Merge pull request #685 from garylin2099/llm_mock

Reduce test time with a global LLM mock

Author: geekan

metagpt/actions/invoice_ocr.py

@@ -88,6 +88,8 @@ async def _unzip(file_path: Path) -> Path:
     async def _ocr(invoice_file_path: Path):
         ocr = PaddleOCR(use_angle_cls=True, lang="ch", page_num=1)
         ocr_result = ocr.ocr(str(invoice_file_path), cls=True)
+        for result in ocr_result[0]:
+            result[1] = (result[1][0], round(result[1][1], 2))  # round long confidence scores to reduce token costs
         return ocr_result
 
     async def run(self, file_path: Path, *args, **kwargs) -> list:

tests/conftest.py

@@ -12,57 +12,20 @@
 import os
 import re
 import uuid
-from typing import Optional
 
 import pytest
 
 from metagpt.config import CONFIG, Config
 from metagpt.const import DEFAULT_WORKSPACE_ROOT, TEST_DATA_PATH
 from metagpt.llm import LLM
 from metagpt.logs import logger
-from metagpt.provider.openai_api import OpenAILLM
 from metagpt.utils.git_repository import GitRepository
+from tests.mock.mock_llm import MockLLM
 
-
-class MockLLM(OpenAILLM):
-    rsp_cache: dict = {}
-
-    async def original_aask(
-        self,
-        msg: str,
-        system_msgs: Optional[list[str]] = None,
-        format_msgs: Optional[list[dict[str, str]]] = None,
-        timeout=3,
-        stream=True,
-    ):
-        """A copy of metagpt.provider.base_llm.BaseLLM.aask, we can't use super().aask because it will be mocked"""
-        if system_msgs:
-            message = self._system_msgs(system_msgs)
-        else:
-            message = [self._default_system_msg()] if self.use_system_prompt else []
-        if format_msgs:
-            message.extend(format_msgs)
-        message.append(self._user_msg(msg))
-        rsp = await self.acompletion_text(message, stream=stream, timeout=timeout)
-        return rsp
-
-    async def aask(
-        self,
-        msg: str,
-        system_msgs: Optional[list[str]] = None,
-        format_msgs: Optional[list[dict[str, str]]] = None,
-        timeout=3,
-        stream=True,
-    ) -> str:
-        if msg not in self.rsp_cache:
-            # Call the original unmocked method
-            rsp = await self.original_aask(msg, system_msgs, format_msgs, timeout, stream)
-            logger.info(f"Added '{rsp[:20]}' ... to response cache")
-            self.rsp_cache[msg] = rsp
-            return rsp
-        else:
-            logger.info("Use response cache")
-            return self.rsp_cache[msg]
+RSP_CACHE_NEW = {}  # used globally for producing new and useful only response cache
+ALLOW_OPENAI_API_CALL = os.environ.get(
+    "ALLOW_OPENAI_API_CALL", True
+)  # NOTE: should change to default False once mock is complete
 
 
 @pytest.fixture(scope="session")
@@ -76,16 +39,37 @@ def rsp_cache():
     else:
         rsp_cache_json = {}
     yield rsp_cache_json
-    with open(new_rsp_cache_file_path, "w") as f2:
+    with open(rsp_cache_file_path, "w") as f2:
         json.dump(rsp_cache_json, f2, indent=4, ensure_ascii=False)
+    with open(new_rsp_cache_file_path, "w") as f2:
+        json.dump(RSP_CACHE_NEW, f2, indent=4, ensure_ascii=False)
 
 
-@pytest.fixture(scope="function")
-def llm_mock(rsp_cache, mocker):
-    llm = MockLLM()
+# Hook to capture the test result
+@pytest.hookimpl(tryfirst=True, hookwrapper=True)
+def pytest_runtest_makereport(item, call):
+    outcome = yield
+    rep = outcome.get_result()
+    if rep.when == "call":
+        item.test_outcome = rep
+
+
+@pytest.fixture(scope="function", autouse=True)
+def llm_mock(rsp_cache, mocker, request):
+    llm = MockLLM(allow_open_api_call=ALLOW_OPENAI_API_CALL)
     llm.rsp_cache = rsp_cache
     mocker.patch("metagpt.provider.base_llm.BaseLLM.aask", llm.aask)
+    mocker.patch("metagpt.provider.base_llm.BaseLLM.aask_batch", llm.aask_batch)
     yield mocker
+    if hasattr(request.node, "test_outcome") and request.node.test_outcome.passed:
+        if llm.rsp_candidates:
+            for rsp_candidate in llm.rsp_candidates:
+                cand_key = list(rsp_candidate.keys())[0]
+                cand_value = list(rsp_candidate.values())[0]
+                if cand_key not in llm.rsp_cache:
+                    logger.info(f"Added '{cand_key[:100]} ... -> {cand_value[:20]} ...' to response cache")
+                    llm.rsp_cache.update(rsp_candidate)
+                RSP_CACHE_NEW.update(rsp_candidate)
 
 
 class Context:
@@ -173,6 +157,13 @@ def init_config():
     Config()
 
 
+@pytest.fixture(scope="function")
+def new_filename(mocker):
+    # NOTE: Mock new filename to make reproducible llm aask, should consider changing after implementing requirement segmentation
+    mocker.patch("metagpt.utils.file_repository.FileRepository.new_filename", lambda: "20240101")
+    yield mocker
+
+
 @pytest.fixture
 def aiohttp_mocker(mocker):
     class MockAioResponse:

tests/data/rsp_cache.json

@@ -117,7 +117,6 @@ def test_player_calculate_score_with_multiple_aces(self):
 
 
 @pytest.mark.asyncio
-@pytest.mark.usefixtures("llm_mock")
 async def test_debug_error():
     CONFIG.src_workspace = CONFIG.git_repo.workdir / uuid.uuid4().hex
     ctx = RunCodeContext(

tests/metagpt/actions/test_debug_error.py

@@ -17,7 +17,6 @@
 
 
 @pytest.mark.asyncio
-@pytest.mark.usefixtures("llm_mock")
 async def test_design_api():
     inputs = ["我们需要一个音乐播放器，它应该有播放、暂停、上一曲、下一曲等功能。", PRD_SAMPLE]
     for prd in inputs:

tests/metagpt/actions/test_design_api.py

@@ -11,7 +11,6 @@
 
 
 @pytest.mark.asyncio
-@pytest.mark.usefixtures("llm_mock")
 async def test_design_api_review():
     prd = "我们需要一个音乐播放器，它应该有播放、暂停、上一曲、下一曲等功能。"
     api_design = """

tests/metagpt/actions/test_design_api_review.py

@@ -20,7 +20,6 @@
 
 
 @pytest.mark.asyncio
-@pytest.mark.usefixtures("llm_mock")
 async def test_generate_questions():
     action = GenerateQuestions()
     rsp = await action.run(context)

tests/metagpt/actions/test_generate_questions.py

@@ -54,7 +54,6 @@ async def test_generate_table(invoice_path: Path, expected_result: dict):
     ("invoice_path", "query", "expected_result"),
     [(Path("invoices/invoice-1.pdf"), "Invoicing date", "2023年02月03日")],
 )
-@pytest.mark.usefixtures("llm_mock")
 async def test_reply_question(invoice_path: Path, query: dict, expected_result: str):
     invoice_path = TEST_DATA_PATH / invoice_path
     ocr_result = await InvoiceOCR().run(file_path=Path(invoice_path))

tests/metagpt/actions/test_invoice_ocr.py

@@ -12,7 +12,6 @@
 
 
 @pytest.mark.asyncio
-@pytest.mark.usefixtures("llm_mock")
 async def test_prepare_interview():
     action = PrepareInterview()
     rsp = await action.run("I just graduated and hope to find a job as a Python engineer")

tests/metagpt/actions/test_prepare_interview.py

@@ -18,7 +18,6 @@
 
 
 @pytest.mark.asyncio
-@pytest.mark.usefixtures("llm_mock")
 async def test_design_api():
     await FileRepository.save_file("1.txt", content=str(PRD), relative_path=PRDS_FILE_REPO)
     await FileRepository.save_file("1.txt", content=str(DESIGN), relative_path=SYSTEM_DESIGN_FILE_REPO)