AutoRAG 0.3.26__tar.gz → 0.3.27__tar.gz

{autorag-0.3.26 → autorag-0.3.27}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: AutoRAG
-Version: 0.3.26
+Version: 0.3.27
 Summary: Automatically Evaluate RAG pipelines with your own data. Find optimal structure for new RAG product.
 Project-URL: Homepage, https://github.com/Marker-Inc-Korea/AutoRAG
 Author-email: Marker-Inc <vkehfdl1@gmail.com>
@@ -217,7 +217,7 @@ Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
 Classifier: Topic :: Software Development :: Libraries
 Classifier: Topic :: Software Development :: Libraries :: Python Modules
 Requires-Python: <3.13,>=3.10
-Requires-Dist: aiohttp<3.14,>=3.12.13
+Requires-Dist: aiohttp<3.15,>=3.14.3
 Requires-Dist: banks>=2.4.2
 Requires-Dist: chromadb>=1.0.0
 Requires-Dist: click>=8.2.1
@@ -300,8 +300,8 @@ Requires-Dist: sudachidict-core; extra == 'all'
 Requires-Dist: sudachipy>=0.6.8; extra == 'all'
 Requires-Dist: torch>=2.7.1; extra == 'all'
 Requires-Dist: transformers>=4.51.3; extra == 'all'
-Requires-Dist: unstructured[pdf]>=0.17.2; extra == 'all'
-Requires-Dist: vllm<0.30.0,>=0.11.0; extra == 'all'
+Requires-Dist: unstructured[pdf]>=0.24.0; (python_version >= '3.11') and extra == 'all'
+Requires-Dist: vllm<0.31.0,>=0.11.0; extra == 'all'
 Requires-Dist: xgrammar<0.3.0,>=0.1.32; extra == 'all'
 Provides-Extra: gpu
 Requires-Dist: bert-score>=0.3.13; extra == 'gpu'
@@ -330,9 +330,9 @@ Requires-Dist: pdfminer-six>=20250506; extra == 'parse'
 Requires-Dist: pdfplumber>=0.11.7; extra == 'parse'
 Requires-Dist: pymupdf>=1.26.1; extra == 'parse'
 Requires-Dist: pypdf2>=3.0.1; extra == 'parse'
-Requires-Dist: unstructured[pdf]>=0.17.2; extra == 'parse'
+Requires-Dist: unstructured[pdf]>=0.24.0; (python_version >= '3.11') and extra == 'parse'
 Provides-Extra: vllm
-Requires-Dist: vllm<0.30.0,>=0.11.0; extra == 'vllm'
+Requires-Dist: vllm<0.31.0,>=0.11.0; extra == 'vllm'
 Description-Content-Type: text/markdown
 
 # AutoRAG (Legacy)

{autorag-0.3.26 → autorag-0.3.27}/autorag/embedding/vllm.py

@@ -81,7 +81,6 @@ class VllmEmbedding(MultiModalEmbedding):
 			)
 		self._client = VLLModel(
 			model=model_name,
-			task="embed",
 			max_num_seqs=embed_batch_size,
 			tensor_parallel_size=tensor_parallel_size,
 			trust_remote_code=trust_remote_code,

{autorag-0.3.26 → autorag-0.3.27}/autorag/nodes/passagereranker/jina.py

@@ -30,15 +30,35 @@ class JinaReranker(BasePassageReranker):
 					"API key is not provided."
 					"You can set it as an argument or as an environment variable 'JINAAI_API_KEY'"
 				)
-		self.session = aiohttp.ClientSession(loop=get_event_loop())
-		self.session.headers.update(
-			{"Authorization": f"Bearer {api_key}", "Accept-Encoding": "identity"}
-		)
+		# Open the session lazily on the loop that runs the requests. aiohttp's loop=
+		# kwarg is a deprecated no-op, so a session created here would bind to whatever
+		# loop happens to exist at construction time instead of the request loop.
+		self._headers = {
+			"Authorization": f"Bearer {api_key}",
+			"Accept-Encoding": "identity",
+		}
+		self.session = None
+
+	def _ensure_session(self) -> aiohttp.ClientSession:
+		if self.session is None or self.session.closed:
+			self.session = aiohttp.ClientSession(headers=self._headers)
+		return self.session
 
 	def __del__(self):
-		self.session.close()
-		del self.session
-		super().__del__()
+		# Best-effort safety net. _pure closes the session on the request loop; calling
+		# session.close() without awaiting it leaks the session and warns at shutdown.
+		try:
+			session = getattr(self, "session", None)
+			if session is not None and not session.closed:
+				loop = get_event_loop()
+				if not loop.is_running() and not loop.is_closed():
+					loop.run_until_complete(session.close())
+		except Exception:
+			pass
+		try:
+			super().__del__()
+		except Exception:
+			pass
 
 	@result_to_dataframe(["retrieved_contents", "retrieved_ids", "retrieve_scores"])
 	def pure(self, previous_result: pd.DataFrame, *args, **kwargs):
@@ -71,14 +91,25 @@ class JinaReranker(BasePassageReranker):
 		:param batch: The number of queries to be processed in a batch
 		:return: Tuple of lists containing the reranked contents, ids, and scores
 		"""
-		tasks = [
-			jina_reranker_pure(
-				self.session, query, contents, ids, top_k=top_k, model=model
-			)
-			for query, contents, ids in zip(queries, contents_list, ids_list)
-		]
 		loop = get_event_loop()
-		results = loop.run_until_complete(process_batch(tasks, batch))
+
+		async def _run_batch():
+			# Create the session here, on the loop that runs the requests. aiohttp 3.13
+			# resolves the session loop from the running loop and rejects loop=.
+			session = self._ensure_session()
+			tasks = [
+				jina_reranker_pure(
+					session, query, contents, ids, top_k=top_k, model=model
+				)
+				for query, contents, ids in zip(queries, contents_list, ids_list)
+			]
+			try:
+				return await process_batch(tasks, batch)
+			finally:
+				if not session.closed:
+					await session.close()
+
+		results = loop.run_until_complete(_run_batch())
 
 		content_result, id_result, score_result = zip(*results)
 

{autorag-0.3.26 → autorag-0.3.27}/autorag/nodes/passagereranker/nvidia.py

@@ -29,21 +29,36 @@ class NvidiaReranker(BasePassageReranker):
 				"or directly set it on the config YAML file."
 			)
 		self.invoke_url = "https://ai.api.nvidia.com/v1/retrieval/nvidia/reranking"
-		self.session = aiohttp.ClientSession(loop=get_event_loop())
-		self.session.headers.update(
-			{"Authorization": f"Bearer {self.api_key}", "Accept": "application/json"}
-		)
+		# Open the session lazily on the loop that runs the requests. aiohttp's loop=
+		# kwarg is a deprecated no-op, so a session created here would bind to whatever
+		# loop happens to exist at construction time instead of the request loop.
+		self._headers = {
+			"Authorization": f"Bearer {self.api_key}",
+			"Accept": "application/json",
+		}
+		self.session = None
+
+	def _ensure_session(self) -> aiohttp.ClientSession:
+		if self.session is None or self.session.closed:
+			self.session = aiohttp.ClientSession(headers=self._headers)
+		return self.session
 
 	def __del__(self):
-		if hasattr(self, "session"):
-			if not self.session.closed:
+		# Best-effort safety net. _pure closes the session on the request loop; scheduling
+		# a fire-and-forget close here is dropped during interpreter shutdown and aiohttp
+		# then warns about an unclosed session.
+		try:
+			session = getattr(self, "session", None)
+			if session is not None and not session.closed:
 				loop = get_event_loop()
-				if loop.is_running():
-					loop.create_task(self.session.close())
-				else:
-					loop.run_until_complete(self.session.close())
-			del self.session
-		super().__del__()
+				if not loop.is_running() and not loop.is_closed():
+					loop.run_until_complete(session.close())
+		except Exception:
+			pass
+		try:
+			super().__del__()
+		except Exception:
+			pass
 
 	@result_to_dataframe(["retrieved_contents", "retrieved_ids", "retrieve_scores"])
 	def pure(self, previous_result: pd.DataFrame, *args, **kwargs):
@@ -88,21 +103,32 @@ class NvidiaReranker(BasePassageReranker):
 		if not queries:
 			return [], [], []
 
-		tasks = [
-			nvidia_rerank_pure(
-				self.session,
-				self.invoke_url,
-				model,
-				query,
-				document,
-				ids,
-				top_k,
-				truncate=truncate,
-			)
-			for query, document, ids in zip(queries, contents_list, ids_list)
-		]
 		loop = get_event_loop()
-		results = loop.run_until_complete(process_batch(tasks, batch_size=batch))
+
+		async def _run_batch():
+			# Create the session here, on the loop that runs the requests. aiohttp 3.13
+			# resolves the session loop from the running loop and rejects loop=.
+			session = self._ensure_session()
+			tasks = [
+				nvidia_rerank_pure(
+					session,
+					self.invoke_url,
+					model,
+					query,
+					document,
+					ids,
+					top_k,
+					truncate=truncate,
+				)
+				for query, document, ids in zip(queries, contents_list, ids_list)
+			]
+			try:
+				return await process_batch(tasks, batch_size=batch)
+			finally:
+				if not session.closed:
+					await session.close()
+
+		results = loop.run_until_complete(_run_batch())
 		if len(results) != len(queries):
 			raise AssertionError(
 				"NVIDIA rerank returned unexpected number of results. "

{autorag-0.3.26 → autorag-0.3.27}/autorag/nodes/promptmaker/run.py

@@ -146,11 +146,22 @@ def run_prompt_maker_node(
 		for result in results:
 			all_prompts.extend(result["prompts"].tolist())
 
+		# Each module must be scored on its own generations. `list * k`
+		# repeats references to the same MetricInput objects, so writing one
+		# module's generated_texts would overwrite every other module's inputs
+		# (every prompt maker ends up scored on the last module's generations).
+		# Deep-copy one fresh MetricInput set per module instead.
+		metric_inputs_all = [
+			deepcopy(metric_input)
+			for _ in range(len(results))
+			for metric_input in metric_inputs
+		]
+
 		evaluation_result_all = evaluate_one_prompt_maker_node(
 			all_prompts,
 			generator_callables,
 			generator_params,
-			metric_inputs * len(results),
+			metric_inputs_all,
 			general_strategy["metrics"],
 			project_dir,
 			strategy_name=strategies.get("strategy", "mean"),

{autorag-0.3.26 → autorag-0.3.27}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "AutoRAG"
-version = "0.3.26"
+version = "0.3.27"
 authors = [
     { name = "Marker-Inc", email = "vkehfdl1@gmail.com" }
 ]
@@ -39,7 +39,7 @@ dependencies = [
     "click>=8.2.1",  # for cli
     "cohere>=5.18.0",  # for cohere services
     "tokenlog>=0.0.3",  # for token logging
-    "aiohttp>=3.12.13,<3.14",  # aioresponses (dev) incompatible with aiohttp 3.14+ (stream_writer)
+    "aiohttp>=3.14.3,<3.15",  # 3.14.3 clears the aiohttp advisories; tests shim aioresponses 0.7.9
     "voyageai>=0.3.2",  # for voyageai reranker
     "mixedbread-ai>=2.2.6",  # for mixedbread-ai reranker
     "scikit-learn>=1.7.0",
@@ -126,12 +126,12 @@ dev = [
 
 [project.optional-dependencies]
 ko = ["kiwipiepy >= 0.18.0", "konlpy>=0.6.0"]
-parse = ["PyMuPDF>=1.26.1", "pdfminer.six>=20250506", "pdfplumber>=0.11.7", "jq>=1.9.1", "unstructured[pdf]>=0.17.2", "PyPDF2>=3.0.1", "pdf2image>=1.17.0"]
+parse = ["PyMuPDF>=1.26.1", "pdfminer.six>=20250506", "pdfplumber>=0.11.7", "jq>=1.9.1", "unstructured[pdf]>=0.24.0; python_version >= '3.11'", "PyPDF2>=3.0.1", "pdf2image>=1.17.0"]
 ja = ["sudachipy>=0.6.8", "sudachidict_core"]
 gpu = ["torch>=2.7.1", "sentencepiece>=0.2.0", "bert_score>=0.3.13", "peft>=0.15.2", "llmlingua>=0.2.2", "FlagEmbedding>=1.2.11",
     "sentence-transformers>=4.1.0", "transformers>=4.51.3", "llama-index-llms-ollama>=0.6.0", "llama-index-embeddings-huggingface>=0.5.4",
     "llama-index-llms-huggingface>=0.5.0", "onnxruntime>=1.22.0", "xgrammar>=0.1.32,<0.3.0"]
-vllm = ["vllm>=0.11.0,<0.30.0"]
+vllm = ["vllm>=0.11.0,<0.31.0"]
 all = ["AutoRAG[gpu]", "AutoRAG[ko]", "AutoRAG[parse]", "AutoRAG[ja]", "AutoRAG[vllm]"]
 
 [project.scripts]

{autorag-0.3.26 → autorag-0.3.27}/tests/autorag/data/qa/evolve/test_openai_query_evolve.py

@@ -17,7 +17,7 @@ async def mock_gen_gt_response(*args, **kwargs):
 	return SimpleNamespace(output_parsed=Response(evolved_query="mock answer"))
 
 
-client = AsyncOpenAI()
+client = AsyncOpenAI(api_key="mock_openai_api_key")
 
 
 @patch.object(

{autorag-0.3.26 → autorag-0.3.27}/tests/autorag/data/qa/filter/test_dontknow.py

@@ -111,7 +111,7 @@ def test_dontknow_filter_rule_based():
 	mock_openai_response,
 )
 def test_dontknow_filter_openai():
-	client = AsyncOpenAI()
+	client = AsyncOpenAI(api_key="mock_openai_api_key")
 	en_qa = QA(en_qa_df)
 	result_en_qa = en_qa.batch_filter(
 		dontknow_filter_openai, client=client, lang="en"

{autorag-0.3.26 → autorag-0.3.27}/tests/autorag/data/qa/filter/test_passage_dependency.py

@@ -113,7 +113,7 @@ async def mock_llama_index_response(*args, **kwargs) -> ChatResponse:
 	mock_openai_response,
 )
 def test_passage_dependency_filter_openai():
-	client = AsyncOpenAI()
+	client = AsyncOpenAI(api_key="mock_openai_api_key")
 	en_qa = QA(en_qa_df)
 	result_en_qa = en_qa.batch_filter(
 		passage_dependency_filter_openai, client=client, lang="en"

{autorag-0.3.26 → autorag-0.3.27}/tests/autorag/data/qa/generation_gt/test_openai_gen_gt.py

@@ -15,7 +15,7 @@ from tests.autorag.data.qa.generation_gt.base_test_generation_gt import (
 	check_generation_gt,
 )
 
-client = AsyncOpenAI()
+client = AsyncOpenAI(api_key="mock_openai_api_key")
 
 
 async def mock_gen_gt_response(*args, **kwargs):

{autorag-0.3.26 → autorag-0.3.27}/tests/autorag/data/qa/query/test_openai_gen_query.py

@@ -14,7 +14,7 @@ from autorag.data.qa.query.openai_gen_query import (
 from autorag.data.qa.schema import QA
 from tests.autorag.data.qa.query.base_test_query_gen import qa_df, multi_hop_qa_df
 
-client = AsyncOpenAI()
+client = AsyncOpenAI(api_key="mock_openai_api_key")
 
 
 async def mock_gen_gt_response(*args, **kwargs):

autorag-0.3.27/tests/autorag/embedding/test_vllm_embedding.py

@@ -0,0 +1,37 @@
+import sys
+import types
+from types import SimpleNamespace
+
+from autorag.embedding.vllm import VllmEmbedding
+
+
+class FakeLLM:
+	def __init__(self, model, **kwargs):
+		if "task" in kwargs:
+			raise TypeError("task is not a supported vLLM constructor argument")
+		self.model = model
+		self.kwargs = kwargs
+		self.llm_engine = SimpleNamespace()
+
+	def embed(self, inputs):
+		return [SimpleNamespace(outputs=SimpleNamespace(embedding=[0.1, 0.2])) for _ in inputs]
+
+
+def install_fake_vllm(monkeypatch):
+	fake_vllm = types.ModuleType("vllm")
+	setattr(fake_vllm, "LLM", FakeLLM)
+	monkeypatch.setitem(sys.modules, "vllm", fake_vllm)
+
+
+def test_vllm_embedding_uses_supported_constructor_and_preserves_kwargs(monkeypatch):
+	install_fake_vllm(monkeypatch)
+
+	embedding = VllmEmbedding(
+		model_name="fake-model",
+		embed_batch_size=4,
+		vllm_kwargs={"gpu_memory_utilization": 0.5},
+	)
+
+	assert embedding._client.kwargs["gpu_memory_utilization"] == 0.5
+	assert "task" not in embedding._client.kwargs
+	assert embedding._get_text_embedding("hello") == [0.1, 0.2]

{autorag-0.3.26 → autorag-0.3.27}/tests/autorag/nodes/passagereranker/test_jina_reranker.py

@@ -5,6 +5,8 @@ import aiohttp
 import pytest
 from aioresponses import aioresponses
 
+from autorag.utils.util import get_event_loop
+
 import autorag
 from autorag.nodes.passagereranker import JinaReranker
 from autorag.nodes.passagereranker.jina import jina_reranker_pure, JINA_API_URL
@@ -49,6 +51,7 @@ async def test_jina_reranker_pure():
 
         # check if the scores are sorted
         assert score_result[0] >= score_result[1]
+        await session.close()
 
 
 async def mock_jina_reranker_pure(session, query, contents, ids, top_k, **kwargs):
@@ -110,3 +113,27 @@ def test_jina_reranker_node():
         project_dir=project_dir, previous_result=previous_result, top_k=top_k
     )
     base_reranker_node_test(result_df, top_k)
+
+
+def test_jina_session_is_lazy_and_not_loop_pinned(monkeypatch):
+    captured = []
+    real_session = aiohttp.ClientSession
+
+    def spy(*args, **kwargs):
+        captured.append(kwargs)
+        return real_session(*args, **kwargs)
+
+    monkeypatch.setattr(aiohttp, "ClientSession", spy)
+    reranker = JinaReranker(project_dir, "mock_api_key")
+    # A session opened at construction time binds to the constructor loop, which is
+    # not the loop that later runs the requests (aiohttp's loop= kwarg is a no-op).
+    assert captured == []
+
+    async def _open():
+        reranker._ensure_session()
+
+    get_event_loop().run_until_complete(_open())
+    assert len(captured) == 1
+    assert "loop" not in captured[0]
+    if reranker.session is not None and not reranker.session.closed:
+        get_event_loop().run_until_complete(reranker.session.close())

{autorag-0.3.26 → autorag-0.3.27}/tests/autorag/nodes/passagereranker/test_nvidia_reranker.py

@@ -30,7 +30,7 @@ MOCK_RESPONSE = {
 def nvidia_reranker_instance():
 	reranker = NvidiaReranker(project_dir=project_dir, api_key="test")
 	yield reranker
-	if hasattr(reranker, "session") and not reranker.session.closed:
+	if getattr(reranker, "session", None) is not None and not reranker.session.closed:
 		loop = get_event_loop()
 		if loop.is_running():
 			loop.create_task(reranker.session.close())
@@ -140,3 +140,27 @@ def test_nvidia_reranker_node():
             api_key="test",
         )
         base_reranker_node_test(result_df, top_k)
+
+
+def test_nvidia_session_is_lazy_and_not_loop_pinned(monkeypatch):
+    captured = []
+    real_session = aiohttp.ClientSession
+
+    def spy(*args, **kwargs):
+        captured.append(kwargs)
+        return real_session(*args, **kwargs)
+
+    monkeypatch.setattr(aiohttp, "ClientSession", spy)
+    reranker = NvidiaReranker(project_dir=project_dir, api_key="test")
+    # A session opened at construction time binds to the constructor loop, which is
+    # not the loop that later runs the requests (aiohttp's loop= kwarg is a no-op).
+    assert captured == []
+
+    async def _open():
+        reranker._ensure_session()
+
+    get_event_loop().run_until_complete(_open())
+    assert len(captured) == 1
+    assert "loop" not in captured[0]
+    if reranker.session is not None and not reranker.session.closed:
+        get_event_loop().run_until_complete(reranker.session.close())

{autorag-0.3.26 → autorag-0.3.27}/tests/autorag/nodes/promptmaker/test_prompt_maker_run.py

@@ -185,6 +185,74 @@ def test_run_prompt_maker_node(node_line_dir):
     assert os.path.exists(os.path.join(node_line_dir, "prompt_maker", "1.parquet"))
 
 
+def test_run_prompt_maker_node_metric_inputs_not_aliased(node_line_dir):
+    """Regression test for the MetricInput aliasing bug (#1707).
+
+    `metric_inputs * len(results)` used to repeat references to the same N
+    objects, so `evaluate_generation` overwrote every module's
+    `generated_texts` with the last module's generations. Each module's block
+    passed to `evaluate_one_prompt_maker_node` must now be distinct objects
+    carrying the same ground truth.
+    """
+    generator_models["mock"] = MockLLM
+    modules = [Fstring, ChatFstring]
+    params = [
+        {
+            "prompt": "Tell me something about the question: {query} \n\n {retrieved_contents}"
+        },
+        {
+            "prompt": [
+                {"role": "system", "content": "You are a helpful assistant."},
+                {
+                    "role": "user",
+                    "content": "Question: {query} \n Something to read: {retrieved_contents} \n What's your answer?",
+                },
+            ]
+        },
+    ]
+    strategies = {
+        "metrics": metrics,
+        "generator_modules": [
+            {
+                "module_type": "llama_index_llm",
+                "llm": "mock",
+                "model": "gpt-3.5-turbo",
+            }
+        ],
+    }
+    with patch(
+        "autorag.nodes.promptmaker.run.evaluate_one_prompt_maker_node"
+    ) as mock_evaluate:
+        mock_evaluate.return_value = pd.DataFrame(
+            {
+                "generated_texts": [
+                    "answer one",
+                    "answer two",
+                    "answer three",
+                    "answer four",
+                ],
+                "bleu": [0.1, 0.2, 0.3, 0.4],
+                "rouge": [0.4, 0.3, 0.2, 0.1],
+            }
+        )
+        run_prompt_maker_node(
+            modules, params, previous_result, node_line_dir, strategies
+        )
+
+    passed_metric_inputs = mock_evaluate.call_args[0][3]
+    half = len(passed_metric_inputs) // 2
+    assert len(passed_metric_inputs) == 2 * len(previous_result)
+    assert all(
+        passed_metric_inputs[i] is not passed_metric_inputs[i + half]
+        for i in range(half)
+    )
+    assert all(isinstance(mi, MetricInput) for mi in passed_metric_inputs)
+    assert (
+        passed_metric_inputs[0].generation_gt
+        == passed_metric_inputs[half].generation_gt
+    )
+
+
 async def acomplete_qa_creation(*args, **kwargs):
     return CompletionResponse(text="This is the test answer.")
 

autorag-0.3.27/tests/autorag/test_optional_extras.py

@@ -0,0 +1,84 @@
+import re
+from pathlib import Path
+
+LEGACY_ROOT = Path(__file__).resolve().parents[2]
+REPO_ROOT = Path(__file__).resolve().parents[3]
+
+
+def _extra_items(name: str) -> list[str]:
+	text = (LEGACY_ROOT / "pyproject.toml").read_text()
+	match = re.search(rf"^{name} = \[", text, re.M)
+	assert match is not None, f"missing optional extra {name}"
+	index = match.end()
+	depth = 1
+	cursor = index
+	while cursor < len(text) and depth:
+		char = text[cursor]
+		if char == "[":
+			depth += 1
+		elif char == "]":
+			depth -= 1
+		cursor += 1
+	body = text[index : cursor - 1]
+	return [
+		item.strip().strip("\"'")
+		for item in body.split(",")
+		if item.strip()
+	]
+
+
+def test_gpu_extra_does_not_include_vllm():
+	items = _extra_items("gpu")
+	assert not any(item.startswith("vllm") for item in items)
+
+
+def test_vllm_is_its_own_optional_extra():
+	assert any(item.startswith("vllm") for item in _extra_items("vllm"))
+	assert any("vllm" in item for item in _extra_items("all"))
+
+
+def test_legacy_unit_test_job_skips_vllm_install_and_bounds_runtime():
+	text = (REPO_ROOT / ".github/workflows/test.yml").read_text()
+	assert re.search(r"timeout-minutes:\s*[1-9]\d*", text)
+	sync_lines = [
+		line.strip() for line in text.splitlines() if "uv sync" in line
+	]
+	assert sync_lines, "legacy-test must install with uv sync"
+	assert all("--all-extras" not in line for line in sync_lines)
+	assert all("--extra vllm" not in line for line in sync_lines)
+	assert any("--extra gpu" in line for line in sync_lines)
+
+
+def test_legacy_unit_test_job_bounds_the_pytest_step():
+	# A stalled test reached 98% and then sat silent for 35 minutes until the
+	# 40-minute job wall killed it. The job bound alone cannot surface that:
+	# the step must carry its own bound so a stall fails fast and loudly.
+	text = (REPO_ROOT / ".github/workflows/test.yml").read_text()
+	steps = text.split("- name:")
+	pytest_step = next((step for step in steps if step.lstrip().startswith("Run AutoRAG tests")), None)
+	assert pytest_step is not None, "legacy-test must run pytest"
+	assert "pytest" in pytest_step, "Run AutoRAG tests must invoke pytest"
+	assert re.search(r"timeout-minutes:\s*[1-9]\d*", pytest_step), "pytest step must bound its runtime"
+	assert "faulthandler_timeout=300" in pytest_step, "pytest step must dump stalled-test tracebacks"
+
+
+def _extra_vllm_specifier() -> str:
+	# _extra_items() splits on commas, so read the specifier straight off the line.
+	match = re.search(r'^vllm = \["vllm([^"]*)"\]', (LEGACY_ROOT / "pyproject.toml").read_text(), re.M)
+	assert match is not None, "missing optional extra vllm"
+	return match.group(1)
+
+
+def _locked_vllm_specifier() -> str:
+	for line in (LEGACY_ROOT / "uv.lock").read_text().splitlines():
+		if 'name = "vllm"' in line and "extra == 'vllm'" in line:
+			match = re.search(r'specifier = "([^"]*)"', line)
+			assert match is not None, f"vllm lock entry has no specifier: {line.strip()}"
+			return match.group(1)
+	raise AssertionError("legacy/uv.lock has no vllm entry for the 'vllm' extra")
+
+
+def test_vllm_extra_specifier_matches_the_lock():
+	# Dependabot rewrites the pyproject extra but never uv.lock, so the specifier
+	# recorded in the lock drifts and CI silently re-locks. Keep them equal.
+	assert _locked_vllm_specifier() == _extra_vllm_specifier()

autorag-0.3.27/tests/conftest.py

@@ -0,0 +1,51 @@
+import os
+from types import SimpleNamespace
+
+import aiohttp.client_reqrep as client_reqrep
+
+
+def pytest_sessionstart(session):
+    os.environ["BM25"] = "bm25"
+    _allow_aioresponses_without_stream_writer()
+
+
+def _allow_aioresponses_without_stream_writer():
+    """Let aioresponses 0.7.9 construct aiohttp 3.14 responses.
+
+    ClientResponse now requires stream_writer and reads output_size when the
+    request writer is already finished. Real sessions pass a writer. This
+    default exists only for the mock, which still omits the argument.
+    """
+    original = client_reqrep.ClientResponse.__init__
+
+    def init(
+        self,
+        method,
+        url,
+        *,
+        writer,
+        continue100,
+        timer,
+        request_info,
+        traces,
+        loop,
+        session,
+        stream_writer=None,
+    ):
+        if stream_writer is None:
+            stream_writer = SimpleNamespace(output_size=0)
+        return original(
+            self,
+            method,
+            url,
+            writer=writer,
+            continue100=continue100,
+            timer=timer,
+            request_info=request_info,
+            traces=traces,
+            loop=loop,
+            session=session,
+            stream_writer=stream_writer,
+        )
+
+    client_reqrep.ClientResponse.__init__ = init

{autorag-0.3.26 → autorag-0.3.27}/tests/resources/result_project/resources/chroma/6595d304-5270-4d9f-a267-c191366804ce/data_level0.bin

@@ -1,4 +1,4 @@
 [diffend] Oversized file quarantined before diffing.
-name: autorag-0.3.26/tests/resources/result_project/resources/chroma/6595d304-5270-4d9f-a267-c191366804ce/data_level0.bin
+name: autorag-0.3.27/tests/resources/result_project/resources/chroma/6595d304-5270-4d9f-a267-c191366804ce/data_level0.bin
 size: 62840000 bytes
 sha256: bb76ae0f6ca830a9a048b0cf53962a78c88c5c7fcda63fc846077d3456eb3890