比较提交

..

25 次代码提交

作者 SHA1 备注 提交日期
binary-husky
b3670e7a6a Update requirements.txt 2023-04-13 12:27:28 +08:00
binary-husky
daef235c5d Update README.md 2023-04-13 12:20:54 +08:00
binary-husky
8c24271a0a Update README.md 2023-04-13 12:17:59 +08:00
binary-husky
e25189b2f9 Update 高级功能函数模板.py 2023-04-13 12:02:52 +08:00
binary-husky
1dd6cf1ae9 Update README.md 2023-04-13 11:53:52 +08:00
binary-husky
5927f0d6de Merge pull request #436 from DDreame/patch-1
[fix]Update requirements.txt For fix
2023-04-13 11:47:19 +08:00
binary-husky
58fa339c75 Merge pull request #439 from mrhblfx/patch-2
增加解析Lua项目
2023-04-13 11:45:42 +08:00
binary-husky
cceabd0f7d Merge pull request #438 from mrhblfx/patch-1
增加了解析Go项目匹配的文件: `go.mod`, `go.sum`, `go.work`
2023-04-13 11:44:03 +08:00
qingxu fu
1c6c29e6dd 更新注释 2023-04-13 11:18:44 +08:00
qingxu fu
74f26a7d64 改善chatpdf的功能 2023-04-13 11:08:53 +08:00
mrhblfx
84e94e50e2 Update crazy_functional.py 2023-04-12 22:44:07 +08:00
mrhblfx
1d3f595fb5 新增解析一个Lua项目 2023-04-12 22:39:31 +08:00
mrhblfx
2498833ee1 增加了解析Go项目匹配的文件: go.mod, go.sum, go.work 2023-04-12 22:33:34 +08:00
无知之人
bb079d9b81 [fix]Update requirements.txt For fix
Modify the version of Gradio, which does not support the color button when it is lower than version 3.24. On version 3.25, it fixes the issue https://github.com/gradio-app/gradio/issues/3716 and #371 .

修改一下gradio的版本,低于 3.24版本时不支持 color button,3.25则修复了 issue: #371https://github.com/gradio-app/gradio/issues/3716
2023-04-12 21:32:03 +08:00
fuqingxu
1c96ed39c3 启动ChatGPT+ChatGLM 2023-04-12 17:41:33 +08:00
fuqingxu
5b040d552e Merge branch 'master' of https://github.com/binary-husky/chatgpt_academic 2023-04-12 16:57:56 +08:00
fuqingxu
c563ac2274 UTF8 Ignore read file errors 2023-04-12 16:57:01 +08:00
binary-husky
d28af7611a Update README.md 2023-04-12 12:04:57 +08:00
binary-husky
9714d3ea2d Update README.md 2023-04-12 11:21:20 +08:00
binary-husky
65c51eb05b Update README.md 2023-04-12 02:46:04 +08:00
binary-husky
5d98d82526 Update README.md 2023-04-12 02:41:43 +08:00
binary-husky
9215199ae1 Update README.md 2023-04-12 02:27:55 +08:00
binary-husky
7ded328970 Update README.md 2023-04-12 02:27:10 +08:00
binary-husky
057b8cd943 Update README.md 2023-04-12 00:12:42 +08:00
binary-husky
2ab6acc6de Update README.md 2023-04-12 00:08:31 +08:00
共有 16 个文件被更改,包括 338 次插入329 次删除

50
Dockerfile+ChatGLM 普通文件
查看文件

@@ -0,0 +1,50 @@
# How to build | 如何构建: docker build -t gpt-academic --network=host -f Dockerfile+ChatGLM .
# How to run | 如何运行 (1) 直接运行: docker run --rm -it --net=host --gpus=all gpt-academic
# How to run | 如何运行 (2) 我想运行之前进容器做一些调整: docker run --rm -it --net=host --gpus=all gpt-academic bash
# 从NVIDIA源,从而支持显卡运损检查宿主的nvidia-smi中的cuda版本必须>=11.3
FROM nvidia/cuda:11.3.1-runtime-ubuntu20.04
ARG useProxyNetwork=''
RUN apt-get update
RUN apt-get install -y curl proxychains curl
RUN apt-get install -y git python python3 python-dev python3-dev --fix-missing
# 配置代理网络构建Docker镜像时使用
# # comment out below if you do not need proxy network | 如果不需要翻墙 - 从此行向下删除
RUN $useProxyNetwork curl cip.cc
RUN sed -i '$ d' /etc/proxychains.conf
RUN sed -i '$ d' /etc/proxychains.conf
RUN echo "socks5 127.0.0.1 10880" >> /etc/proxychains.conf
ARG useProxyNetwork=proxychains
# # comment out above if you do not need proxy network | 如果不需要翻墙 - 从此行向上删除
# use python3 as the system default python
RUN curl -sS https://bootstrap.pypa.io/get-pip.py | python3.8
# 下载分支
WORKDIR /gpt
RUN $useProxyNetwork git clone https://github.com/binary-husky/chatgpt_academic.git -b v3.0
WORKDIR /gpt/chatgpt_academic
RUN $useProxyNetwork python3 -m pip install -r requirements.txt
RUN $useProxyNetwork python3 -m pip install -r request_llm/requirements_chatglm.txt
RUN $useProxyNetwork python3 -m pip install torch --extra-index-url https://download.pytorch.org/whl/cu113
# 预热CHATGLM参数非必要 可选步骤)
RUN echo ' \n\
from transformers import AutoModel, AutoTokenizer \n\
chatglm_tokenizer = AutoTokenizer.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True) \n\
chatglm_model = AutoModel.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True).float() ' >> warm_up_chatglm.py
RUN python3 -u warm_up_chatglm.py
RUN $useProxyNetwork git pull
# 为chatgpt-academic配置代理和API-KEY (非必要 可选步骤)
RUN echo ' \n\
API_KEY = "sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx" \n\
USE_PROXY = True \n\
LLM_MODEL = "chatglm" \n\
LOCAL_MODEL_DEVICE = "cuda" \n\
proxies = { "http": "socks5h://localhost:10880", "https": "socks5h://localhost:10880", } ' >> config_private.py
# 启动
CMD ["python3", "-u", "main.py"]

查看文件

@@ -1,9 +1,3 @@
> **Note**
>
> 紧急很抱歉2.60版本的一部分代码重构出错,目前2.67及以上版本已经解决,请您及时更新。
>
# ChatGPT 学术优化
**如果喜欢这个项目,请给它一个Star;如果你发明了更好用的快捷键或函数插件,欢迎发issue或者pull requestsdev分支**
@@ -12,7 +6,7 @@ If you like this project, please give it a Star. If you've come up with more use
> **Note**
>
> 1.请注意只有红颜色标识的函数插件(按钮)才支持读取文件。目前对pdf/word格式文件的支持插件正在逐步完善中,需要更多developer的帮助。
> 1.请注意只有**红颜色**标识的函数插件(按钮)才支持读取文件,部分插件位于插件区的**下拉菜单**中。另外我们以**最高优先级**欢迎和处理任何新插件的PR
>
> 2.本项目中每个文件的功能都在自译解[`self_analysis.md`](https://github.com/binary-husky/chatgpt_academic/wiki/chatgpt-academic%E9%A1%B9%E7%9B%AE%E8%87%AA%E8%AF%91%E8%A7%A3%E6%8A%A5%E5%91%8A)详细说明。随着版本的迭代,您也可以随时自行点击相关函数插件,调用GPT重新生成项目的自我解析报告。常见问题汇总在[`wiki`](https://github.com/binary-husky/chatgpt_academic/wiki/%E5%B8%B8%E8%A7%81%E9%97%AE%E9%A2%98)当中。
>
@@ -30,9 +24,10 @@ If you like this project, please give it a Star. If you've come up with more use
[自定义快捷键](https://www.bilibili.com/video/BV14s4y1E7jN) | 支持自定义快捷键
[配置代理服务器](https://www.bilibili.com/video/BV1rc411W7Dr) | 支持配置代理服务器
模块化设计 | 支持自定义高阶的实验性功能与[函数插件],插件支持[热更新](https://github.com/binary-husky/chatgpt_academic/wiki/%E5%87%BD%E6%95%B0%E6%8F%92%E4%BB%B6%E6%8C%87%E5%8D%97)
[自我程序剖析](https://www.bilibili.com/video/BV1cj411A7VW) | [函数插件] 一键读懂本项目的源代码
[程序剖析](https://www.bilibili.com/video/BV1cj411A7VW) | [函数插件] 一键可以剖析其他Python/C/C++/Java项目树
[自我程序剖析](https://www.bilibili.com/video/BV1cj411A7VW) | [函数插件] [一键读懂](https://github.com/binary-husky/chatgpt_academic/wiki/chatgpt-academic%E9%A1%B9%E7%9B%AE%E8%87%AA%E8%AF%91%E8%A7%A3%E6%8A%A5%E5%91%8A)本项目的源代码
[程序剖析](https://www.bilibili.com/video/BV1cj411A7VW) | [函数插件] 一键可以剖析其他Python/C/C++/Java/Golang/Lua/Rect项目树
读论文 | [函数插件] 一键解读latex论文全文并生成摘要
Latex全文翻译、润色 | [函数插件] 一键翻译或润色latex论文
批量注释生成 | [函数插件] 一键批量生成函数注释
chat分析报告生成 | [函数插件] 运行后自动生成总结汇报
[arxiv小助手](https://www.bilibili.com/video/BV1LM4y1279X) | [函数插件] 输入arxiv文章url即可一键翻译摘要+下载PDF
@@ -44,12 +39,14 @@ chat分析报告生成 | [函数插件] 运行后自动生成总结汇报
支持GPT输出的markdown表格 | 可以输出支持GPT的markdown表格
启动暗色gradio[主题](https://github.com/binary-husky/chatgpt_academic/issues/173) | 在浏览器url后面添加```/?__dark-theme=true```可以切换dark主题
huggingface免科学上网[在线体验](https://huggingface.co/spaces/qingxu98/gpt-academic) | 登陆huggingface后复制[此空间](https://huggingface.co/spaces/qingxu98/gpt-academic)
[多LLM模型](https://www.bilibili.com/video/BV1EM411K7VH/)混合支持([v3.0分支](https://github.com/binary-husky/chatgpt_academic/tree/v3.0)测试中) | 同时被ChatGPT和[清华ChatGLM](https://github.com/THUDM/ChatGLM-6B)伺候的感觉一定会很不错吧?
兼容[TGUI](https://github.com/oobabooga/text-generation-webui)接入更多样的语言模型 | 接入opt-1.3b, galactica-1.3b等模型([v3.0分支](https://github.com/binary-husky/chatgpt_academic/tree/v3.0)测试中)
…… | ……
</div>
<!-- - 新界面master主分支, 右dev开发前沿 -->
- 新界面
- 新界面修改config.py中的LAYOUT选项即可实现“左右布局”和“上下布局”的切换
<div align="center">
<img src="https://user-images.githubusercontent.com/96192199/230361456-61078362-a966-4eb5-b49e-3c62ef18b860.gif" width="700" >
</div>
@@ -84,6 +81,14 @@ huggingface免科学上网[在线体验](https://huggingface.co/spaces/qingxu98/
<img src="https://user-images.githubusercontent.com/96192199/226935232-6b6a73ce-8900-4aee-93f9-733c7e6fef53.png" width="700" >
</div>
- 多种大语言模型混合调用([v3.0分支](https://github.com/binary-husky/chatgpt_academic/tree/v3.0)测试中)
<div align="center">
<img src="https://user-images.githubusercontent.com/96192199/231222778-34776885-a7f0-4f2c-b5f4-7cc2ef3ecb58.png" width="700" >
</div>
## 直接运行 (Windows, Linux or MacOS)
### 1. 下载项目
@@ -287,8 +292,7 @@ python check_proxy.py
# 借鉴项目1借鉴了ChuanhuChatGPT中读取OpenAI json的方法、记录历史问询记录的方法以及gradio queue的使用技巧
https://github.com/GaiZhenbiao/ChuanhuChatGPT
# 借鉴项目2借鉴了mdtex2html中公式处理的方法
https://github.com/polarwinkel/mdtex2html
# 借鉴项目2
https://github.com/THUDM/ChatGLM-6B
```

查看文件

@@ -76,7 +76,6 @@ def get_crazy_functions():
from crazy_functions.总结word文档 import 总结word文档
from crazy_functions.批量翻译PDF文档_多线程 import 批量翻译PDF文档
from crazy_functions.谷歌检索小助手 import 谷歌检索小助手
from crazy_functions.理解PDF文档内容 import 理解PDF文档内容
from crazy_functions.理解PDF文档内容 import 理解PDF文档内容标准文件输入
from crazy_functions.Latex全文润色 import Latex中文润色
from crazy_functions.Latex全文翻译 import Latex中译英
@@ -108,12 +107,7 @@ def get_crazy_functions():
"Color": "stop",
"Function": HotReload(总结word文档)
},
# "[测试功能] 理解PDF文档内容Tk文件选择接口,仅本地": {
# # HotReload 的意思是热更新,修改函数插件代码后,不需要重启程序,代码直接生效
# "AsButton": False, # 加入下拉菜单中
# "Function": HotReload(理解PDF文档内容)
# },
"[测试功能] 理解PDF文档内容通用接口,读取文件输入区": {
"理解PDF文档内容 模仿ChatPDF": {
# HotReload 的意思是热更新,修改函数插件代码后,不需要重启程序,代码直接生效
"Color": "stop",
"AsButton": False, # 加入下拉菜单中
@@ -131,7 +125,6 @@ def get_crazy_functions():
"AsButton": False, # 加入下拉菜单中
"Function": HotReload(Latex中文润色)
},
"[测试功能] Latex项目全文中译英输入路径或上传压缩包": {
# HotReload 的意思是热更新,修改函数插件代码后,不需要重启程序,代码直接生效
"Color": "stop",
@@ -163,5 +156,14 @@ def get_crazy_functions():
except Exception as err:
print(f'[下载arxiv论文并翻译摘要] 插件导入失败 {str(err)}')
from crazy_functions.解析项目源代码 import 解析一个Lua项目
function_plugins.update({
"解析整个Lua项目": {
"Color": "stop", # 按钮颜色
"AsButton": False, # 加入下拉菜单中
"Function": HotReload(解析一个Lua项目)
},
})
###################### 第n组插件 ###########################
return function_plugins

查看文件

@@ -45,7 +45,7 @@ def 多文件润色(file_manifest, project_folder, llm_kwargs, plugin_kwargs, ch
pfg = PaperFileGroup()
for index, fp in enumerate(file_manifest):
with open(fp, 'r', encoding='utf-8') as f:
with open(fp, 'r', encoding='utf-8', errors='replace') as f:
file_content = f.read()
# 定义注释的正则表达式
comment_pattern = r'%.*'

查看文件

@@ -44,7 +44,7 @@ def 多文件翻译(file_manifest, project_folder, llm_kwargs, plugin_kwargs, ch
pfg = PaperFileGroup()
for index, fp in enumerate(file_manifest):
with open(fp, 'r', encoding='utf-8') as f:
with open(fp, 'r', encoding='utf-8', errors='replace') as f:
file_content = f.read()
# 定义注释的正则表达式
comment_pattern = r'%.*'

查看文件

@@ -360,3 +360,171 @@ def breakdown_txt_to_satisfy_token_limit_for_pdf(txt, get_token_fn, limit):
# 这个中文的句号是故意的,作为一个标识而存在
res = cut(txt.replace('.', '\n'), must_break_at_empty_line=False)
return [r.replace('\n', '.') for r in res]
def read_and_clean_pdf_text(fp):
"""
这个函数用于分割pdf,用了很多trick,逻辑较乱,效果奇好
**输入参数说明**
- `fp`需要读取和清理文本的pdf文件路径
**输出参数说明**
- `meta_txt`:清理后的文本内容字符串
- `page_one_meta`:第一页清理后的文本内容列表
**函数功能**
读取pdf文件并清理其中的文本内容,清理规则包括
- 提取所有块元的文本信息,并合并为一个字符串
- 去除短块字符数小于100并替换为回车符
- 清理多余的空行
- 合并小写字母开头的段落块并替换为空格
- 清除重复的换行
- 将每个换行符替换为两个换行符,使每个段落之间有两个换行符分隔
"""
import fitz, copy
import re
import numpy as np
from colorful import print亮黄, print亮绿
fc = 0
fs = 1
fb = 2
REMOVE_FOOT_NOTE = True
REMOVE_FOOT_FFSIZE_PERCENT = 0.95
def primary_ffsize(l):
fsize_statiscs = {}
for wtf in l['spans']:
if wtf['size'] not in fsize_statiscs: fsize_statiscs[wtf['size']] = 0
fsize_statiscs[wtf['size']] += len(wtf['text'])
return max(fsize_statiscs, key=fsize_statiscs.get)
def ffsize_same(a,b):
return abs((a-b)/max(a,b)) < 0.02
# file_content = ""
with fitz.open(fp) as doc:
meta_txt = []
meta_font = []
meta_line = []
meta_span = []
for index, page in enumerate(doc):
# file_content += page.get_text()
text_areas = page.get_text("dict") # 获取页面上的文本信息
for t in text_areas['blocks']:
if 'lines' in t:
pf = 998
for l in t['lines']:
txt_line = "".join([wtf['text'] for wtf in l['spans']])
pf = primary_ffsize(l)
meta_line.append([txt_line, pf, l['bbox'], l])
for wtf in l['spans']: # for l in t['lines']:
meta_span.append([wtf['text'], wtf['size'], len(wtf['text'])])
# meta_line.append(["NEW_BLOCK", pf])
# 块元提取 for each word segment with in line for each line cross-line words for each block
meta_txt.extend([" ".join(["".join([wtf['text'] for wtf in l['spans']]) for l in t['lines']]).replace(
'- ', '') for t in text_areas['blocks'] if 'lines' in t])
meta_font.extend([np.mean([np.mean([wtf['size'] for wtf in l['spans']])
for l in t['lines']]) for t in text_areas['blocks'] if 'lines' in t])
if index == 0:
page_one_meta = [" ".join(["".join([wtf['text'] for wtf in l['spans']]) for l in t['lines']]).replace(
'- ', '') for t in text_areas['blocks'] if 'lines' in t]
# 获取正文主字体
fsize_statiscs = {}
for span in meta_span:
if span[1] not in fsize_statiscs: fsize_statiscs[span[1]] = 0
fsize_statiscs[span[1]] += span[2]
main_fsize = max(fsize_statiscs, key=fsize_statiscs.get)
if REMOVE_FOOT_NOTE:
give_up_fize_threshold = main_fsize * REMOVE_FOOT_FFSIZE_PERCENT
# 切分和重新整合
mega_sec = []
sec = []
for index, line in enumerate(meta_line):
if index == 0:
sec.append(line[fc])
continue
if REMOVE_FOOT_NOTE:
if meta_line[index][fs] <= give_up_fize_threshold:
continue
if ffsize_same(meta_line[index][fs], meta_line[index-1][fs]):
# 尝试识别段落
if meta_line[index][fc].endswith('.') and\
(meta_line[index-1][fc] != 'NEW_BLOCK') and \
(meta_line[index][fb][2] - meta_line[index][fb][0]) < (meta_line[index-1][fb][2] - meta_line[index-1][fb][0]) * 0.7:
sec[-1] += line[fc]
sec[-1] += "\n\n"
else:
sec[-1] += " "
sec[-1] += line[fc]
else:
if (index+1 < len(meta_line)) and \
meta_line[index][fs] > main_fsize:
# 单行 + 字体大
mega_sec.append(copy.deepcopy(sec))
sec = []
sec.append("# " + line[fc])
else:
# 尝试识别section
if meta_line[index-1][fs] > meta_line[index][fs]:
sec.append("\n" + line[fc])
else:
sec.append(line[fc])
mega_sec.append(copy.deepcopy(sec))
finals = []
for ms in mega_sec:
final = " ".join(ms)
final = final.replace('- ', ' ')
finals.append(final)
meta_txt = finals
def 把字符太少的块清除为回车(meta_txt):
for index, block_txt in enumerate(meta_txt):
if len(block_txt) < 100:
meta_txt[index] = '\n'
return meta_txt
meta_txt = 把字符太少的块清除为回车(meta_txt)
def 清理多余的空行(meta_txt):
for index in reversed(range(1, len(meta_txt))):
if meta_txt[index] == '\n' and meta_txt[index-1] == '\n':
meta_txt.pop(index)
return meta_txt
meta_txt = 清理多余的空行(meta_txt)
def 合并小写开头的段落块(meta_txt):
def starts_with_lowercase_word(s):
pattern = r"^[a-z]+"
match = re.match(pattern, s)
if match:
return True
else:
return False
for _ in range(100):
for index, block_txt in enumerate(meta_txt):
if starts_with_lowercase_word(block_txt):
if meta_txt[index-1] != '\n':
meta_txt[index-1] += ' '
else:
meta_txt[index-1] = ''
meta_txt[index-1] += meta_txt[index]
meta_txt[index] = '\n'
return meta_txt
meta_txt = 合并小写开头的段落块(meta_txt)
meta_txt = 清理多余的空行(meta_txt)
meta_txt = '\n'.join(meta_txt)
# 清除重复的换行
for _ in range(5):
meta_txt = meta_txt.replace('\n\n', '\n')
# 换行 -> 双换行
meta_txt = meta_txt.replace('\n', '\n\n')
for f in finals:
print亮黄(f)
print亮绿('***************************')
return meta_txt, page_one_meta

查看文件

@@ -49,7 +49,7 @@ def 全项目切换英文(txt, llm_kwargs, plugin_kwargs, chatbot, history, sys_
# 第4步随便显示点什么防止卡顿的感觉
for index, fp in enumerate(file_manifest):
# if 'test_project' in fp: continue
with open(fp, 'r', encoding='utf-8') as f:
with open(fp, 'r', encoding='utf-8', errors='replace') as f:
file_content = f.read()
i_say_show_user =f'[{index}/{len(file_manifest)}] 接下来请将以下代码中包含的所有中文转化为英文,只输出转化后的英文代码,请用代码块输出代码: {os.path.abspath(fp)}'
i_say_show_user_buffer.append(i_say_show_user)
@@ -72,7 +72,7 @@ def 全项目切换英文(txt, llm_kwargs, plugin_kwargs, chatbot, history, sys_
if index > 10:
time.sleep(60)
print('Openai 限制免费用户每分钟20次请求,降低请求频率中。')
with open(fp, 'r', encoding='utf-8') as f:
with open(fp, 'r', encoding='utf-8', errors='replace') as f:
file_content = f.read()
i_say_template = lambda fp, file_content: f'接下来请将以下代码中包含的所有中文转化为英文,只输出代码,文件名是{fp},文件代码是 ```{file_content}```'
try:

查看文件

@@ -68,7 +68,7 @@ def 解析Paper(file_manifest, project_folder, llm_kwargs, plugin_kwargs, chatbo
print('begin analysis on:', file_manifest)
for index, fp in enumerate(file_manifest):
if ".tex" in fp:
with open(fp, 'r', encoding='utf-8') as f:
with open(fp, 'r', encoding='utf-8', errors='replace') as f:
file_content = f.read()
if ".pdf" in fp.lower():
file_content = readPdf(fp)

查看文件

@@ -2,174 +2,9 @@ from toolbox import CatchException, report_execption, write_results_to_file
from toolbox import update_ui
from .crazy_utils import request_gpt_model_in_new_thread_with_ui_alive
from .crazy_utils import request_gpt_model_multi_threads_with_very_awesome_ui_and_high_efficiency
from .crazy_utils import read_and_clean_pdf_text
from colorful import *
def read_and_clean_pdf_text(fp):
"""
这个函数用于分割pdf,用了很多trick,逻辑较乱,效果奇好,不建议任何人去读这个函数
**输入参数说明**
- `fp`需要读取和清理文本的pdf文件路径
**输出参数说明**
- `meta_txt`:清理后的文本内容字符串
- `page_one_meta`:第一页清理后的文本内容列表
**函数功能**
读取pdf文件并清理其中的文本内容,清理规则包括
- 提取所有块元的文本信息,并合并为一个字符串
- 去除短块字符数小于100并替换为回车符
- 清理多余的空行
- 合并小写字母开头的段落块并替换为空格
- 清除重复的换行
- 将每个换行符替换为两个换行符,使每个段落之间有两个换行符分隔
"""
import fitz, copy
import re
import numpy as np
fc = 0
fs = 1
fb = 2
REMOVE_FOOT_NOTE = True
REMOVE_FOOT_FFSIZE_PERCENT = 0.95
def primary_ffsize(l):
fsize_statiscs = {}
for wtf in l['spans']:
if wtf['size'] not in fsize_statiscs: fsize_statiscs[wtf['size']] = 0
fsize_statiscs[wtf['size']] += len(wtf['text'])
return max(fsize_statiscs, key=fsize_statiscs.get)
def ffsize_same(a,b):
return abs((a-b)/max(a,b)) < 0.02
# file_content = ""
with fitz.open(fp) as doc:
meta_txt = []
meta_font = []
meta_line = []
meta_span = []
for index, page in enumerate(doc):
# file_content += page.get_text()
text_areas = page.get_text("dict") # 获取页面上的文本信息
for t in text_areas['blocks']:
if 'lines' in t:
pf = 998
for l in t['lines']:
txt_line = "".join([wtf['text'] for wtf in l['spans']])
pf = primary_ffsize(l)
meta_line.append([txt_line, pf, l['bbox'], l])
for wtf in l['spans']: # for l in t['lines']:
meta_span.append([wtf['text'], wtf['size'], len(wtf['text'])])
# meta_line.append(["NEW_BLOCK", pf])
# 块元提取 for each word segment with in line for each line cross-line words for each block
meta_txt.extend([" ".join(["".join([wtf['text'] for wtf in l['spans']]) for l in t['lines']]).replace(
'- ', '') for t in text_areas['blocks'] if 'lines' in t])
meta_font.extend([np.mean([np.mean([wtf['size'] for wtf in l['spans']])
for l in t['lines']]) for t in text_areas['blocks'] if 'lines' in t])
if index == 0:
page_one_meta = [" ".join(["".join([wtf['text'] for wtf in l['spans']]) for l in t['lines']]).replace(
'- ', '') for t in text_areas['blocks'] if 'lines' in t]
# 获取正文主字体
fsize_statiscs = {}
for span in meta_span:
if span[1] not in fsize_statiscs: fsize_statiscs[span[1]] = 0
fsize_statiscs[span[1]] += span[2]
main_fsize = max(fsize_statiscs, key=fsize_statiscs.get)
if REMOVE_FOOT_NOTE:
give_up_fize_threshold = main_fsize * REMOVE_FOOT_FFSIZE_PERCENT
# 切分和重新整合
mega_sec = []
sec = []
for index, line in enumerate(meta_line):
if index == 0:
sec.append(line[fc])
continue
if REMOVE_FOOT_NOTE:
if meta_line[index][fs] <= give_up_fize_threshold:
continue
if ffsize_same(meta_line[index][fs], meta_line[index-1][fs]):
# 尝试识别段落
if meta_line[index][fc].endswith('.') and\
(meta_line[index-1][fc] != 'NEW_BLOCK') and \
(meta_line[index][fb][2] - meta_line[index][fb][0]) < (meta_line[index-1][fb][2] - meta_line[index-1][fb][0]) * 0.7:
sec[-1] += line[fc]
sec[-1] += "\n\n"
else:
sec[-1] += " "
sec[-1] += line[fc]
else:
if (index+1 < len(meta_line)) and \
meta_line[index][fs] > main_fsize:
# 单行 + 字体大
mega_sec.append(copy.deepcopy(sec))
sec = []
sec.append("# " + line[fc])
else:
# 尝试识别section
if meta_line[index-1][fs] > meta_line[index][fs]:
sec.append("\n" + line[fc])
else:
sec.append(line[fc])
mega_sec.append(copy.deepcopy(sec))
finals = []
for ms in mega_sec:
final = " ".join(ms)
final = final.replace('- ', ' ')
finals.append(final)
meta_txt = finals
def 把字符太少的块清除为回车(meta_txt):
for index, block_txt in enumerate(meta_txt):
if len(block_txt) < 100:
meta_txt[index] = '\n'
return meta_txt
meta_txt = 把字符太少的块清除为回车(meta_txt)
def 清理多余的空行(meta_txt):
for index in reversed(range(1, len(meta_txt))):
if meta_txt[index] == '\n' and meta_txt[index-1] == '\n':
meta_txt.pop(index)
return meta_txt
meta_txt = 清理多余的空行(meta_txt)
def 合并小写开头的段落块(meta_txt):
def starts_with_lowercase_word(s):
pattern = r"^[a-z]+"
match = re.match(pattern, s)
if match:
return True
else:
return False
for _ in range(100):
for index, block_txt in enumerate(meta_txt):
if starts_with_lowercase_word(block_txt):
if meta_txt[index-1] != '\n':
meta_txt[index-1] += ' '
else:
meta_txt[index-1] = ''
meta_txt[index-1] += meta_txt[index]
meta_txt[index] = '\n'
return meta_txt
meta_txt = 合并小写开头的段落块(meta_txt)
meta_txt = 清理多余的空行(meta_txt)
meta_txt = '\n'.join(meta_txt)
# 清除重复的换行
for _ in range(5):
meta_txt = meta_txt.replace('\n\n', '\n')
# 换行 -> 双换行
meta_txt = meta_txt.replace('\n', '\n\n')
for f in finals:
print亮黄(f)
print亮绿('***************************')
return meta_txt, page_one_meta
@CatchException
def 批量翻译PDF文档(txt, llm_kwargs, plugin_kwargs, chatbot, history, sys_prompt, web_port):
import glob

查看文件

@@ -1,142 +1,67 @@
from toolbox import update_ui
from toolbox import CatchException, report_execption
import re
import unicodedata
from .crazy_utils import read_and_clean_pdf_text
from .crazy_utils import request_gpt_model_in_new_thread_with_ui_alive
fast_debug = False
def is_paragraph_break(match):
"""
根据给定的匹配结果来判断换行符是否表示段落分隔。
如果换行符前为句子结束标志(句号,感叹号,问号),且下一个字符为大写字母,则换行符更有可能表示段落分隔。
也可以根据之前的内容长度来判断段落是否已经足够长。
"""
prev_char, next_char = match.groups()
# 句子结束标志
sentence_endings = ".!?"
# 设定一个最小段落长度阈值
min_paragraph_length = 140
if prev_char in sentence_endings and next_char.isupper() and len(match.string[:match.start(1)]) > min_paragraph_length:
return "\n\n"
else:
return " "
def normalize_text(text):
"""
通过把连字ligatures等文本特殊符号转换为其基本形式来对文本进行归一化处理。
例如,将连字 "fi" 转换为 "f""i"
"""
# 对文本进行归一化处理,分解连字
normalized_text = unicodedata.normalize("NFKD", text)
# 替换其他特殊字符
cleaned_text = re.sub(r'[^\x00-\x7F]+', '', normalized_text)
return cleaned_text
def clean_text(raw_text):
"""
对从 PDF 提取出的原始文本进行清洗和格式化处理。
1. 对原始文本进行归一化处理。
2. 替换跨行的连词,例如 “Espe-\ncially” 转换为 “Especially”。
3. 根据 heuristic 规则判断换行符是否是段落分隔,并相应地进行替换。
"""
# 对文本进行归一化处理
normalized_text = normalize_text(raw_text)
# 替换跨行的连词
text = re.sub(r'(\w+-\n\w+)', lambda m: m.group(1).replace('-\n', ''), normalized_text)
# 根据前后相邻字符的特点,找到原文本中的换行符
newlines = re.compile(r'(\S)\n(\S)')
# 根据 heuristic 规则,用空格或段落分隔符替换原换行符
final_text = re.sub(newlines, lambda m: m.group(1) + is_paragraph_break(m) + m.group(2), text)
return final_text.strip()
def 解析PDF(file_name, llm_kwargs, plugin_kwargs, chatbot, history, system_prompt):
import time, glob, os, fitz
import tiktoken
print('begin analysis on:', file_name)
file_content, page_one = read_and_clean_pdf_text(file_name)
with fitz.open(file_name) as doc:
file_content = ""
for page in doc:
file_content += page.get_text()
file_content = clean_text(file_content)
# print(file_content)
split_number = 10000
split_group = (len(file_content)//split_number)+1
for i in range(0,split_group):
if i==0:
prefix = "接下来请你仔细分析下面的论文,学习里面的内容(专业术语、公式、数学概念).并且注意:由于论文内容较多,将分批次发送,每次发送完之后,你只需要回答“接受完成”"
i_say = prefix + f'文件名是{file_name},文章内容第{i+1}部分是 ```{file_content[i*split_number:(i+1)*split_number]}```'
i_say_show_user = f'文件名是:\n{file_name},\n由于论文内容过长,将分批请求(共{len(file_content)}字符,将分为{split_group}批,每批{split_number}字符)。\n当前发送{i+1}/{split_group}部分'
elif i==split_group-1:
i_say = f'你只需要回答“所有论文接受完成,请进行下一步”。文章内容第{i+1}/{split_group}部分是 ```{file_content[i*split_number:]}```'
i_say_show_user = f'当前发送{i+1}/{split_group}部分'
else:
i_say = f'你只需要回答“接受完成”。文章内容第{i+1}/{split_group}部分是 ```{file_content[i*split_number:(i+1)*split_number]}```'
i_say_show_user = f'当前发送{i+1}/{split_group}部分'
chatbot.append((i_say_show_user, "[Local Message] waiting gpt response."))
gpt_say = yield from request_gpt_model_in_new_thread_with_ui_alive(i_say, i_say_show_user, llm_kwargs, chatbot, history=[], sys_prompt="") # 带超时倒计时
while "完成" not in gpt_say:
i_say = f'你只需要回答“接受完成”。文章内容第{i+1}/{split_group}部分是 ```{file_content[i*split_number:(i+1)*split_number]}```'
i_say_show_user = f'出现error,重新发送{i+1}/{split_group}部分'
gpt_say = yield from request_gpt_model_in_new_thread_with_ui_alive(i_say, i_say_show_user, llm_kwargs, chatbot, history=[], sys_prompt="") # 带超时倒计时
time.sleep(1)
chatbot[-1] = (i_say_show_user, gpt_say)
history.append(i_say_show_user); history.append(gpt_say)
yield from update_ui(chatbot=chatbot, history=history) # 刷新界面
time.sleep(2)
############################## <第零步,从摘要中提取高价值信息,放到history中> ##################################
# 递归地切割PDF文件,每一块尽量是完整的一个section,比如introduction,experiment等,必要时再进行切割
# 的长度必须小于 2500 个 Token
TOKEN_LIMIT_PER_FRAGMENT = 2500
i_say = f'接下来,请你扮演一名专业的学术教授,利用你的所有知识并且结合这篇文章,回答我的问题。请牢记1.直到我说“退出”,你才能结束任务;2.所有问题需要紧密围绕文章内容;3.如果有公式,请使用tex渲染)'
chatbot.append((i_say, "[Local Message] waiting gpt response."))
yield from update_ui(chatbot=chatbot, history=history) # 刷新界面
from .crazy_utils import breakdown_txt_to_satisfy_token_limit_for_pdf
from toolbox import get_conf
enc = tiktoken.encoding_for_model(*get_conf('LLM_MODEL'))
def get_token_num(txt): return len(enc.encode(txt))
paper_fragments = breakdown_txt_to_satisfy_token_limit_for_pdf(
txt=file_content, get_token_fn=get_token_num, limit=TOKEN_LIMIT_PER_FRAGMENT)
page_one_fragments = breakdown_txt_to_satisfy_token_limit_for_pdf(
txt=str(page_one), get_token_fn=get_token_num, limit=TOKEN_LIMIT_PER_FRAGMENT//4)
# 为了更好的效果,我们剥离Introduction之后的部分如果有
paper_meta = page_one_fragments[0].split('introduction')[0].split('Introduction')[0].split('INTRODUCTION')[0]
# ** gpt request **
gpt_say = yield from request_gpt_model_in_new_thread_with_ui_alive(i_say, i_say, llm_kwargs, chatbot, history=history, sys_prompt="") # 带超时倒计时
chatbot[-1] = (i_say, gpt_say)
history.append(i_say); history.append(gpt_say)
yield from update_ui(chatbot=chatbot, history=history) # 刷新界面
############################## <第一步,从摘要中提取高价值信息,放到history中> ##################################
final_results = []
final_results.append(paper_meta)
############################## <第二步,迭代地历遍整个文章,提取精炼信息> ##################################
i_say_show_user = f'首先你在英文语境下通读整篇论文。'; gpt_say = "[Local Message] 收到。" # 用户提示
chatbot.append([i_say_show_user, gpt_say]); yield from update_ui(chatbot=chatbot, history=[]) # 更新UI
@CatchException
def 理解PDF文档内容(txt, llm_kwargs, plugin_kwargs, chatbot, history, system_prompt, web_port):
import glob, os
iteration_results = []
last_iteration_result = paper_meta # 初始值是摘要
MAX_WORD_TOTAL = 4096
n_fragment = len(paper_fragments)
if n_fragment >= 20: print('文章极长,不能达到预期效果')
for i in range(n_fragment):
NUM_OF_WORD = MAX_WORD_TOTAL // n_fragment
i_say = f"Read this section, recapitulate the content of this section with less than {NUM_OF_WORD} words: {paper_fragments[i]}"
i_say_show_user = f"[{i+1}/{n_fragment}] Read this section, recapitulate the content of this section with less than {NUM_OF_WORD} words: {paper_fragments[i][:200]}"
gpt_say = yield from request_gpt_model_in_new_thread_with_ui_alive(i_say, i_say_show_user, # i_say=真正给chatgpt的提问, i_say_show_user=给用户看的提问
llm_kwargs, chatbot,
history=["The main idea of the previous section is?", last_iteration_result], # 迭代上一次的结果
sys_prompt="Extract the main idea of this section." # 提示
)
iteration_results.append(gpt_say)
last_iteration_result = gpt_say
# 基本信息:功能、贡献者
chatbot.append([
"函数插件功能?",
"理解PDF论文内容,并且将结合上下文内容,进行学术解答。函数插件贡献者: Hanzoe。"])
yield from update_ui(chatbot=chatbot, history=history) # 刷新界面
import tkinter as tk
from tkinter import filedialog
root = tk.Tk()
root.withdraw()
txt = filedialog.askopenfilename()
# 尝试导入依赖,如果缺少依赖,则给出安装建议
try:
import fitz
except:
report_execption(chatbot, history,
a = f"解析项目: {txt}",
b = f"导入软件依赖失败。使用该模块需要额外依赖,安装方法```pip install --upgrade pymupdf```。")
yield from update_ui(chatbot=chatbot, history=history) # 刷新界面
return
# 清空历史,以免输入溢出
history = []
# 开始正式执行任务
yield from 解析PDF(txt, llm_kwargs, plugin_kwargs, chatbot, history, system_prompt)
############################## <第三步,整理history> ##################################
final_results.extend(iteration_results)
final_results.append(f'接下来,你是一名专业的学术教授,利用以上信息,使用中文回答我的问题。')
# 接下来两句话只显示在界面上,不起实际作用
i_say_show_user = f'接下来,你是一名专业的学术教授,利用以上信息,使用中文回答我的问题。'; gpt_say = "[Local Message] 收到。"
chatbot.append([i_say_show_user, gpt_say])
############################## <第四步,设置一个token上限,防止回答时Token溢出> ##################################
from .crazy_utils import input_clipping
_, final_results = input_clipping("", final_results, max_token_limit=3200)
yield from update_ui(chatbot=chatbot, history=final_results) # 注意这里的历史记录被替代了
@CatchException
@@ -146,7 +71,7 @@ def 理解PDF文档内容标准文件输入(txt, llm_kwargs, plugin_kwargs, chat
# 基本信息:功能、贡献者
chatbot.append([
"函数插件功能?",
"理解PDF论文内容,并且将结合上下文内容,进行学术解答。函数插件贡献者: Hanzoe"])
"理解PDF论文内容,并且将结合上下文内容,进行学术解答。函数插件贡献者: Hanzoe, binary-husky"])
yield from update_ui(chatbot=chatbot, history=history) # 刷新界面
# 尝试导入依赖,如果缺少依赖,则给出安装建议

查看文件

@@ -7,7 +7,7 @@ def 生成函数注释(file_manifest, project_folder, llm_kwargs, plugin_kwargs,
import time, os
print('begin analysis on:', file_manifest)
for index, fp in enumerate(file_manifest):
with open(fp, 'r', encoding='utf-8') as f:
with open(fp, 'r', encoding='utf-8', errors='replace') as f:
file_content = f.read()
i_say = f'请对下面的程序文件做一个概述,并对文件中的所有函数生成注释,使用markdown表格输出结果,文件名是{os.path.relpath(fp, project_folder)},文件内容是 ```{file_content}```'

查看文件

@@ -14,7 +14,7 @@ def 解析源代码新(file_manifest, project_folder, llm_kwargs, plugin_kwargs,
############################## <第一步,逐个文件分析,多线程> ##################################
for index, fp in enumerate(file_manifest):
with open(fp, 'r', encoding='utf-8') as f:
with open(fp, 'r', encoding='utf-8', errors='replace') as f:
file_content = f.read()
prefix = "接下来请你逐文件分析下面的工程" if index==0 else ""
i_say = prefix + f'请对下面的程序文件做一个概述文件名是{os.path.relpath(fp, project_folder)},文件代码是 ```{file_content}```'
@@ -225,9 +225,34 @@ def 解析一个Golang项目(txt, llm_kwargs, plugin_kwargs, chatbot, history, s
report_execption(chatbot, history, a=f"解析项目: {txt}", b=f"找不到本地项目或无权访问: {txt}")
yield from update_ui(chatbot=chatbot, history=history) # 刷新界面
return
file_manifest = [f for f in glob.glob(f'{project_folder}/**/*.go', recursive=True)]
file_manifest = [f for f in glob.glob(f'{project_folder}/**/*.go', recursive=True)] + \
[f for f in glob.glob(f'{project_folder}/**/go.mod', recursive=True)] + \
[f for f in glob.glob(f'{project_folder}/**/go.sum', recursive=True)] + \
[f for f in glob.glob(f'{project_folder}/**/go.work', recursive=True)]
if len(file_manifest) == 0:
report_execption(chatbot, history, a=f"解析项目: {txt}", b=f"找不到任何golang文件: {txt}")
yield from update_ui(chatbot=chatbot, history=history) # 刷新界面
return
yield from 解析源代码新(file_manifest, project_folder, llm_kwargs, plugin_kwargs, chatbot, history, system_prompt)
@CatchException
def 解析一个Lua项目(txt, llm_kwargs, plugin_kwargs, chatbot, history, system_prompt, web_port):
history = [] # 清空历史,以免输入溢出
import glob, os
if os.path.exists(txt):
project_folder = txt
else:
if txt == "": txt = '空空如也的输入栏'
report_execption(chatbot, history, a = f"解析项目: {txt}", b = f"找不到本地项目或无权访问: {txt}")
yield from update_ui(chatbot=chatbot, history=history) # 刷新界面
return
file_manifest = [f for f in glob.glob(f'{project_folder}/**/*.lua', recursive=True)] + \
[f for f in glob.glob(f'{project_folder}/**/*.xml', recursive=True)] + \
[f for f in glob.glob(f'{project_folder}/**/*.json', recursive=True)] + \
[f for f in glob.glob(f'{project_folder}/**/*.toml', recursive=True)]
if len(file_manifest) == 0:
report_execption(chatbot, history, a = f"解析项目: {txt}", b = f"找不到任何lua文件: {txt}")
yield from update_ui(chatbot=chatbot, history=history) # 刷新界面
return
yield from 解析源代码新(file_manifest, project_folder, llm_kwargs, plugin_kwargs, chatbot, history, system_prompt)

查看文件

@@ -8,7 +8,7 @@ def 解析Paper(file_manifest, project_folder, llm_kwargs, plugin_kwargs, chatbo
import time, glob, os
print('begin analysis on:', file_manifest)
for index, fp in enumerate(file_manifest):
with open(fp, 'r', encoding='utf-8') as f:
with open(fp, 'r', encoding='utf-8', errors='replace') as f:
file_content = f.read()
prefix = "接下来请你逐文件分析下面的论文文件,概括其内容" if index==0 else ""

查看文件

@@ -6,7 +6,7 @@ def 高阶功能模板函数(txt, llm_kwargs, plugin_kwargs, chatbot, history, s
"""
txt 输入栏用户输入的文本,例如需要翻译的一段话,再例如一个包含了待处理文件的路径
llm_kwargs gpt模型参数,如温度和top_p等,一般原样传递下去就行
plugin_kwargs 插件模型的参数,如温度和top_p等,一般原样传递下去就行
plugin_kwargs 插件模型的参数,暂时没有用武之地
chatbot 聊天显示框的句柄,用于显示给用户
history 聊天历史,前情提要
system_prompt 给gpt的静默提醒

查看文件

@@ -1,4 +1,5 @@
gradio>=3.23
gradio>=3.25.0
tiktoken>=0.3.3
requests[socks]
transformers
python-markdown-math
@@ -7,7 +8,6 @@ latex2mathml
python-docx
mdtex2html
colorama
tiktoken
Markdown
pygments
pymupdf

查看文件

@@ -1,5 +1,5 @@
{
"version": 2.67,
"version": 2.68,
"show_feature": true,
"new_feature": "现可通过输入区更新临时api-key <-> 增强多线程稳定性涉及代码解析、PDF翻译、自译解等 <-> 修复Token计数错误解决PDF翻译的分割不合理的问题 <-> 如果一键更新失败,可前往github手动更新"
"new_feature": "改善理解pdfchatpdf功能 <-> 如果一键更新失败,可前往github手动更新"
}