gangyan/langchain-chat/server/knowledge_base/utils.py

import asyncio
import os
import re
from configs import (
    KB_ROOT_PATH,
    CHUNK_SIZE,
    OVERLAP_SIZE,
    ZH_TITLE_ENHANCE,
    logger,
    log_verbose,
    text_splitter_dict,
    LLM_MODELS,
    TEXT_SPLITTER_NAME,
    TEXT_SPLITTER_MAP
)
import importlib
from server.chat.policy_fun_iast import get_llm_model_response_async
from server.knowledge_base import kb_service as tr
from server.knowledge_base.TexkRank import TextRank
from text_splitter import zh_title_enhance as func_zh_title_enhance
import langchain.document_loaders
from langchain.docstore.document import Document
from langchain.text_splitter import TextSplitter
from pathlib import Path
from server.utils import run_in_thread_pool, get_model_worker_config
import json
from typing import List, Union,Dict, Tuple, Generator
import chardet
import time

def get_split_time(f):
    def inner(*arg,**kwarg):
        s_time = time.time()
        res = f(*arg,**kwarg)
        e_time = time.time()
        print('切片耗时：{}秒'.format(e_time - s_time))
        return res
    return inner

def validate_kb_name(knowledge_base_id: str) -> bool:
    # 检查是否包含预期外的字符或路径攻击关键字
    if "../" in knowledge_base_id:
        return False
    return True


def get_kb_path(knowledge_base_name: str):
    return os.path.join(KB_ROOT_PATH, knowledge_base_name)


def get_doc_path(knowledge_base_name: str):
    return os.path.join(get_kb_path(knowledge_base_name), "content")


def get_vs_path(knowledge_base_name: str, vector_name: str):
    return os.path.join(get_kb_path(knowledge_base_name), "vector_store", vector_name)


def get_file_path(knowledge_base_name: str, doc_name: str):
    return os.path.join(get_doc_path(knowledge_base_name), doc_name)


def list_kbs_from_folder():
    return [f for f in os.listdir(KB_ROOT_PATH)
            if os.path.isdir(os.path.join(KB_ROOT_PATH, f))]


def list_files_from_folder(kb_name: str):
    doc_path = get_doc_path(kb_name)
    result = []

    def is_skiped_path(path: str):
        tail = os.path.basename(path).lower()
        for x in ["temp", "tmp", ".", "~$"]:
            if tail.startswith(x):
                return True
        return False

    def process_entry(entry):
        if is_skiped_path(entry.path):
            return

        if entry.is_symlink():
            target_path = os.path.realpath(entry.path)
            with os.scandir(target_path) as target_it:
                for target_entry in target_it:
                    process_entry(target_entry)
        elif entry.is_file():
            file_path = (Path(os.path.relpath(entry.path, doc_path)).as_posix()) # 路径统一为 posix 格式
            result.append(file_path)
        elif entry.is_dir():
            with os.scandir(entry.path) as it:
                for sub_entry in it:
                    process_entry(sub_entry)

    with os.scandir(doc_path) as it:
        for entry in it:
            process_entry(entry)

    return result


LOADER_DICT = {"GCYHTMLLoader": ['.html', '.htm'],
               "GCYWordLoader2": ['.docx', '.doc'],
            #    "GCYWordLoader": ['.docx'], # .doc 解析目前有点问题，暂时关掉
               "MHTMLLoader": ['.mhtml'],
               "TextLoader": ['.md', '.txt'],
               "JSONLoader": [".json"],
               "JSONLinesLoader": [".jsonl"],
               "RapidOCRCSVLoader": [".csv"],
            #    "CSVLoader": [".csv"],
               # "FilteredCSVLoader": [".csv"], 如果使用自定义分割csv
               "PyMuPDFLoader": [".pdf"],
               #"RapidOCRDocLoader": ['.docx', '.doc'],
               "RapidOCRPPTLoader": ['.ppt', '.pptx', ],
               "RapidOCRLoader": ['.png', '.jpg', '.jpeg', '.bmp'],
               "UnstructuredFileLoader": ['.eml', '.msg', '.rst',
                                          '.rtf', '.xml',
                                          '.epub', '.odt','.tsv'],
               "UnstructuredEmailLoader": ['.eml', '.msg'],
               "UnstructuredEPubLoader": ['.epub'],
               "ExcelLoader": ['.xlsx', '.xls', '.xlsd'],
               "NotebookLoader": ['.ipynb'],
               "UnstructuredODTLoader": ['.odt'],
               "PythonLoader": ['.py'],
               "UnstructuredRSTLoader": ['.rst'],
               "UnstructuredRTFLoader": ['.rtf'],
               "SRTLoader": ['.srt'],
               "TomlLoader": ['.toml'],
               "UnstructuredTSVLoader": ['.tsv'],
               "UnstructuredXMLLoader": ['.xml'],
               "UnstructuredPowerPointLoader": ['.ppt', '.pptx'],
               "EverNoteLoader": ['.enex'],
               }
SUPPORTED_EXTS = [ext for sublist in LOADER_DICT.values() for ext in sublist]


# patch json.dumps to disable ensure_ascii
def _new_json_dumps(obj, **kwargs):
    kwargs["ensure_ascii"] = False
    return _origin_json_dumps(obj, **kwargs)

if json.dumps is not _new_json_dumps:
    _origin_json_dumps = json.dumps
    json.dumps = _new_json_dumps


class JSONLinesLoader(langchain.document_loaders.JSONLoader):
    '''
    行式 Json 加载器，要求文件扩展名为 .jsonl
    '''
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self._json_lines = True


langchain.document_loaders.JSONLinesLoader = JSONLinesLoader


def get_LoaderClass(file_extension):
    for LoaderClass, extensions in LOADER_DICT.items():
        if file_extension in extensions:
            return LoaderClass

def get_SplitterClass(file_extension):
    """
    根据文件类型获取文本分块器类型
    """
    # print('get Splitter Class', file_extension)
    for SplitterClass, extensions in TEXT_SPLITTER_MAP.items():
        if file_extension in extensions:
            return SplitterClass
    print(f'未找到文件类型"{file_extension}"对应的切分器，使用默认值')
    return TEXT_SPLITTER_NAME

def get_loader(loader_name: str, file_path: str, loader_kwargs: Dict = None):
    '''
    根据loader_name和文件路径或内容返回文档加载器。
    '''
    loader_kwargs = loader_kwargs or {}
    try:
        # print(loader_name)
        if loader_name in ["RapidOCRLoader", "FilteredCSVLoader",
                        "GCYWordLoader","GCYWordLoader2", "GCYHTMLLoader",
                        "RapidOCRPPTLoader", "RapidOCRCSVLoader","ExcelLoader"]:

            document_loaders_module = importlib.import_module('document_loaders')
        else:
            document_loaders_module = importlib.import_module('langchain.document_loaders')
        DocumentLoader = getattr(document_loaders_module, loader_name)
    except Exception as e:
        msg = f"为文件{file_path}查找加载器{loader_name}时出错：{e}"
        logger.error(f'{e.__class__.__name__}: {msg}',
                     exc_info=e if log_verbose else None)
        document_loaders_module = importlib.import_module('langchain.document_loaders')
        DocumentLoader = getattr(document_loaders_module, "UnstructuredFileLoader")

    if loader_name in ["UnstructuredFileLoader", "TextLoader"]:
        loader_kwargs.setdefault("autodetect_encoding", True)
    elif loader_name == "CSVLoader":
        if not loader_kwargs.get("encoding"):
            # 如果未指定 encoding，自动识别文件编码类型，避免langchain loader 加载文件报编码错误
            with open(file_path, 'rb') as struct_file:
                encode_detect = chardet.detect(struct_file.read())
            if encode_detect is None:
                encode_detect = {"encoding": "utf-8"}
            loader_kwargs["encoding"] = encode_detect["encoding"]

    elif loader_name == "JSONLoader":
        loader_kwargs.setdefault("jq_schema", ".")
        loader_kwargs.setdefault("text_content", False)
    elif loader_name == "JSONLinesLoader":
        loader_kwargs.setdefault("jq_schema", ".")
        loader_kwargs.setdefault("text_content", False)

    loader = DocumentLoader(file_path, **loader_kwargs)
    return loader


def make_text_splitter(
        splitter_name: str = TEXT_SPLITTER_NAME,
        chunk_size: int = CHUNK_SIZE,
        chunk_overlap: int = OVERLAP_SIZE,
        llm_model: str = LLM_MODELS[0],
):
    """
    根据参数获取特定的分词器
    """
    # print('spliter name', splitter_name)
    splitter_name = splitter_name or "SpacyTextSplitter"
    try:
        # if splitter_name == "GCYMarkdownTextSplitter":  # MarkdownHeaderTextSplitter特殊判定
        if splitter_name == "MarkdownTextSplitter":  # MarkdownHeaderTextSplitter特殊判定
            text_splitter_module = importlib.import_module('text_splitter')
            TextSplitter = getattr(text_splitter_module, splitter_name)
            headers_to_split_on = text_splitter_dict[splitter_name]['headers_to_split_on']
            text_splitter = TextSplitter(
                headers_to_split_on=headers_to_split_on,
                strip_headers=False, # 不要将标题从分块文本中去掉
                promote_headers=True
            )
        else:

            try:  ## 优先使用用户自定义的text_splitter
                text_splitter_module = importlib.import_module('text_splitter')
                TextSplitter = getattr(text_splitter_module, splitter_name)
            except:  ## 否则使用langchain的text_splitter
                text_splitter_module = importlib.import_module('langchain.text_splitter')
                TextSplitter = getattr(text_splitter_module, splitter_name)

            if text_splitter_dict[splitter_name]["source"] == "tiktoken":  ## 从tiktoken加载
                try:
                    text_splitter = TextSplitter.from_tiktoken_encoder(
                        encoding_name=text_splitter_dict[splitter_name]["tokenizer_name_or_path"],
                        pipeline="zh_core_web_sm",
                        chunk_size=chunk_size,
                        chunk_overlap=chunk_overlap
                    )
                except:
                    text_splitter = TextSplitter.from_tiktoken_encoder(
                        encoding_name=text_splitter_dict[splitter_name]["tokenizer_name_or_path"],
                        chunk_size=chunk_size,
                        chunk_overlap=chunk_overlap
                    )
            elif text_splitter_dict[splitter_name]["source"] == "huggingface":  ## 从huggingface加载
                if text_splitter_dict[splitter_name]["tokenizer_name_or_path"] == "":
                    config = get_model_worker_config(llm_model)
                    text_splitter_dict[splitter_name]["tokenizer_name_or_path"] = \
                        config.get("model_path")

                if text_splitter_dict[splitter_name]["tokenizer_name_or_path"] == "gpt2":
                    from transformers import GPT2TokenizerFast
                    from langchain.text_splitter import CharacterTextSplitter
                    tokenizer = GPT2TokenizerFast.from_pretrained("gpt2")
                else:  ## 字符长度加载
                    from transformers import AutoTokenizer
                    tokenizer = AutoTokenizer.from_pretrained(
                        text_splitter_dict[splitter_name]["tokenizer_name_or_path"],
                        trust_remote_code=True)
                text_splitter = TextSplitter.from_huggingface_tokenizer(
                    tokenizer=tokenizer,
                    chunk_size=chunk_size,
                    chunk_overlap=chunk_overlap
                )
            elif text_splitter_dict[splitter_name]["source"] == "no_tokenizer": # IAST 0429: 目前不需要使用分词器
                    text_splitter = TextSplitter(
                        chunk_size=chunk_size,
                        chunk_overlap=chunk_overlap
                    )
            else:
                try:
                    text_splitter = TextSplitter(
                        pipeline="zh_core_web_sm",
                        chunk_size=chunk_size,
                        chunk_overlap=chunk_overlap
                    )
                except:
                    text_splitter = TextSplitter(
                        chunk_size=chunk_size,
                        chunk_overlap=chunk_overlap
                    )
    except Exception as e:
        print(e)
        text_splitter_module = importlib.import_module('langchain.text_splitter')
        TextSplitter = getattr(text_splitter_module, "RecursiveCharacterTextSplitter")
        text_splitter = TextSplitter(chunk_size=chunk_size, chunk_overlap=chunk_overlap)

    # If you use SpacyTextSplitter you can use GPU to do split likes Issue #1287
    # text_splitter._tokenizer.max_length = 37016792
    # text_splitter._tokenizer.prefer_gpu()
    return text_splitter


class KnowledgeFile:
    def __init__(
            self,
            filename: str,
            knowledge_base_name: str,
            loader_kwargs: Dict = {},
    ):
        '''
        对应知识库目录中的文件，必须是磁盘上存在的才能进行向量化等操作。
        '''
        self.kb_name = knowledge_base_name
        self.filename = str(Path(filename).as_posix())
        self.ext = os.path.splitext(filename)[-1].lower()
        if self.ext not in SUPPORTED_EXTS:
            raise ValueError(f"暂未支持的文件格式 {self.filename}")
        self.loader_kwargs = loader_kwargs
        self.filepath = get_file_path(knowledge_base_name, filename)
        self.docs = None
        self.splited_docs = None
        self.document_loader_name = get_LoaderClass(self.ext)
        self.text_splitter_name = get_SplitterClass(self.ext)
    def get_full_text(self) -> Dict[str, str]:
        """
        获取文件的全文内容，并返回文件名和全文内容的结构。
        """
        try:
            docs = self.file2docs()
            full_text = "".join([doc.page_content for doc in docs])
            result = json.dumps( {
                "filename": self.filename,
                "full_text": full_text
            }, ensure_ascii=False)
            return result
        except Exception as e:
            logger.error(f"获取文件全文内容时出错：{e}", exc_info=e if log_verbose else None)
            return {
                "filename": self.filename,
                "full_text": "加载文件失败或文件内容为空"
            }
    async def get_llm_result(self) -> Dict[str, str]:
        """
        根据文件的全文内容，异步调用模型生成文章摘要、关键词和章节速览。
        """
        try:
            # full_text_data = self.get_full_text()
            # full_text = full_text_data.get("full_text", "")
            loop = asyncio.get_event_loop()
            full_text_data = await loop.run_in_executor(None, self.get_full_text)
            # full_text = full_text_data.get("full_text", "")
            try:
                # 将 JSON 字符串解析为字典
                full_text_dict = json.loads(full_text_data)
                full_text = full_text_dict.get("full_text", "")
            except json.JSONDecodeError:
                print("解析 JSON 数据时出错")
                full_text = ""
            if len(full_text) > 40000:
                # 判断英文占比
                # english_chars = re.findall(r'[a-zA-Z]', full_text)
                # english_ratio = len(english_chars) / len(full_text) if len(full_text) > 0 else 0
                # if english_ratio > 0.9 and len(full_text) > 50000:
                full_text = full_text[:40000]

                # logger.info(f'=============文章长度{len(full_text)}')
                # full_text_80 = TextRank(full_text, 80)
                # logger.info(f'=============按80句压缩后文章长度{len(full_text_80)}')
                # if len(full_text_80) > 55000:
                #     full_text_10 = TextRank(full_text_80, num_sentences=10)
                #     logger.info(f'=============按10句压缩后文章长度{len(full_text_10)}')
                #     full_text = full_text_10
                # else:
                #     full_text = full_text_80
            else:
                pass
            # 异步调用模型
            from asyncio import gather
            llm_time = time.time()
            abstract_task = get_llm_model_response_async(
                strategy_name="gen_abstract",
                llm_model_name=LLM_MODELS[0],
                template_prompt_name="gen_abstract",
                prompt_param_dict={"context": full_text},
                temperature=0.7,
                max_tokens=4096
            )

            keywords_task = get_llm_model_response_async(
                strategy_name="gen_keywords",
                llm_model_name=LLM_MODELS[0],
                template_prompt_name="gen_keywords",
                prompt_param_dict={"context": full_text},
                temperature=0.7,
                max_tokens=512
            )

            paragraph_task = get_llm_model_response_async(
                strategy_name="gen_paragraph",
                llm_model_name=LLM_MODELS[0],
                template_prompt_name="gen_paragraph",
                prompt_param_dict={"context": full_text},
                temperature=0.7,
                max_tokens=8192
            )

            # 并行执行任务
            article_abstract, article_keywords, article_paragraph = await gather(
                abstract_task, keywords_task, paragraph_task
            )

            logger.info(f'生成导读用时：{time.time() - llm_time}')
            return {
                "filename": self.filename,
                "full_text": full_text,
                "article_abstract": article_abstract,
                "article_keywords": article_keywords,
                "article_paragraph": article_paragraph
            }
        except Exception as e:
            logger.error(f"生成LLM结果时出错：{e}", exc_info=e if log_verbose else None)
            return {
                "filename": self.filename,
                "article_abstract": "生成摘要失败",
                "article_keywords": "生成关键词失败",
                "article_paragraph": "生成章节速览失败"
            }
    def file2docs(self, refresh: bool = False):
        if self.docs is None or refresh:
            try:
                logger.info(f"{self.document_loader_name} used for {self.filepath}")
                loader = get_loader(loader_name=self.document_loader_name,
                                    file_path=self.filepath,
                                    loader_kwargs=self.loader_kwargs)
                self.docs = loader.load()
            except Exception as e:
                if self.document_loader_name == 'GCYWordLoader':
                    loader = get_loader(loader_name='GCYWordLoader2',
                                        file_path=self.filepath,
                                        loader_kwargs=self.loader_kwargs)
                else:
                    logger.error(f"加载文件 {self.filepath} 时出错：{e}", exc_info=e if log_verbose else None)
                self.docs = loader.load()
        return self.docs

    @get_split_time
    def docs2texts(
            self,
            docs: List[Document] = None,
            zh_title_enhance: bool = ZH_TITLE_ENHANCE,
            refresh: bool = False,
            chunk_size: int = CHUNK_SIZE,
            chunk_overlap: int = OVERLAP_SIZE,
            text_splitter: TextSplitter = None,
    ):
        docs = docs or self.file2docs(refresh=refresh)
        # debug 0429
        # print('docs2texts',docs )
        if not docs:
            return []
        
        if text_splitter is None:
            self.text_splitter_name = get_SplitterClass(self.ext)
            text_splitter = make_text_splitter(splitter_name=self.text_splitter_name, chunk_size=chunk_size,
                                                chunk_overlap=chunk_overlap)
        # if self.text_splitter_name == "GCYMarkdownTextSplitter":
        if self.text_splitter_name == "MarkdownTextSplitter":
            doc_source = (docs[0].metadata)["source"]
            docs = text_splitter.split_markdown_text(docs[0].page_content, doc_source)
        else:
            docs = text_splitter.split_documents(docs)

        if not docs:
            return []
        
        # 检查切分好的文档是否有'h1'标题字段，如果没有，就加上。为之后入库其它有h1的文件做准备
        if 'h1' not in docs[0].metadata:
            for doc in docs:
                doc.metadata['h1'] = ''

        print(f"文档切分示例：{docs[0]}")
        if zh_title_enhance:
            docs = func_zh_title_enhance(docs)
        self.splited_docs = docs
        return self.splited_docs

    def file2text(
            self,
            zh_title_enhance: bool = ZH_TITLE_ENHANCE,
            refresh: bool = False,
            chunk_size: int = CHUNK_SIZE,
            chunk_overlap: int = OVERLAP_SIZE,
            text_splitter: TextSplitter = None,
    ):
        if self.splited_docs is None or refresh:
            docs = self.file2docs()
            self.splited_docs = self.docs2texts(docs=docs,
                                                zh_title_enhance=zh_title_enhance,
                                                refresh=refresh,
                                                chunk_size=chunk_size,
                                                chunk_overlap=chunk_overlap,
                                                text_splitter=text_splitter)
        return self.splited_docs

    def file_exist(self):
        return os.path.isfile(self.filepath)

    def get_mtime(self):
        return os.path.getmtime(self.filepath)

    def get_size(self):
        return os.path.getsize(self.filepath)


def files2docs_in_thread(
        files: List[Union[KnowledgeFile, Tuple[str, str], Dict]],
        chunk_size: int = CHUNK_SIZE,
        chunk_overlap: int = OVERLAP_SIZE,
        zh_title_enhance: bool = ZH_TITLE_ENHANCE,
) -> Generator:
    '''
    利用多线程批量将磁盘文件转化成langchain Document.
    如果传入参数是Tuple，形式为(filename, kb_name)
    生成器返回值为 status, (kb_name, file_name, docs | error)
    '''

    def file2docs(*, file: KnowledgeFile, **kwargs) -> Tuple[bool, Tuple[str, str, List[Document]]]:
        try:
            return True, (file.kb_name, file.filename, file.file2text(**kwargs))
        except Exception as e:
            msg = f"从文件 {file.kb_name}/{file.filename} 加载文档时出错：{e}"
            logger.error(f'{e.__class__.__name__}: {msg}',
                         exc_info=e if log_verbose else None)
            return False, (file.kb_name, file.filename, msg)

    kwargs_list = []
    for i, file in enumerate(files):
        kwargs = {}
        try:
            if isinstance(file, tuple) and len(file) >= 2:
                filename = file[0]
                kb_name = file[1]
                file = KnowledgeFile(filename=filename, knowledge_base_name=kb_name)
            elif isinstance(file, dict):
                filename = file.pop("filename")
                kb_name = file.pop("kb_name")
                kwargs.update(file)
                file = KnowledgeFile(filename=filename, knowledge_base_name=kb_name)
            kwargs["file"] = file
            kwargs["chunk_size"] = chunk_size
            kwargs["chunk_overlap"] = chunk_overlap
            kwargs["zh_title_enhance"] = zh_title_enhance
            kwargs_list.append(kwargs)
        except Exception as e:
            yield False, (kb_name, filename, str(e))

    for result in run_in_thread_pool(func=file2docs, params=kwargs_list):
        yield result


if __name__ == "__main__":
    from pprint import pprint

    kb_file = KnowledgeFile(
        filename="/home/congyin/Code/Project_Langchain_0814/Langchain-Chatchat/knowledge_base/csv1/content/gm.csv",
        knowledge_base_name="samples")
    # kb_file.text_splitter_name = "RecursiveCharacterTextSplitter"
    docs = kb_file.file2docs()
    # pprint(docs[-1])