ("payload_model", "valid_payload"),
+    [
+        (ExtractionResult, {"summary": "s", "action_items": []}),
+        (ExtractionPayload, {"objects": [], "relations": []}),
+        (GroundedAnswerPayload, {"answer": "a", "cited_email_ids": []}),
+    ],
+)
+def test_structured_payload_models_forbid_unknown_top_level_fields(
+    payload_model: type[BaseModel],
+    valid_payload: dict[str, object],
+) -> None:
+    """Reject fields outside the product-owned structured payload schema."""
+
+    schema = payload_model.model_json_schema()
+
+    assert schema["additionalProperties"] is False
+    with pytest.raises(ValidationError):
+        payload_model.model_validate({**valid_payload, "unexpected_field": True})
+
+
+@pytest.mark.parametrize(
+    "invalid_payload",
+    [
+        {
+            "objects": [
+                {
+                    "object_type": "requirement",
+                    "title": "t",
+                    "summary": "s",
+                    "source_segment_uids": ["segment-1"],
+                    "confidence": 0.8,
+                    "unexpected_field": True,
+                }
+            ],
+            "relations": [],
+        },
+        {
+            "objects": [],
+            "relations": [
+                {
+                    "source_local_key": "a",
+                    "target_local_key": "b",
+                    "relation_type": "depends_on",
+                    "confidence": 0.8,
+                    "unexpected_field": True,
+                }
+            ],
+        },
+    ],
+)
+def test_project_graph_nested_payloads_forbid_unknown_fields(
+    invalid_payload: dict[str, object],
+) -> None:
+    """Keep nested project-graph object and relation schemas fail closed."""
+
+    with pytest.raises(ValidationError):
+        ExtractionPayload.model_validate(invalid_payload)
+
+
+def test_structured_payload_models_reject_scalar_coercion() -> None:
+    """Keep machine-readable identifiers strict instead of coercing strings."""
+
+    with pytest.raises(ValidationError):
+        GroundedAnswerPayload.model_validate(
+            {"answer": "grounded", "cited_email_ids": ["123"]}
+        )
+
+    with pytest.raises(ValidationError):
+        ExtractionResult.model_validate(
+            {"summary": "s", "action_items": [], "confidence": "90"}
+        )
+
+
+@pytest.mark.parametrize("confidence", [-0.1, 1.1, float("nan"), float("inf")])
+@pytest.mark.parametrize("payload_kind", ["object", "relation"])
+def test_project_graph_payload_rejects_invalid_confidence(
+    confidence: float,
+    payload_kind: str,
+) -> None:
+    """Reject out-of-range or non-finite model confidence at the wire boundary."""
+
+    object_payload = {
+        "object_type": "requirement",
+        "title": "t",
+        "summary": "s",
+        "source_segment_uids": ["segment-1"],
+        "confidence": confidence if payload_kind == "object" else 0.8,
+    }
+    relation_payload = {
+        "source_local_key": "a",
+        "target_local_key": "b",
+        "relation_type": "depends_on",
+        "confidence": confidence if payload_kind == "relation" else 0.8,
+    }
+
+    with pytest.raises(ValidationError):
+        ExtractionPayload.model_validate(
+            {"objects": [object_payload], "relations": [relation_payload]}
+        )

backend/tests/test_project_graph_llm_extractor.py

@@ -133,7 +133,7 @@ async def test_objects_citing_unknown_segments_are_dropped(monkeypatch):
 
 
 @pytest.mark.asyncio
-async def test_unknown_type_dropped_and_confidence_clamped(monkeypatch):
+async def test_unknown_type_is_dropped(monkeypatch):
     monkeypatch.setattr(
         llm_extractor,
         "_call_llm",
@@ -148,10 +148,10 @@ async def test_unknown_type_dropped_and_confidence_clamped(monkeypatch):
                 ),
                 llm_extractor.ExtractedObjectPayload(
                     object_type="milestone",
-                    title="Overconfident",
+                    title="Grounded milestone",
                     summary="Due next week.",
                     source_segment_uids=["seg1"],
-                    confidence=7.5,
+                    confidence=0.75,
                 ),
             )
         ),
@@ -163,7 +163,7 @@ async def test_unknown_type_dropped_and_confidence_clamped(monkeypatch):
 
     assert len(result.objects) == 1
     assert result.objects[0].object_type.value == "milestone"
-    assert result.objects[0].confidence == 1.0
+    assert result.objects[0].confidence == 0.75
 
 
 @pytest.mark.asyncio