AutoRAG 0.3.25__tar.gz → 0.3.27__tar.gz

autorag-0.3.27/AGENTS.md

@@ -0,0 +1,14 @@
+# Legacy AutoRAG contribution guide
+
+## Type checking
+
+`ty` is required for legacy product code only. Do not run `ty` against
+`legacy/tests`, and do not add test files to the `ty` command or CI scope.
+
+Tests are validated with pytest and ruff. Test doubles often intentionally
+implement only the runtime surface needed by a test and are therefore outside
+the legacy product-code typing contract.
+
+The current clean product scope is documented in
+`legacy/docs/source/ty.md`. Extend that scope with product modules as they are
+made clean; keep tests excluded.

{autorag-0.3.25 → autorag-0.3.27}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: AutoRAG
-Version: 0.3.25
+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
@@ -229,10 +229,10 @@ Requires-Dist: evaluate>=0.4.4
 Requires-Dist: fastapi>=0.115.13
 Requires-Dist: fastparquet>=2024.11.0
 Requires-Dist: gradio>=6.15.0
-Requires-Dist: grpcio-health-checking<1.84.0,>=1.66.2
-Requires-Dist: grpcio-status<1.84.0,>=1.66.2
-Requires-Dist: grpcio-tools<1.84.0,>=1.66.2
-Requires-Dist: grpcio<1.84.0,>=1.66.2
+Requires-Dist: grpcio-health-checking<1.85.0,>=1.66.2
+Requires-Dist: grpcio-status<1.85.0,>=1.66.2
+Requires-Dist: grpcio-tools<1.85.0,>=1.66.2
+Requires-Dist: grpcio<1.85.0,>=1.66.2
 Requires-Dist: ipykernel>=6.29.5
 Requires-Dist: ipywidgets-bokeh>=1.6.0
 Requires-Dist: ipywidgets>=8.1.7
@@ -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<0.28.0,>=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'
@@ -316,7 +316,6 @@ Requires-Dist: sentence-transformers>=4.1.0; extra == 'gpu'
 Requires-Dist: sentencepiece>=0.2.0; extra == 'gpu'
 Requires-Dist: torch>=2.7.1; extra == 'gpu'
 Requires-Dist: transformers>=4.51.3; extra == 'gpu'
-Requires-Dist: vllm<0.28.0,>=0.11.0; extra == 'gpu'
 Requires-Dist: xgrammar<0.3.0,>=0.1.32; extra == 'gpu'
 Provides-Extra: ja
 Requires-Dist: sudachidict-core; extra == 'ja'
@@ -331,7 +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<0.31.0,>=0.11.0; extra == 'vllm'
 Description-Content-Type: text/markdown
 
 # AutoRAG (Legacy)

{autorag-0.3.25 → autorag-0.3.27}/autorag/data/chunk/langchain_chunk.py

@@ -31,6 +31,7 @@ def langchain_chunk(
 	:param metadata_list: The list of dict of metadata from the parsed result
 	:return: tuple of lists containing the chunked doc_id, contents, path, start_idx, end_idx and metadata
 	"""
+	metadata_list = metadata_list or [{} for _ in texts]
 	results = [
 		langchain_chunk_pure(text, chunker, file_name_language, meta)
 		for text, meta in zip(texts, metadata_list)
@@ -50,7 +51,9 @@ def langchain_chunk_pure(
 	_metadata: Optional[Dict[str, str]] = None,
 ):
 	# chunk
-	chunk_results = chunker.create_documents([text], metadatas=[_metadata])
+	chunk_results = chunker.create_documents(
+		[text], metadatas=[_metadata or {}]
+	)
 
 	# make doc_id
 	doc_id = list(str(uuid.uuid4()) for _ in range(len(chunk_results)))

{autorag-0.3.25 → autorag-0.3.27}/autorag/data/chunk/llama_index_chunk.py

@@ -37,6 +37,7 @@ def llama_index_chunk(
 	:param batch: The batch size for chunk texts. Default is 8
 	:return: tuple of lists containing the chunked doc_id, contents, path, start_idx, end_idx and metadata
 	"""
+	metadata_list = metadata_list or [{} for _ in texts]
 	tasks = [
 		llama_index_chunk_pure(text, chunker, file_name_language, meta)
 		for text, meta in zip(texts, metadata_list)
@@ -58,7 +59,7 @@ async def llama_index_chunk_pure(
 	_metadata: Optional[Dict[str, str]] = None,
 ):
 	# set document
-	document = [Document(text=text, metadata=_metadata)]
+	document = [Document(text=text, metadata=_metadata or {})]
 
 	# chunk document
 	chunk_results = await chunker.aget_nodes_from_documents(documents=document)
@@ -72,7 +73,7 @@ async def llama_index_chunk_pure(
 	# make contents and start_end_idx
 	if file_name_language:
 		chunked_file_names = list(map(lambda x: os.path.basename(x), path_lst))
-		chunked_texts = list(map(lambda x: x.text, chunk_results))
+		chunked_texts = list(map(lambda x: x.get_content(), chunk_results))
 		start_end_idx = list(
 			map(
 				lambda x: get_start_end_idx(text, x),
@@ -81,7 +82,7 @@ async def llama_index_chunk_pure(
 		)
 		contents = add_file_name(file_name_language, chunked_file_names, chunked_texts)
 	else:
-		contents = list(map(lambda x: x.text, chunk_results))
+		contents = list(map(lambda x: x.get_content(), chunk_results))
 		start_end_idx = list(map(lambda x: get_start_end_idx(text, x), contents))
 
 	metadata = list(

{autorag-0.3.25 → autorag-0.3.27}/autorag/data/parse/base.py

@@ -48,6 +48,8 @@ def parser_node(func):
 		)
 
 		if func.__name__ == "langchain_parse":
+			if parse_method is None:
+				raise ValueError("parse_method is required for langchain_parse")
 			parse_method = parse_method.lower()
 			if parse_method == "directory":
 				path_split_list = data_path_glob.split("/")

{autorag-0.3.25 → autorag-0.3.27}/autorag/data/parse/clova.py

@@ -2,7 +2,7 @@ import base64
 import itertools
 import json
 import os
-from typing import List, Optional, Tuple
+from typing import List, Optional, Tuple, TypedDict
 
 import aiohttp
 import fitz  # PyMuPDF
@@ -11,6 +11,12 @@ from autorag.data.parse.base import parser_node
 from autorag.utils.util import process_batch, get_event_loop
 
 
+class TableCell(TypedDict):
+	text: str
+	rowSpan: int
+	colSpan: int
+
+
 @parser_node
 def clova_ocr(
 	data_path_list: List[str],
@@ -152,7 +158,9 @@ def json_to_html_table(json_data):
 	max_row = max(cell["rowIndex"] + cell["rowSpan"] for cell in json_data)
 	max_col = max(cell["columnIndex"] + cell["columnSpan"] for cell in json_data)
 	# Create a 2D array to keep track of merged cells
-	table = [["" for _ in range(max_col)] for _ in range(max_row)]
+	table: list[list[TableCell | None | str]] = [
+		["" for _ in range(max_col)] for _ in range(max_row)
+	]
 	# Fill the table with cell data
 	for cell in json_data:
 		row = cell["rowIndex"]
@@ -182,6 +190,8 @@ def json_to_html_table(json_data):
 			if cell == "":
 				html += "    <td></td>\n"
 			else:
+				if not isinstance(cell, dict):
+					raise TypeError(f"Unexpected table cell value: {cell!r}")
 				row_span_attr = (
 					f' rowspan="{cell["rowSpan"]}"' if cell["rowSpan"] > 1 else ""
 				)

{autorag-0.3.25 → 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-0.3.25 → 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-0.3.25 → 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-0.3.25 → autorag-0.3.27}/autorag/nodes/passagereranker/tart/modeling_enc_t5.py

@@ -1,14 +1,34 @@
 # Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
 
 import copy
+from math import ceil
 
 from transformers.modeling_outputs import SequenceClassifierOutput
 from transformers.models.t5.modeling_t5 import T5Config, T5PreTrainedModel, T5Stack
-from transformers.utils.model_parallel_utils import assert_device_map, get_device_map
 
 from autorag.utils.util import empty_cuda_cache
 
 
+def assert_device_map(device_map, num_blocks):
+	blocks = list(range(num_blocks))
+	device_map_blocks = [block for blocks in device_map.values() for block in blocks]
+	duplicate_blocks = [block for block in device_map_blocks if device_map_blocks.count(block) > 1]
+	missing_blocks = [block for block in blocks if block not in device_map_blocks]
+	extra_blocks = [block for block in device_map_blocks if block not in blocks]
+	if duplicate_blocks or missing_blocks or extra_blocks:
+		raise ValueError("device_map must contain each encoder block exactly once")
+
+
+def get_device_map(n_layers, devices):
+	layers = list(range(n_layers))
+	layers_per_device = int(ceil(n_layers / len(devices)))
+	layers_list = [
+		layers[index : index + layers_per_device]
+		for index in range(0, n_layers, layers_per_device)
+	]
+	return dict(zip(devices, layers_list))
+
+
 class EncT5ForSequenceClassification(T5PreTrainedModel):
 	_keys_to_ignore_on_load_missing = [
 		r"encoder\.embed_tokens\.weight",
@@ -28,7 +48,8 @@ class EncT5ForSequenceClassification(T5PreTrainedModel):
 		encoder_config = copy.deepcopy(config)
 		encoder_config.use_cache = False
 		encoder_config.is_encoder_decoder = False
-		self.encoder = T5Stack(encoder_config, self.shared)
+		self.encoder = T5Stack(encoder_config)
+		self.encoder.embed_tokens = self.shared
 
 		self.dropout = nn.Dropout(dropout)
 		self.classifier = nn.Linear(config.hidden_size, config.num_labels)

{autorag-0.3.25 → autorag-0.3.27}/autorag/nodes/passagereranker/tart/tart.py

@@ -8,6 +8,7 @@ from autorag.nodes.passagereranker.tart.modeling_enc_t5 import (
 	EncT5ForSequenceClassification,
 )
 from autorag.nodes.passagereranker.tart.tokenization_enc_t5 import EncT5Tokenizer
+from transformers import T5Config
 from autorag.utils.util import (
 	make_batch,
 	sort_by_scores,
@@ -28,7 +29,15 @@ class Tart(BasePassageReranker):
 				"torch is not installed. Please install torch first to use TART reranker."
 			)
 		model_name = "facebook/tart-full-flan-t5-xl"
-		self.model = EncT5ForSequenceClassification.from_pretrained(model_name)
+		config_dict, _ = T5Config.get_config_dict(model_name)
+		config_dict["id2label"] = {
+			int(key): str(value) for key, value in config_dict["id2label"].items()
+		}
+		config = T5Config.from_dict(config_dict)
+		self.model = EncT5ForSequenceClassification.from_pretrained(
+			model_name,
+			config=config,
+		)
 		self.tokenizer = EncT5Tokenizer.from_pretrained(model_name)
 		self.device = "cuda" if torch.cuda.is_available() else "cpu"
 		self.model = self.model.to(self.device)

{autorag-0.3.25 → 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-0.3.25 → autorag-0.3.27}/autorag/nodes/util.py

@@ -4,6 +4,7 @@ from autorag.support import get_support_modules
 
 
 def make_generator_callable_param(generator_dict: Optional[Dict]):
+	generator_dict = generator_dict or {}
 	if "generator_module_type" not in generator_dict.keys():
 		generator_dict = {
 			"generator_module_type": "llama_index_llm",

{autorag-0.3.25 → autorag-0.3.27}/docs/source/index.rst

@@ -83,6 +83,7 @@ Also, feel free to ask your question at our `github issue <https://github.com/Ma
    azure_openai.md
    migration.md
    test_your_rag.md
+   ty.md
 
 .. toctree::
     :maxdepth: 1

autorag-0.3.27/docs/source/ty.md

@@ -0,0 +1,30 @@
+# Legacy AutoRAG type checking
+
+Legacy AutoRAG uses `ty` for static type checking. The checks are intentionally
+introduced in small product-code scopes so optional integrations and third-party
+stub limitations do not hide errors in the core modules.
+
+`legacy/tests` is intentionally excluded from `ty`. Tests and test doubles are
+validated with pytest and ruff instead; they are not part of the product-code
+typing contract.
+
+## Reproducible commands
+
+From `legacy/`:
+
+```bash
+uv sync --extra ko --extra parse --extra ja
+uv run --locked ty check autorag/data/chunk autorag/data/parse/base.py \
+  autorag/data/parse/clova.py autorag/nodes/util.py
+uv run --locked ruff check autorag tests
+uv run --locked pytest -q tests/autorag/data/chunk \
+  tests/autorag/data/parse/test_langchain_parse.py \
+  tests/autorag/data/parse/test_clova.py
+```
+
+The focused product-code `ty` scope above is clean and runs in CI. Do not
+replace it with `ty check autorag tests`: tests are intentionally excluded.
+The complete product-code command, `uv run --locked ty check autorag`, remains
+broader than this initial scope because it includes optional integrations and
+third-party library APIs that are being migrated in subsequent issue-sized
+changes.

{autorag-0.3.25 → autorag-0.3.27}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "AutoRAG"
-version = "0.3.25"
+version = "0.3.27"
 authors = [
     { name = "Marker-Inc", email = "vkehfdl1@gmail.com" }
 ]
@@ -39,17 +39,17 @@ 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",
     "emoji>=2.14.1",
     "fastapi>=0.115.13",
     "banks>=2.4.2",
-    "grpcio>=1.66.2,<1.84.0",
-    "grpcio-health-checking>=1.66.2,<1.84.0",
-    "grpcio-status>=1.66.2,<1.84.0",
-    "grpcio-tools>=1.66.2,<1.84.0",
+    "grpcio>=1.66.2,<1.85.0",
+    "grpcio-health-checking>=1.66.2,<1.85.0",
+    "grpcio-status>=1.66.2,<1.85.0",
+    "grpcio-tools>=1.66.2,<1.85.0",
     "datasets>=3.5.1",
     # Vector DB
     "pymilvus>=2.6.0b0,<4",  # for using milvus vectordb (3.x drops milvus-lite)
@@ -120,20 +120,19 @@ dev = [
     "pytest-mock",
     "aioresponses>=0.7.9",
     "asyncstdlib",
-    "vllm-mock>=0.0.3; sys_platform == 'linux'",
     "ty",
     "deptry",
 ]
 
 [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>=0.11.0,<0.28.0"]
-all = ["AutoRAG[gpu]", "AutoRAG[ko]", "AutoRAG[parse]", "AutoRAG[ja]"]
+    "llama-index-llms-huggingface>=0.5.0", "onnxruntime>=1.22.0", "xgrammar>=0.1.32,<0.3.0"]
+vllm = ["vllm>=0.11.0,<0.31.0"]
+all = ["AutoRAG[gpu]", "AutoRAG[ko]", "AutoRAG[parse]", "AutoRAG[ja]", "AutoRAG[vllm]"]
 
 [project.scripts]
 autorag = "autorag.cli:cli"

{autorag-0.3.25 → autorag-0.3.27}/tests/autorag/data/parse/test_langchain_parse.py

@@ -1,3 +1,5 @@
+import pytest
+
 from autorag.data.parse import langchain_parse
 
 from tests.autorag.data.parse.test_parse_base import (
@@ -30,6 +32,16 @@ def test_langchain_parse_single_pdf():
     assert pages == [-1]
 
 
+def test_langchain_parse_requires_parse_method():
+    with pytest.raises(ValueError, match="parse_method is required"):
+        langchain_parse(
+            single_pdf_path_list[0].replace(
+                single_pdf_path_list[0].split("/")[-1], "*"
+            ),
+            file_type="pdf",
+        )
+
+
 def test_langchain_parse_single_pdf_node():
     result_df = langchain_parse(
         korean_text_glob, file_type="pdf", parse_method="pdfminer"

{autorag-0.3.25 → 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-0.3.25 → 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-0.3.25 → 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-0.3.25 → 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-0.3.25 → 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):