refactor(ai): OCR sidecar canonical naming cleanup — typhoon→np-dms, remove hardcoded keys, asyncio.to_thread, ADR-040/041

tests/integration/ocr-sidecar/test_active_prompt.py

@@ -0,0 +1,96 @@
+# File: tests/integration/ocr-sidecar/test_active_prompt.py
+# Change Log:
+# - 2026-06-20: Initial creation for US3 active prompt integration tests.
+
+import sys
+from pathlib import Path
+from types import SimpleNamespace
+from unittest.mock import patch
+
+from fastapi.testclient import TestClient
+
+UNIT_DIR = Path(__file__).resolve().parents[2] / "unit" / "ocr-sidecar"
+if str(UNIT_DIR) not in sys.path:
+    sys.path.insert(0, str(UNIT_DIR))
+
+from test_path_traversal import FakeDocument, load_app
+
+
+class FakeAsyncResponse:
+    def raise_for_status(self) -> None:
+        return None
+
+    def json(self) -> dict:
+        return {"choices": [{"message": {"content": "{\"natural_text\": \"prompt result\"}"}}]}
+
+
+class FakeAsyncClient:
+    last_payload = None
+
+    def __init__(self, *args, **kwargs) -> None:
+        pass
+
+    async def post(self, url: str, json: dict, headers: dict) -> FakeAsyncResponse:
+        FakeAsyncClient.last_payload = json
+        return FakeAsyncResponse()
+
+    async def aclose(self) -> None:
+        pass
+
+
+def test_ocr_injects_system_prompt_and_dms_tags(tmp_path: Path) -> None:
+    upload_base = tmp_path / "uploads"
+    upload_base.mkdir()
+    pdf_path = upload_base / "document.pdf"
+    pdf_path.write_bytes(b"%PDF-1.4\n")
+
+    app_module = load_app(upload_base)
+    client = TestClient(app_module.app)
+
+    decision = SimpleNamespace(keep_alive_seconds=120, reason="headroom-sufficient", vram_headroom_mb=9000.0)
+    fake_client = FakeAsyncClient()
+    FakeAsyncClient.last_payload = None
+
+    # Prepare dummy message structure
+    initial_messages = [{"role": "user", "content": [{"type": "text", "text": "OCR Page content"}]}]
+
+    with patch.object(app_module, "calculate_ocr_residency", return_value=decision), \
+         patch.object(app_module, "prepare_ocr_messages", return_value=initial_messages), \
+         patch.object(app_module.fitz, "open", return_value=FakeDocument()), \
+         patch.object(app_module, "ollama_client", fake_client):
+
+        response = client.post(
+            "/ocr",
+            json={
+                "pdfPath": str(pdf_path),
+                "engine": "np-dms-ocr",
+                "system_prompt": "Custom system instruction",
+                "dms_tags": {
+                    "document_number": "true",
+                    "document_date": "true"
+                },
+                "runtime_params": {
+                    "temperature": 0.1,
+                    "top_p": 0.5,
+                    "repeat_penalty": 1.0,
+                    "max_tokens": 4096
+                }
+            },
+            headers={"X-API-Key": "test-key"}
+        )
+
+    assert response.status_code == 200
+
+    # Verify the message content in last payload sent to Ollama
+    sent_messages = FakeAsyncClient.last_payload["messages"]
+
+    # We expect system_prompt to be appended to messages[0]["content"]
+    content_list = sent_messages[0]["content"]
+
+    # Verify system prompt exists
+    system_prompt_found = any(c.get("type") == "text" and c.get("text") == "Custom system instruction" for c in content_list)
+    assert system_prompt_found, "System prompt was not injected into message content"
+
+    # Verify DMS tags instruction exists
+    dms_tags_instruction = any(c.get("type") == "text" and "<document_number>" in c.get("text") and "<document_date>" in c.get("text") for c in content_list)
+    assert dms_tags_instruction, "DMS tags instructions were not injected correctly"

tests/integration/ocr-sidecar/test_async_performance.py

@@ -0,0 +1,129 @@
+# File: tests/integration/ocr-sidecar/test_async_performance.py
+# Change Log:
+# - 2026-06-20: Added ADR-040 US4 async I/O performance tests for process_ocr and lifespan.
+
+import asyncio
+import inspect
+import sys
+from pathlib import Path
+from types import SimpleNamespace
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+
+UNIT_DIR = Path(__file__).resolve().parents[2] / "unit" / "ocr-sidecar"
+if str(UNIT_DIR) not in sys.path:
+    sys.path.insert(0, str(UNIT_DIR))
+
+from test_path_traversal import load_app
+
+
+class FakeAsyncResponse:
+    """จำลอง httpx.AsyncClient response"""
+
+    def raise_for_status(self) -> None:
+        return None
+
+    def json(self) -> dict:
+        return {"choices": [{"message": {"content": '{"natural_text": "ok"}'}}]}
+
+
+class FakeAsyncClient:
+    """จำลอง httpx.AsyncClient สำหรับ async process_ocr"""
+
+    def __init__(self, *args, **kwargs) -> None:
+        self.payload = None
+        FakeAsyncClient.last_payload = None
+
+    async def post(self, url: str, json: dict, headers: dict) -> FakeAsyncResponse:
+        self.payload = json
+        FakeAsyncClient.last_payload = json
+        return FakeAsyncResponse()
+
+    async def aclose(self) -> None:
+        pass
+
+
+FakeAsyncClient.last_payload = None
+
+
+def test_process_ocr_is_coroutine_function(tmp_path: Path) -> None:
+    """T042: process_ocr ต้องเป็น async def (coroutine function)"""
+    app_module = load_app(tmp_path)
+    assert inspect.iscoroutinefunction(app_module.process_ocr), (
+        "process_ocr must be async def per ADR-040 US4"
+    )
+
+
+def test_process_pdf_doc_is_coroutine_function(tmp_path: Path) -> None:
+    """T042: _process_pdf_doc ต้องเป็น async def เพราะเรียก process_ocr"""
+    app_module = load_app(tmp_path)
+    assert inspect.iscoroutinefunction(app_module._process_pdf_doc), (
+        "_process_pdf_doc must be async def per ADR-040 US4"
+    )
+
+
+def test_app_uses_lifespan_not_startup_event(tmp_path: Path) -> None:
+    """T045: app ต้องใช้ lifespan context manager ไม่ใช่ @app.on_event('startup')"""
+    app_module = load_app(tmp_path)
+    app_obj = app_module.app
+    # FastAPI เก็บ lifespan ใน app.router.lifespan_context
+    assert hasattr(app_obj.router, "lifespan_context"), (
+        "App must use lifespan parameter, not @app.on_event('startup')"
+    )
+    # ตรวจสอบว่าไม่มี startup event handlers แบบเดิม
+    startup_handlers = app_obj.router.on_startup
+    assert len(startup_handlers) == 0, (
+        "App must not register @app.on_event('startup') handlers"
+    )
+
+
+def test_app_has_async_client_global(tmp_path: Path) -> None:
+    """T043: app module ต้องมี ollama_client global สำหรับ AsyncClient"""
+    app_module = load_app(tmp_path)
+    assert hasattr(app_module, "ollama_client"), (
+        "app module must have ollama_client global for shared AsyncClient"
+    )
+
+
+def test_normalize_endpoint_removed(tmp_path: Path) -> None:
+    """T054: /normalize endpoint ต้องถูกลบออกแล้ว"""
+    app_module = load_app(tmp_path)
+    routes = [r.path for r in app_module.app.routes]
+    assert "/normalize" not in routes, (
+        "/normalize endpoint must be removed per ADR-040 D2"
+    )
+
+
+def test_concurrent_ocr_requests_dont_block(tmp_path: Path) -> None:
+    """T041: concurrent OCR requests ต้องไม่ block กัน (async I/O)"""
+    app_module = load_app(tmp_path)
+
+    decision = SimpleNamespace(
+        keep_alive_seconds=60,
+        reason="headroom-sufficient",
+        vram_headroom_mb=9000.0,
+    )
+
+    fake_client = FakeAsyncClient()
+
+    async def run_concurrent() -> list[str]:
+        """รัน process_ocr 3 ครั้งพร้อมกัน วัดว่าไม่ block"""
+        with (
+            patch.object(app_module, "calculate_ocr_residency", return_value=decision),
+            patch.object(app_module, "prepare_ocr_messages", return_value=[{"content": []}]),
+            patch.object(app_module, "ollama_client", fake_client),
+        ):
+            tasks = [
+                app_module.process_ocr("/tmp/test.pdf", page_num=i + 1)
+                for i in range(3)
+            ]
+            results = await asyncio.gather(*tasks)
+            return results
+
+    results = asyncio.run(run_concurrent())
+    assert len(results) == 3
+    assert all(r == "ok" for r in results)
+    # ทุก request ต้องส่ง payload ได้สำเร็จ
+    assert FakeAsyncClient.last_payload is not None
+    assert FakeAsyncClient.last_payload["keep_alive"] == 60

tests/integration/ocr-sidecar/test_cpu_fallback.py

@@ -0,0 +1,49 @@
+# File: tests/integration/ocr-sidecar/test_cpu_fallback.py
+# Change Log:
+# - 2026-06-20: Added ADR-040 CPU fallback integration coverage for retrieval endpoints.
+
+from pathlib import Path
+from unittest.mock import MagicMock, patch
+
+from fastapi.testclient import TestClient
+
+import sys
+
+UNIT_DIR = Path(__file__).resolve().parents[2] / "unit" / "ocr-sidecar"
+if str(UNIT_DIR) not in sys.path:
+    sys.path.insert(0, str(UNIT_DIR))
+
+from test_path_traversal import load_app
+
+
+def test_embed_uses_cpu_when_vram_headroom_is_low(tmp_path: Path) -> None:
+    app_module = load_app(tmp_path)
+    client = TestClient(app_module.app)
+    bge_model = MagicMock()
+    bge_model.encode.return_value = {
+        "dense_vecs": [[0.1, 0.2]],
+        "lexical_weights": [{"101": 0.5}],
+    }
+    headroom = MagicMock(total_mb=16384.0, used_mb=15000.0, available_mb=1000.0, query_success=True)
+    with patch.object(app_module, "bge_model", bge_model), patch.object(app_module, "get_vram_headroom", return_value=headroom):
+        response = client.post("/embed", json={"text": "hello"}, headers={"X-API-Key": "test-key"})
+    assert response.status_code == 200
+    assert response.json()["device"] == "cpu"
+    bge_model.model.to.assert_called_with("cpu")
+
+
+def test_rerank_uses_cpu_when_vram_headroom_is_low(tmp_path: Path) -> None:
+    app_module = load_app(tmp_path)
+    client = TestClient(app_module.app)
+    reranker = MagicMock()
+    reranker.compute_score.return_value = [0.9]
+    headroom = MagicMock(total_mb=16384.0, used_mb=15000.0, available_mb=1000.0, query_success=True)
+    with patch.object(app_module, "reranker", reranker), patch.object(app_module, "get_vram_headroom", return_value=headroom):
+        response = client.post(
+            "/rerank",
+            json={"query": "q", "chunks": ["chunk"]},
+            headers={"X-API-Key": "test-key"},
+        )
+    assert response.status_code == 200
+    assert response.json()["device"] == "cpu"
+    reranker.model.to.assert_called_with("cpu")

tests/integration/ocr-sidecar/test_parameter_governance.py

@@ -0,0 +1,81 @@
+# File: tests/integration/ocr-sidecar/test_parameter_governance.py
+# Change Log:
+# - 2026-06-20: Initial creation for US3 parameter governance integration tests.
+
+import sys
+from pathlib import Path
+from types import SimpleNamespace
+from unittest.mock import patch
+
+from fastapi.testclient import TestClient
+
+UNIT_DIR = Path(__file__).resolve().parents[2] / "unit" / "ocr-sidecar"
+if str(UNIT_DIR) not in sys.path:
+    sys.path.insert(0, str(UNIT_DIR))
+
+from test_path_traversal import FakeDocument, load_app
+
+
+class FakeAsyncResponse:
+    def raise_for_status(self) -> None:
+        return None
+
+    def json(self) -> dict:
+        return {"choices": [{"message": {"content": "{\"natural_text\": \"governed result\"}"}}]}
+
+
+class FakeAsyncClient:
+    last_payload = None
+
+    def __init__(self, *args, **kwargs) -> None:
+        pass
+
+    async def post(self, url: str, json: dict, headers: dict) -> FakeAsyncResponse:
+        FakeAsyncClient.last_payload = json
+        return FakeAsyncResponse()
+
+    async def aclose(self) -> None:
+        pass
+
+
+def test_ocr_uses_governed_runtime_parameters(tmp_path: Path) -> None:
+    upload_base = tmp_path / "uploads"
+    upload_base.mkdir()
+    pdf_path = upload_base / "document.pdf"
+    pdf_path.write_bytes(b"%PDF-1.4\n")
+
+    app_module = load_app(upload_base)
+    client = TestClient(app_module.app)
+
+    decision = SimpleNamespace(keep_alive_seconds=120, reason="headroom-sufficient", vram_headroom_mb=9000.0)
+    fake_client = FakeAsyncClient()
+    FakeAsyncClient.last_payload = None
+
+    with patch.object(app_module, "calculate_ocr_residency", return_value=decision), \
+         patch.object(app_module, "prepare_ocr_messages", return_value=[{"content": []}]), \
+         patch.object(app_module.fitz, "open", return_value=FakeDocument()), \
+         patch.object(app_module, "ollama_client", fake_client):
+
+        response = client.post(
+            "/ocr",
+            json={
+                "pdfPath": str(pdf_path),
+                "engine": "np-dms-ocr",
+                "runtime_params": {
+                    "temperature": 0.7,
+                    "top_p": 0.9,
+                    "repeat_penalty": 1.1,
+                    "max_tokens": 4096
+                }
+            },
+            headers={"X-API-Key": "test-key"}
+        )
+
+    assert response.status_code == 200
+    assert response.json()["text"] == "governed result"
+
+    # Check that parameters were passed to Ollama payload
+    assert FakeAsyncClient.last_payload["temperature"] == 0.7
+    assert FakeAsyncClient.last_payload["top_p"] == 0.9
+    assert FakeAsyncClient.last_payload["repetition_penalty"] == 1.1
+    assert FakeAsyncClient.last_payload["max_tokens"] == 4096