Fix Dockerfile build issue

autorag/data/chunk/__init__.py

@@ -0,0 +1,2 @@
+from .llama_index_chunk import llama_index_chunk
+from .langchain_chunk import langchain_chunk

autorag/data/chunk/base.py

@@ -0,0 +1,128 @@
+import functools
+import logging
+from typing import Tuple, List, Dict, Any
+
+import pandas as pd
+
+from autorag.embedding.base import EmbeddingModel
+from autorag.data import chunk_modules, sentence_splitter_modules
+from autorag.utils import result_to_dataframe
+
+logger = logging.getLogger("AutoRAG")
+
+
+def chunker_node(func):
+	@functools.wraps(func)
+	@result_to_dataframe(["doc_id", "contents", "path", "start_end_idx", "metadata"])
+	def wrapper(
+		parsed_result: pd.DataFrame, chunk_method: str, **kwargs
+	) -> Tuple[
+		List[str], List[str], List[str], List[Tuple[int, int]], List[Dict[str, Any]]
+	]:
+		logger.info(f"Running chunker - {func.__name__} module...")
+
+		# get texts from parsed_result
+		texts = parsed_result["texts"].tolist()
+
+		# get filenames from parsed_result when 'add_file_name' is setting
+		file_name_language = kwargs.pop("add_file_name", None)
+		metadata_list = make_metadata_list(parsed_result)
+
+		# run chunk module
+		if func.__name__ in ["llama_index_chunk", "langchain_chunk"]:
+			chunk_instance = __get_chunk_instance(
+				func.__name__, chunk_method.lower(), **kwargs
+			)
+			result = func(
+				texts=texts,
+				chunker=chunk_instance,
+				file_name_language=file_name_language,
+				metadata_list=metadata_list,
+			)
+			del chunk_instance
+			return result
+		else:
+			raise ValueError(f"Unsupported module_type: {func.__name__}")
+
+	return wrapper
+
+
+def make_metadata_list(parsed_result: pd.DataFrame) -> List[Dict[str, str]]:
+	metadata_list = [{} for _ in range(len(parsed_result["texts"]))]
+
+	def _make_metadata_pure(
+		lst: List[str], key: str, metadata_lst: List[Dict[str, str]]
+	):
+		for value, metadata in zip(lst, metadata_lst):
+			metadata[key] = value
+
+	for column in ["page", "last_modified_datetime", "path"]:
+		if column in parsed_result.columns:
+			_make_metadata_pure(parsed_result[column].tolist(), column, metadata_list)
+	return metadata_list
+
+
+def __get_chunk_instance(module_type: str, chunk_method: str, **kwargs):
+	# Add sentence_splitter to kwargs
+	sentence_available_methods = [
+		"semantic_llama_index",
+		"semanticdoublemerging",
+		"sentencewindow",
+	]
+	if chunk_method in sentence_available_methods:
+		# llama index default sentence_splitter is 'nltk -PunktSentenceTokenizer'
+		if "sentence_splitter" in kwargs.keys():
+			sentence_splitter_str = kwargs.pop("sentence_splitter")
+			sentence_splitter_func = sentence_splitter_modules[sentence_splitter_str]()
+			kwargs.update({"sentence_splitter": sentence_splitter_func})
+
+	def get_embedding_model(_embed_model_str: str, _module_type: str):
+		if _embed_model_str == "openai":
+			if _module_type == "langchain_chunk":
+				_embed_model_str = "openai_langchain"
+		return EmbeddingModel.load(_embed_model_str)()
+
+	# Add embed_model to kwargs
+	embedding_available_methods = ["semantic_llama_index", "semantic_langchain"]
+	if chunk_method in embedding_available_methods:
+		# there is no default embed_model, so we have to get it parameter and add it.
+		if "embed_model" not in kwargs.keys():
+			raise ValueError(f"embed_model is required for {chunk_method} method.")
+		embed_model_str = kwargs.pop("embed_model")
+		embed_model = get_embedding_model(embed_model_str, module_type)
+		if chunk_method == "semantic_llama_index":
+			kwargs.update({"embed_model": embed_model})
+		elif chunk_method == "semantic_langchain":
+			kwargs.update({"embeddings": embed_model})
+
+	return chunk_modules[chunk_method](**kwargs)
+
+
+def add_file_name(
+	file_name_language: str, file_names: List[str], chunk_texts: List[str]
+) -> List[str]:
+	if file_name_language == "en":
+		return list(
+			map(
+				lambda x: f"file_name: {x[1]}\n contents: {x[0]}",
+				zip(chunk_texts, file_names),
+			)
+		)
+	elif file_name_language == "ko":
+		return list(
+			map(
+				lambda x: f"파일 제목: {x[1]}\n 내용: {x[0]}",
+				zip(chunk_texts, file_names),
+			)
+		)
+	elif file_name_language == "ja":
+		return list(
+			map(
+				lambda x: f"ファイル名: {x[1]}\n 内容: {x[0]}",
+				zip(chunk_texts, file_names),
+			)
+		)
+	else:
+		raise ValueError(
+			f"Unsupported file_name_language: {file_name_language}. Choose from 'en' ,'ko' or 'ja."
+		)

autorag/data/chunk/langchain_chunk.py

@@ -0,0 +1,76 @@
+import os
+from itertools import chain
+import uuid
+from typing import Tuple, List, Dict, Any, Optional
+
+from langchain_text_splitters import TextSplitter
+
+from autorag.data.chunk.base import chunker_node, add_file_name
+from autorag.data.utils.util import add_essential_metadata, get_start_end_idx
+
+
+@chunker_node
+def langchain_chunk(
+	texts: List[str],
+	chunker: TextSplitter,
+	file_name_language: Optional[str] = None,
+	metadata_list: Optional[List[Dict[str, str]]] = None,
+) -> Tuple[
+	List[str], List[str], List[str], List[Tuple[int, int]], List[Dict[str, Any]]
+]:
+	"""
+	Chunk texts from the parsed result to use langchain chunk method
+
+	:param texts: The list of texts to chunk from the parsed result
+	:param chunker: A langchain TextSplitter(Chunker) instance.
+	:param file_name_language: The language to use 'add_file_name' feature.
+	    You need to set one of 'English' and 'Korean'
+	    The 'add_file_name' feature is to add a file_name to chunked_contents.
+	    This is used to prevent hallucination by retrieving contents from the wrong document.
+	    Default form of 'English' is "file_name: {file_name}\n contents: {content}"
+	: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
+	"""
+	results = [
+		langchain_chunk_pure(text, chunker, file_name_language, meta)
+		for text, meta in zip(texts, metadata_list)
+	]
+
+	doc_id, contents, path, start_end_idx, metadata = (
+		list(chain.from_iterable(item)) for item in zip(*results)
+	)
+
+	return doc_id, contents, path, start_end_idx, metadata
+
+
+def langchain_chunk_pure(
+	text: str,
+	chunker: TextSplitter,
+	file_name_language: Optional[str] = None,
+	_metadata: Optional[Dict[str, str]] = None,
+):
+	# chunk
+	chunk_results = chunker.create_documents([text], metadatas=[_metadata])
+
+	# make doc_id
+	doc_id = list(str(uuid.uuid4()) for _ in range(len(chunk_results)))
+
+	# make path
+	path_lst = list(map(lambda x: x.metadata.get("path", ""), chunk_results))
+
+	# 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.page_content, chunk_results))
+		start_end_idx = list(map(lambda x: get_start_end_idx(text, x), chunked_texts))
+		contents = add_file_name(file_name_language, chunked_file_names, chunked_texts)
+	else:
+		contents = list(map(lambda node: node.page_content, chunk_results))
+		start_end_idx = list(map(lambda x: get_start_end_idx(text, x), contents))
+
+	# make metadata
+	metadata = list(
+		map(lambda node: add_essential_metadata(node.metadata), chunk_results)
+	)
+
+	return doc_id, contents, path_lst, start_end_idx, metadata

autorag/data/chunk/llama_index_chunk.py

@@ -0,0 +1,96 @@
+import os.path
+from itertools import chain
+from typing import Tuple, List, Dict, Any, Optional
+
+from llama_index.core import Document
+from llama_index.core.node_parser.interface import NodeParser
+
+from autorag.utils.util import process_batch, get_event_loop
+from autorag.data.chunk.base import chunker_node, add_file_name
+from autorag.data.utils.util import (
+	add_essential_metadata_llama_text_node,
+	get_start_end_idx,
+)
+
+
+@chunker_node
+def llama_index_chunk(
+	texts: List[str],
+	chunker: NodeParser,
+	file_name_language: Optional[str] = None,
+	metadata_list: Optional[List[Dict[str, str]]] = None,
+	batch: int = 8,
+) -> Tuple[
+	List[str], List[str], List[str], List[Tuple[int, int]], List[Dict[str, Any]]
+]:
+	"""
+	Chunk texts from the parsed result to use llama index chunk method
+
+	:param texts: The list of texts to chunk from the parsed result
+	:param chunker: A llama index NodeParser(Chunker) instance.
+	:param file_name_language: The language to use 'add_file_name' feature.
+	    You need to set one of 'English' and 'Korean'
+	    The 'add_file_name' feature is to add a file_name to chunked_contents.
+	    This is used to prevent hallucination by retrieving contents from the wrong document.
+	    Default form of 'English' is "file_name: {file_name}\n contents: {content}"
+	:param metadata_list: The list of dict of metadata from the parsed result
+	: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
+	"""
+	tasks = [
+		llama_index_chunk_pure(text, chunker, file_name_language, meta)
+		for text, meta in zip(texts, metadata_list)
+	]
+	loop = get_event_loop()
+	results = loop.run_until_complete(process_batch(tasks, batch))
+
+	doc_id, contents, path, start_end_idx, metadata = (
+		list(chain.from_iterable(item)) for item in zip(*results)
+	)
+
+	return list(doc_id), list(contents), list(path), list(start_end_idx), list(metadata)
+
+
+async def llama_index_chunk_pure(
+	text: str,
+	chunker: NodeParser,
+	file_name_language: Optional[str] = None,
+	_metadata: Optional[Dict[str, str]] = None,
+):
+	# set document
+	document = [Document(text=text, metadata=_metadata)]
+
+	# chunk document
+	chunk_results = await chunker.aget_nodes_from_documents(documents=document)
+
+	# make doc_id
+	doc_id = list(map(lambda node: node.node_id, chunk_results))
+
+	# make path
+	path_lst = list(map(lambda x: x.metadata.get("path", ""), chunk_results))
+
+	# 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))
+		start_end_idx = list(
+			map(
+				lambda x: get_start_end_idx(text, x),
+				chunked_texts,
+			)
+		)
+		contents = add_file_name(file_name_language, chunked_file_names, chunked_texts)
+	else:
+		contents = list(map(lambda x: x.text, chunk_results))
+		start_end_idx = list(map(lambda x: get_start_end_idx(text, x), contents))
+
+	metadata = list(
+		map(
+			lambda node: add_essential_metadata_llama_text_node(
+				node.metadata, node.relationships
+			),
+			chunk_results,
+		)
+	)
+
+	return doc_id, contents, path_lst, start_end_idx, metadata

autorag/data/chunk/run.py

@@ -0,0 +1,38 @@
+import os
+from typing import Callable, List, Dict
+import pandas as pd
+
+from autorag.strategy import measure_speed
+
+
+def run_chunker(
+	modules: List[Callable],
+	module_params: List[Dict],
+	parsed_result: pd.DataFrame,
+	project_dir: str,
+):
+	results, execution_times = zip(
+		*map(
+			lambda x: measure_speed(x[0], parsed_result=parsed_result, **x[1]),
+			zip(modules, module_params),
+		)
+	)
+	average_times = list(map(lambda x: x / len(results[0]), execution_times))
+
+	# save results to parquet files
+	filepaths = list(
+		map(lambda x: os.path.join(project_dir, f"{x}.parquet"), range(len(modules)))
+	)
+	list(map(lambda x: x[0].to_parquet(x[1], index=False), zip(results, filepaths)))
+	filenames = list(map(lambda x: os.path.basename(x), filepaths))
+
+	summary_df = pd.DataFrame(
+		{
+			"filename": filenames,
+			"module_name": list(map(lambda module: module.__name__, modules)),
+			"module_params": module_params,
+			"execution_time": average_times,
+		}
+	)
+	summary_df.to_csv(os.path.join(project_dir, "summary.csv"), index=False)
+	return summary_df