fufa/chatgpt
0
1# 借鉴了 https://github.com/GaiZhenbiao/ChuanhuChatGPT 项目2 3"""4 该文件中主要包含三个函数5 6 不具备多线程能力的函数：7 1. predict: 正常对话时使用，具备完备的交互功能，不可多线程8 9 具备多线程调用能力的函数10 2. predict_no_ui：高级实验性功能模块调用，不会实时显示在界面上，参数简单，可以多线程并行，方便实现复杂的功能逻辑11 3. predict_no_ui_long_connection：在实验过程中发现调用predict_no_ui处理长文档时，和openai的连接容易断掉，这个函数用stream的方式解决这个问题，同样支持多线程12"""13 14import json15import gradio as gr16import logging17import traceback18import requests19import importlib20 21# config_private.py放自己的秘密如API和代理网址22# 读取时首先看是否存在私密的config_private配置文件（不受git管控），如果有，则覆盖原config文件23from toolbox import get_conf24proxies, API_URL, API_KEY, TIMEOUT_SECONDS, MAX_RETRY, LLM_MODEL = \25 get_conf('proxies', 'API_URL', 'API_KEY', 'TIMEOUT_SECONDS', 'MAX_RETRY', 'LLM_MODEL')26 27timeout_bot_msg = '[Local Message] Request timeout. Network error. Please check proxy settings in config.py.' + \28 '网络错误，检查代理服务器是否可用，以及代理设置的格式是否正确，格式须是[协议]://[地址]:[端口]，缺一不可。'29 30def get_full_error(chunk, stream_response):31 """32 获取完整的从Openai返回的报错33 """34 while True:35 try:36 chunk += next(stream_response)37 except:38 break39 return chunk40 41def predict_no_ui(inputs, top_p, temperature, history=[], sys_prompt=""):42 """43 发送至chatGPT，等待回复，一次性完成，不显示中间过程。44 predict函数的简化版。45 用于payload比较大的情况，或者用于实现多线、带嵌套的复杂功能。46 47 inputs 是本次问询的输入48 top_p, temperature是chatGPT的内部调优参数49 history 是之前的对话列表50 （注意无论是inputs还是history，内容太长了都会触发token数量溢出的错误，然后raise ConnectionAbortedError）51 """52 headers, payload = generate_payload(inputs, top_p, temperature, history, system_prompt=sys_prompt, stream=False)53 54 retry = 055 while True:56 try:57 # make a POST request to the API endpoint, stream=False58 response = requests.post(API_URL, headers=headers, proxies=proxies,59 json=payload, stream=False, timeout=TIMEOUT_SECONDS*2); break60 except requests.exceptions.ReadTimeout as e:61 retry += 162 traceback.print_exc()63 if retry > MAX_RETRY: raise TimeoutError64 if MAX_RETRY!=0: print(f'请求超时，正在重试 ({retry}/{MAX_RETRY}) ……')65 66 try:67 result = json.loads(response.text)["choices"][0]["message"]["content"]68 return result69 except Exception as e:70 if "choices" not in response.text: print(response.text)71 raise ConnectionAbortedError("Json解析不合常规，可能是文本过长" + response.text)72 73 74def predict_no_ui_long_connection(inputs, top_p, temperature, history=[], sys_prompt=""):75 """76 发送至chatGPT，等待回复，一次性完成，不显示中间过程。但内部用stream的方法避免有人中途掐网线。77 """78 headers, payload = generate_payload(inputs, top_p, temperature, history, system_prompt=sys_prompt, stream=True)79 80 retry = 081 while True:82 try:83 # make a POST request to the API endpoint, stream=False84 response = requests.post(API_URL, headers=headers, proxies=proxies,85 json=payload, stream=True, timeout=TIMEOUT_SECONDS); break86 except requests.exceptions.ReadTimeout as e:87 retry += 188 traceback.print_exc()89 if retry > MAX_RETRY: raise TimeoutError90 if MAX_RETRY!=0: print(f'请求超时，正在重试 ({retry}/{MAX_RETRY}) ……')91 92 stream_response = response.iter_lines()93 result = ''94 while True:95 try: chunk = next(stream_response).decode()96 except StopIteration: break97 if len(chunk)==0: continue98 if not chunk.startswith('data:'): 99 chunk = get_full_error(chunk.encode('utf8'), stream_response)100 raise ConnectionAbortedError("OpenAI拒绝了请求:" + chunk.decode())101 delta = json.loads(chunk.lstrip('data:'))['choices'][0]["delta"]102 if len(delta) == 0: break103 if "role" in delta: continue104 if "content" in delta: result += delta["content"]; print(delta["content"], end='')105 else: raise RuntimeError("意外Json结构："+delta)106 return result107 108 109def predict(inputs, top_p, temperature, chatbot=[], history=[], system_prompt='', 110 stream = True, additional_fn=None):111 """112 发送至chatGPT，流式获取输出。113 用于基础的对话功能。114 inputs 是本次问询的输入115 top_p, temperature是chatGPT的内部调优参数116 history 是之前的对话列表（注意无论是inputs还是history，内容太长了都会触发token数量溢出的错误）117 chatbot 为WebUI中显示的对话列表，修改它，然后yeild出去，可以直接修改对话界面内容118 additional_fn代表点击的哪个按钮，按钮见functional.py119 """120 if additional_fn is not None:121 import functional122 importlib.reload(functional) # 热更新prompt123 functional = functional.get_functionals()124 if "PreProcess" in functional[additional_fn]: inputs = functional[additional_fn]["PreProcess"](inputs) # 获取预处理函数（如果有的话）125 inputs = functional[additional_fn]["Prefix"] + inputs + functional[additional_fn]["Suffix"]126 127 if stream:128 raw_input = inputs129 logging.info(f'[raw_input] {raw_input}')130 chatbot.append((inputs, ""))131 yield chatbot, history, "等待响应"132 133 headers, payload = generate_payload(inputs, top_p, temperature, history, system_prompt, stream)134 history.append(inputs); history.append(" ")135 136 retry = 0137 while True:138 try:139 # make a POST request to the API endpoint, stream=True140 response = requests.post(API_URL, headers=headers, proxies=proxies,141 json=payload, stream=True, timeout=TIMEOUT_SECONDS);break142 except:143 retry += 1144 chatbot[-1] = ((chatbot[-1][0], timeout_bot_msg))145 retry_msg = f"，正在重试 ({retry}/{MAX_RETRY}) ……" if MAX_RETRY > 0 else ""146 yield chatbot, history, "请求超时"+retry_msg147 if retry > MAX_RETRY: raise TimeoutError148 149 gpt_replying_buffer = ""150 151 is_head_of_the_stream = True152 if stream:153 stream_response = response.iter_lines()154 while True:155 chunk = next(stream_response)156 # print(chunk.decode()[6:])157 if is_head_of_the_stream:158 # 数据流的第一帧不携带content159 is_head_of_the_stream = False; continue160 161 if chunk:162 try:163 if len(json.loads(chunk.decode()[6:])['choices'][0]["delta"]) == 0:164 # 判定为数据流的结束，gpt_replying_buffer也写完了165 logging.info(f'[response] {gpt_replying_buffer}')166 break167 # 处理数据流的主体168 chunkjson = json.loads(chunk.decode()[6:])169 status_text = f"finish_reason: {chunkjson['choices'][0]['finish_reason']}"170 # 如果这里抛出异常，一般是文本过长，详情见get_full_error的输出171 gpt_replying_buffer = gpt_replying_buffer + json.loads(chunk.decode()[6:])['choices'][0]["delta"]["content"]172 history[-1] = gpt_replying_buffer173 chatbot[-1] = (history[-2], history[-1])174 yield chatbot, history, status_text175 176 except Exception as e:177 traceback.print_exc()178 yield chatbot, history, "Json解析不合常规"179 chunk = get_full_error(chunk, stream_response)180 error_msg = chunk.decode()181 if "reduce the length" in error_msg:182 chatbot[-1] = (chatbot[-1][0], "[Local Message] Input (or history) is too long, please reduce input or clear history by refreshing this page.")183 history = []184 elif "Incorrect API key" in error_msg:185 chatbot[-1] = (chatbot[-1][0], "[Local Message] Incorrect API key provided.")186 else:187 from toolbox import regular_txt_to_markdown188 tb_str = regular_txt_to_markdown(traceback.format_exc())189 chatbot[-1] = (chatbot[-1][0], f"[Local Message] Json Error \n\n {tb_str} \n\n {regular_txt_to_markdown(chunk.decode()[4:])}")190 yield chatbot, history, "Json解析不合常规" + error_msg191 return192 193def generate_payload(inputs, top_p, temperature, history, system_prompt, stream):194 """195 整合所有信息，选择LLM模型，生成http请求，为发送请求做准备196 """197 headers = {198 "Content-Type": "application/json",199 "Authorization": f"Bearer {API_KEY}"200 }201 202 conversation_cnt = len(history) // 2203 204 messages = [{"role": "system", "content": system_prompt}]205 if conversation_cnt:206 for index in range(0, 2*conversation_cnt, 2):207 what_i_have_asked = {}208 what_i_have_asked["role"] = "user"209 what_i_have_asked["content"] = history[index]210 what_gpt_answer = {}211 what_gpt_answer["role"] = "assistant"212 what_gpt_answer["content"] = history[index+1]213 if what_i_have_asked["content"] != "":214 if what_gpt_answer["content"] == "": continue215 if what_gpt_answer["content"] == timeout_bot_msg: continue216 messages.append(what_i_have_asked)217 messages.append(what_gpt_answer)218 else:219 messages[-1]['content'] = what_gpt_answer['content']220 221 what_i_ask_now = {}222 what_i_ask_now["role"] = "user"223 what_i_ask_now["content"] = inputs224 messages.append(what_i_ask_now)225 226 payload = {227 "model": LLM_MODEL,228 "messages": messages, 229 "temperature": temperature, # 1.0,230 "top_p": top_p, # 1.0,231 "n": 1,232 "stream": stream,233 "presence_penalty": 0,234 "frequency_penalty": 0,235 }236 237 print(f" {LLM_MODEL} : {conversation_cnt} : {inputs}")238 return headers,payload239 240 241