AutoRAG 0.3.24__tar.gz → 0.3.26__tar.gz

autorag-0.3.26/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-0.3.24 → autorag-0.3.26}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.4
+Metadata-Version: 2.5
 Name: AutoRAG
-Version: 0.3.24
+Version: 0.3.26
 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>
@@ -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<1.83.0,>=1.66.2
-Requires-Dist: grpcio-status<1.83.0,>=1.66.2
-Requires-Dist: grpcio-tools<1.83.0,>=1.66.2
-Requires-Dist: grpcio<1.83.0,>=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
@@ -251,6 +251,7 @@ Requires-Dist: llama-index-llms-openai>=0.4.7
 Requires-Dist: llama-index-readers-file>=0.4.9
 Requires-Dist: llama-index-retrievers-bm25>=0.5.2
 Requires-Dist: llama-index>=0.12.42
+Requires-Dist: llama-parse>=0.6.21
 Requires-Dist: mixedbread-ai>=2.2.6
 Requires-Dist: numpy>=2.2.6
 Requires-Dist: openai>=2.8.0
@@ -259,7 +260,7 @@ Requires-Dist: panel>=1.7.1
 Requires-Dist: pinecone[grpc]
 Requires-Dist: pyarrow>=23.0.1
 Requires-Dist: pydantic>=2.9.2
-Requires-Dist: pymilvus<3,>=2.6.0b0
+Requires-Dist: pymilvus<4,>=2.6.0b0
 Requires-Dist: pyngrok>=7.2.11
 Requires-Dist: pyyaml>=6.0.2
 Requires-Dist: qdrant-client>=1.12.1
@@ -300,7 +301,8 @@ 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>=0.11.0; extra == 'all'
+Requires-Dist: vllm<0.30.0,>=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'
 Requires-Dist: flagembedding>=1.2.11; extra == 'gpu'
@@ -314,7 +316,7 @@ 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>=0.11.0; extra == 'gpu'
+Requires-Dist: xgrammar<0.3.0,>=0.1.32; extra == 'gpu'
 Provides-Extra: ja
 Requires-Dist: sudachidict-core; extra == 'ja'
 Requires-Dist: sudachipy>=0.6.8; extra == 'ja'
@@ -329,6 +331,8 @@ 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'
+Provides-Extra: vllm
+Requires-Dist: vllm<0.30.0,>=0.11.0; extra == 'vllm'
 Description-Content-Type: text/markdown
 
 # AutoRAG (Legacy)

{autorag-0.3.24 → autorag-0.3.26}/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-0.3.24 → autorag-0.3.26}/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-0.3.24 → autorag-0.3.26}/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-0.3.24 → autorag-0.3.26}/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-0.3.24 → autorag-0.3.26}/autorag/evaluator.py

@@ -151,6 +151,9 @@ class Evaluator:
 		)
 		yaml_dict = load_yaml_config(yaml_path)
 		vectordb = yaml_dict.get("vectordb", [])
+		for vectordb_config in vectordb:
+			if "path" in vectordb_config:
+				vectordb_config["path"] = os.path.abspath(vectordb_config["path"])
 
 		vectordb_config_path = os.path.join(
 			self.project_dir, "resources", "vectordb.yaml"

autorag-0.3.26/autorag/nodes/generator/chat.py

@@ -0,0 +1,65 @@
+"""Helpers for truncating chat prompts without changing their message roles."""
+
+from typing import Callable, Dict, List, Sequence, TypeVar, Union
+
+
+Token = TypeVar("Token")
+Encode = Callable[[str], Sequence[Token]]
+Decode = Callable[[Sequence[Token]], str]
+Prompt = Union[str, List[Dict]]
+
+
+def messages_to_string(messages: List[Dict[str, str]]) -> str:
+	"""Render messages only for token counting, never for model input."""
+	formatted_parts = [
+		f"<|im_start|>{message['role']}\n{message['content']}<|im_end|>"
+		for message in messages
+	]
+	formatted_parts.append("<|im_start|>assistant")
+	return "\n".join(formatted_parts)
+
+
+def count_message_tokens(messages: List[Dict], encode: Encode) -> int:
+	"""Count the tokenized representation of a chat prompt."""
+	return len(encode(messages_to_string(messages)))
+
+
+def truncate_messages_by_token(
+	messages: List[Dict], encode: Encode, decode: Decode, max_token_size: int
+) -> List[Dict]:
+	"""Fit a chat prompt in a token budget while preserving message structure."""
+	truncated_messages = [dict(message) for message in messages]
+	if count_message_tokens(truncated_messages, encode) <= max_token_size:
+		return truncated_messages
+
+	result = [{**message, "content": ""} for message in truncated_messages]
+	if count_message_tokens(result, encode) > max_token_size:
+		return result
+
+	for index, message in enumerate(truncated_messages):
+		content_tokens = encode(message.get("content", ""))
+		low, high = 0, len(content_tokens)
+		best_content = ""
+		while low <= high:
+			mid = (low + high) // 2
+			candidate = [dict(item) for item in result]
+			candidate[index]["content"] = decode(content_tokens[:mid])
+			if count_message_tokens(candidate, encode) <= max_token_size:
+				best_content = candidate[index]["content"]
+				low = mid + 1
+			else:
+				high = mid - 1
+
+		result[index]["content"] = best_content
+
+	return result
+
+
+def truncate_prompt_by_token(
+	prompt: Prompt, encode: Encode, decode: Decode, max_token_size: int
+) -> Prompt:
+	"""Truncate text or chat prompts while retaining a chat prompt's roles."""
+	if isinstance(prompt, list):
+		return truncate_messages_by_token(prompt, encode, decode, max_token_size)
+	tokens = encode(prompt)
+	return decode(tokens[:max_token_size])

{autorag-0.3.24 → autorag-0.3.26}/autorag/nodes/generator/llama_index_llm.py

@@ -55,7 +55,12 @@ class LlamaIndexLLM(BaseGenerator):
 						"`model` or `model_name` parameter must be provided for using huggingfacellm."
 					)
 			kwargs["tokenizer_name"] = kwargs["model_name"]
-		self.llm_instance: BaseLLM = llm_class(**pop_params(llm_class.__init__, kwargs))
+		llm_params = pop_params(llm_class.__init__, kwargs)
+		model_fields = getattr(llm_class, "model_fields", {})
+		for key in list(kwargs.keys()):
+			if key in model_fields:
+				llm_params[key] = kwargs.pop(key)
+		self.llm_instance: BaseLLM = llm_class(**llm_params)
 
 	def __del__(self):
 		super().__del__()

{autorag-0.3.24 → autorag-0.3.26}/autorag/nodes/generator/minimax_llm.py

@@ -8,6 +8,11 @@ from openai import AsyncOpenAI
 from tiktoken import Encoding
 
 from autorag.nodes.generator.base import BaseGenerator
+from autorag.nodes.generator.chat import (
+	count_message_tokens as count_message_tokens_with_encoder,
+	messages_to_string as render_messages_to_string,
+	truncate_prompt_by_token,
+)
 from autorag.utils.util import (
 	get_event_loop,
 	process_batch,
@@ -186,63 +191,34 @@ class MiniMaxLLM(BaseGenerator):
 def truncate_by_token(
 	prompt: Union[str, List[Dict]], tokenizer: Encoding, max_token_size: int
 ) -> Union[str, List[Dict]]:
-	if isinstance(prompt, list):
-		return truncate_messages_by_token(prompt, tokenizer, max_token_size)
-	tokens = tokenizer.encode(prompt, allowed_special="all")
-	return tokenizer.decode(tokens[:max_token_size])
+	return truncate_prompt_by_token(
+		prompt,
+		lambda text: tokenizer.encode(text, allowed_special="all"),
+		tokenizer.decode,
+		max_token_size,
+	)
 
 
 def truncate_messages_by_token(
 	messages: List[Dict], tokenizer: Encoding, max_token_size: int
 ) -> List[Dict]:
-	truncated_messages = [dict(message) for message in messages]
-	if count_message_tokens(truncated_messages, tokenizer) <= max_token_size:
-		return truncated_messages
-
-	result: List[Dict] = []
-	for message in truncated_messages:
-		candidate = result + [message]
-		if count_message_tokens(candidate, tokenizer) <= max_token_size:
-			result.append(message)
-			continue
-
-		content_tokens = tokenizer.encode(
-			message.get("content", ""), allowed_special="all"
-		)
-		low, high = 0, len(content_tokens)
-		best_content = ""
-		while low <= high:
-			mid = (low + high) // 2
-			truncated_message = {
-				**message,
-				"content": tokenizer.decode(content_tokens[:mid]),
-			}
-			candidate = result + [truncated_message]
-			if count_message_tokens(candidate, tokenizer) <= max_token_size:
-				best_content = truncated_message["content"]
-				low = mid + 1
-			else:
-				high = mid - 1
-
-		if best_content or not result:
-			result.append({**message, "content": best_content})
-		break
-
-	return result
+	return truncate_prompt_by_token(
+		messages,
+		lambda text: tokenizer.encode(text, allowed_special="all"),
+		tokenizer.decode,
+		max_token_size,
+	)
 
 
 def count_message_tokens(messages: List[Dict], tokenizer: Encoding) -> int:
-	return len(tokenizer.encode(messages_to_string(messages), allowed_special="all"))
+	return count_message_tokens_with_encoder(
+		messages, lambda text: tokenizer.encode(text, allowed_special="all")
+	)
 
 
 def messages_to_string(messages: List[Dict[str, str]]) -> str:
-	"""Convert chat messages to string format for token counting."""
-	formatted_parts = [
-		f"<|im_start|>{message['role']}\n{message['content']}<|im_end|>"
-		for message in messages
-	]
-	formatted_parts.append("<|im_start|>assistant")
-	return "\n".join(formatted_parts)
+	"""Convert chat messages to a string for token counting compatibility."""
+	return render_messages_to_string(messages)
 
 
 def strip_think_tags(text: str) -> str:

{autorag-0.3.24 → autorag-0.3.26}/autorag/nodes/generator/openai_llm.py

@@ -8,6 +8,7 @@ from openai import AsyncOpenAI
 from tiktoken import Encoding
 
 from autorag.nodes.generator.base import BaseGenerator
+from autorag.nodes.generator.chat import messages_to_string, truncate_prompt_by_token
 from autorag.utils.util import (
 	get_event_loop,
 	process_batch,
@@ -333,16 +334,9 @@ class OpenAILLM(BaseGenerator):
 		if not self.llm.startswith("gpt-5"):
 			raise ValueError("get_result_gpt_5 is only for gpt-5 models.")
 		messages = parse_prompt(prompt)
-		instruction = "\n\n".join(
-			[msg["content"] for msg in messages if msg["role"] == "system"]
-		)
-		user_input = "\n\n".join(
-			[msg["content"] for msg in messages if msg["role"] == "user"]
-		)
 		response = await self.client.responses.create(
 			model=self.llm,
-			instructions=instruction,
-			input=user_input,
+			input=messages,
 			**kwargs,
 		)
 		answer: str = response.output_text
@@ -354,21 +348,17 @@ class OpenAILLM(BaseGenerator):
 def truncate_by_token(
 	prompt: Union[str, List[Dict]], tokenizer: Encoding, max_token_size: int
 ):
-	if isinstance(prompt, list):
-		prompt = tiktoken_messages_to_string(prompt)
-	tokens = tokenizer.encode(prompt, allowed_special="all")
-	return tokenizer.decode(tokens[:max_token_size])
+	return truncate_prompt_by_token(
+		prompt,
+		lambda text: tokenizer.encode(text, allowed_special="all"),
+		tokenizer.decode,
+		max_token_size,
+	)
 
 
 def tiktoken_messages_to_string(messages: List[Dict[str, str]]) -> str:
-	"""Convert chat messages to string format for accurate token counting"""
-	formatted_parts = [
-		f"<|im_start|>{message['role']}\n{message['content']}<|im_end|>"
-		for message in messages
-	]
-	formatted_parts.append("<|im_start|>assistant")
-	full_string = "\n".join(formatted_parts)
-	return full_string
+	"""Convert chat messages to a string for token counting compatibility."""
+	return messages_to_string(messages)
 
 
 def parse_prompt(prompt: Union[str, List[Dict]]) -> List[Dict]:

{autorag-0.3.24 → autorag-0.3.26}/autorag/nodes/generator/vllm_api.py

@@ -7,6 +7,7 @@ import requests
 from asyncio import to_thread
 
 from autorag.nodes.generator.base import BaseGenerator
+from autorag.nodes.generator.chat import truncate_prompt_by_token
 from autorag.utils.util import get_event_loop, process_batch, result_to_dataframe
 
 logger = logging.getLogger("AutoRAG")
@@ -84,15 +85,18 @@ class VllmAPI(BaseGenerator):
 		logprob_result = list(map(lambda x: x[2], results))
 		return answer_result, token_result, logprob_result
 
-	def truncate_by_token(self, prompt: Union[str, List[Dict]]) -> str:
+	def truncate_by_token(
+		self, prompt: Union[str, List[Dict]]
+	) -> Union[str, List[Dict]]:
 		"""
 		Function to truncate prompts to fit within the maximum token limit.
 		"""
-		if not isinstance(prompt, str):
-			content_list = [msg["content"] for msg in prompt]
-			prompt = "\n".join(content_list)
-		tokens = self.encoding_for_model(prompt)["tokens"]  # Simple tokenization
-		return self.decoding_for_model(tokens[: self.max_model_len])["prompt"]
+		return truncate_prompt_by_token(
+			prompt,
+			lambda text: self.encoding_for_model(text)["tokens"],
+			lambda tokens: self.decoding_for_model(list(tokens))["prompt"],
+			self.max_model_len,
+		)
 
 	def call_vllm_api(self, prompt: Union[str, List[Dict]], **kwargs) -> dict:
 		"""

{autorag-0.3.24 → autorag-0.3.26}/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-0.3.24 → autorag-0.3.26}/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-0.3.24 → autorag-0.3.26}/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-0.3.24 → autorag-0.3.26}/autorag/schema/node.py

@@ -22,6 +22,12 @@ class Node:
 	run_node: Callable = field(init=False)
 
 	def __post_init__(self):
+		if self.node_type == "hybrid_retrieval" and isinstance(
+			self.node_params.get("top_k"), list
+		):
+			raise ValueError(
+				"hybrid_retrieval top_k must be an integer, not a list."
+			)
 		self.run_node = get_support_nodes(self.node_type)
 		if self.run_node is None:
 			raise ValueError(f"Node type {self.node_type} is not supported.")

{autorag-0.3.24 → autorag-0.3.26}/autorag/vectordb/qdrant.py

@@ -8,7 +8,7 @@ from qdrant_client.models import (
 	PointIdsList,
 	HasIdCondition,
 	Filter,
-	SearchRequest,
+	QueryRequest,
 )
 
 from typing import List, Tuple, Union
@@ -148,18 +148,18 @@ class Qdrant(BaseVectorStore):
 
 		search_queries = list(
 			map(
-				lambda x: SearchRequest(vector=x, limit=top_k, with_vector=True),
+				lambda x: QueryRequest(query=x, limit=top_k, with_vector=True),
 				query_embeddings,
 			)
 		)
 
-		search_result = self.client.search_batch(
+		search_result = self.client.query_batch_points(
 			collection_name=self.collection_name, requests=search_queries
 		)
 
 		# Extract IDs and distances
-		ids = [[str(hit.id) for hit in result] for result in search_result]
-		scores = [[hit.score for hit in result] for result in search_result]
+		ids = [[str(hit.id) for hit in result.points] for result in search_result]
+		scores = [[hit.score for hit in result.points] for result in search_result]
 
 		return ids, scores
 

{autorag-0.3.24 → autorag-0.3.26}/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-0.3.26/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.