支持gpt-4-vision-preview

request_llms/bridge_all.py

@@ -16,6 +16,9 @@ from toolbox import get_conf, trimmed_format_exc
 from .bridge_chatgpt import predict_no_ui_long_connection as chatgpt_noui
 from .bridge_chatgpt import predict as chatgpt_ui
 
+from .bridge_chatgpt_vision import predict_no_ui_long_connection as chatgpt_vision_noui
+from .bridge_chatgpt_vision import predict as chatgpt_vision_ui
+
 from .bridge_chatglm import predict_no_ui_long_connection as chatglm_noui
 from .bridge_chatglm import predict as chatglm_ui
 
@@ -162,6 +165,16 @@ model_info = {
         "token_cnt": get_token_num_gpt4,
     },
     
+    "gpt-4-vision-preview": {
+        "fn_with_ui": chatgpt_vision_ui,
+        "fn_without_ui": chatgpt_vision_noui,
+        "endpoint": openai_endpoint,
+        "max_token": 4096,
+        "tokenizer": tokenizer_gpt4,
+        "token_cnt": get_token_num_gpt4,
+    },
+
+
     # azure openai
     "azure-gpt-3.5":{
         "fn_with_ui": chatgpt_ui,

request_llms/bridge_chatgpt_vision.py

@@ -0,0 +1,329 @@
+"""
+    该文件中主要包含三个函数
+
+    不具备多线程能力的函数：
+    1. predict: 正常对话时使用，具备完备的交互功能，不可多线程
+
+    具备多线程调用能力的函数
+    2. predict_no_ui_long_connection：支持多线程
+"""
+
+import json
+import time
+import logging
+import requests
+import base64
+import os
+import glob
+
+from toolbox import get_conf, update_ui, is_any_api_key, select_api_key, what_keys, clip_history, trimmed_format_exc, is_the_upload_folder, update_ui_lastest_msg, get_max_token
+proxies, TIMEOUT_SECONDS, MAX_RETRY, API_ORG, AZURE_CFG_ARRAY = \
+    get_conf('proxies', 'TIMEOUT_SECONDS', 'MAX_RETRY', 'API_ORG', 'AZURE_CFG_ARRAY')
+
+timeout_bot_msg = '[Local Message] Request timeout. Network error. Please check proxy settings in config.py.' + \
+                  '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。'
+
+def have_any_recent_upload_image_files(chatbot):
+    _5min = 5 * 60
+    if chatbot is None: return False, None    # chatbot is None
+    most_recent_uploaded = chatbot._cookies.get("most_recent_uploaded", None)
+    if not most_recent_uploaded: return False, None   # most_recent_uploaded is None
+    if time.time() - most_recent_uploaded["time"] < _5min: 
+        most_recent_uploaded = chatbot._cookies.get("most_recent_uploaded", None)
+        path = most_recent_uploaded['path']
+        file_manifest = [f for f in glob.glob(f'{path}/**/*.jpg', recursive=True)]
+        file_manifest += [f for f in glob.glob(f'{path}/**/*.jpeg', recursive=True)]
+        file_manifest += [f for f in glob.glob(f'{path}/**/*.png', recursive=True)]
+        if len(file_manifest) == 0: return False, None
+        return True, file_manifest # most_recent_uploaded is new
+    else: 
+        return False, None  # most_recent_uploaded is too old
+
+def report_invalid_key(key):
+    if get_conf("BLOCK_INVALID_APIKEY"): 
+        # 实验性功能，自动检测并屏蔽失效的KEY，请勿使用
+        from request_llms.key_manager import ApiKeyManager
+        api_key = ApiKeyManager().add_key_to_blacklist(key)
+
+def get_full_error(chunk, stream_response):
+    """
+        获取完整的从Openai返回的报错
+    """
+    while True:
+        try:
+            chunk += next(stream_response)
+        except:
+            break
+    return chunk
+
+def decode_chunk(chunk):
+    # 提前读取一些信息 （用于判断异常）
+    chunk_decoded = chunk.decode()
+    chunkjson = None
+    has_choices = False
+    choice_valid = False
+    has_content = False
+    has_role = False
+    try: 
+        chunkjson = json.loads(chunk_decoded[6:])
+        has_choices = 'choices' in chunkjson
+        if has_choices: choice_valid = (len(chunkjson['choices']) > 0)
+        if has_choices and choice_valid: has_content = "content" in chunkjson['choices'][0]["delta"]
+        if has_choices and choice_valid: has_role = "role" in chunkjson['choices'][0]["delta"]
+    except: 
+        pass
+    return chunk_decoded, chunkjson, has_choices, choice_valid, has_content, has_role
+
+from functools import lru_cache
+@lru_cache(maxsize=32)
+def verify_endpoint(endpoint):
+    """
+        检查endpoint是否可用
+    """
+    return endpoint
+
+def predict_no_ui_long_connection(inputs, llm_kwargs, history=[], sys_prompt="", observe_window=None, console_slience=False):
+    raise NotImplementedError
+
+
+def predict(inputs, llm_kwargs, plugin_kwargs, chatbot, history=[], system_prompt='', stream = True, additional_fn=None):
+
+    have_recent_file, image_paths = have_any_recent_upload_image_files(chatbot)
+
+    if is_any_api_key(inputs):
+        chatbot._cookies['api_key'] = inputs
+        chatbot.append(("输入已识别为openai的api_key", what_keys(inputs)))
+        yield from update_ui(chatbot=chatbot, history=history, msg="api_key已导入") # 刷新界面
+        return
+    elif not is_any_api_key(chatbot._cookies['api_key']):
+        chatbot.append((inputs, "缺少api_key。\n\n1. 临时解决方案：直接在输入区键入api_key，然后回车提交。\n\n2. 长效解决方案：在config.py中配置。"))
+        yield from update_ui(chatbot=chatbot, history=history, msg="缺少api_key") # 刷新界面
+        return
+    if not have_recent_file:
+        chatbot.append((inputs, "没有检测到任何近期上传的图像文件，请上传jpg格式的图片，此外，请注意拓展名需要小写"))
+        yield from update_ui(chatbot=chatbot, history=history, msg="等待图片") # 刷新界面
+        return
+    if os.path.exists(inputs):
+        chatbot.append((inputs, "已经接收到您上传的文件，您不需要再重复强调该文件的路径了，请直接输入您的问题。"))
+        yield from update_ui(chatbot=chatbot, history=history, msg="等待指令") # 刷新界面
+        return
+
+
+    user_input = inputs
+    if additional_fn is not None:
+        from core_functional import handle_core_functionality
+        inputs, history = handle_core_functionality(additional_fn, inputs, history, chatbot)
+
+    raw_input = inputs
+    logging.info(f'[raw_input] {raw_input}')
+    def make_media_input(inputs, image_paths): 
+        for image_path in image_paths:
+            inputs = inputs + f'<br/><br/><div align="center"><img src="file={os.path.abspath(image_path)}"></div>'
+        return inputs
+    chatbot.append((make_media_input(inputs, image_paths), ""))
+    yield from update_ui(chatbot=chatbot, history=history, msg="等待响应") # 刷新界面
+
+    # check mis-behavior
+    if is_the_upload_folder(user_input):
+        chatbot[-1] = (inputs, f"[Local Message] 检测到操作错误！当您上传文档之后，需点击“**函数插件区**”按钮进行处理，请勿点击“提交”按钮或者“基础功能区”按钮。")
+        yield from update_ui(chatbot=chatbot, history=history, msg="正常") # 刷新界面
+        time.sleep(2)
+
+    try:
+        headers, payload, api_key = generate_payload(inputs, llm_kwargs, history, system_prompt, image_paths)
+    except RuntimeError as e:
+        chatbot[-1] = (inputs, f"您提供的api-key不满足要求，不包含任何可用于{llm_kwargs['llm_model']}的api-key。您可能选择了错误的模型或请求源。")
+        yield from update_ui(chatbot=chatbot, history=history, msg="api-key不满足要求") # 刷新界面
+        return
+        
+    # 检查endpoint是否合法
+    try:
+        from .bridge_all import model_info
+        endpoint = verify_endpoint(model_info[llm_kwargs['llm_model']]['endpoint'])
+    except:
+        tb_str = '```\n' + trimmed_format_exc() + '```'
+        chatbot[-1] = (inputs, tb_str)
+        yield from update_ui(chatbot=chatbot, history=history, msg="Endpoint不满足要求") # 刷新界面
+        return
+
+    history.append(make_media_input(inputs, image_paths))
+    history.append("")
+
+    retry = 0
+    while True:
+        try:
+            # make a POST request to the API endpoint, stream=True
+            response = requests.post(endpoint, headers=headers, proxies=proxies,
+                                    json=payload, stream=True, timeout=TIMEOUT_SECONDS);break
+        except:
+            retry += 1
+            chatbot[-1] = ((chatbot[-1][0], timeout_bot_msg))
+            retry_msg = f"，正在重试 ({retry}/{MAX_RETRY}) ……" if MAX_RETRY > 0 else ""
+            yield from update_ui(chatbot=chatbot, history=history, msg="请求超时"+retry_msg) # 刷新界面
+            if retry > MAX_RETRY: raise TimeoutError
+
+    gpt_replying_buffer = ""
+    
+    is_head_of_the_stream = True
+    if stream:
+        stream_response =  response.iter_lines()
+        while True:
+            try:
+                chunk = next(stream_response)
+            except StopIteration:
+                # 非OpenAI官方接口的出现这样的报错，OpenAI和API2D不会走这里
+                chunk_decoded = chunk.decode()
+                error_msg = chunk_decoded
+                # 首先排除一个one-api没有done数据包的第三方Bug情形
+                if len(gpt_replying_buffer.strip()) > 0 and len(error_msg) == 0: 
+                    yield from update_ui(chatbot=chatbot, history=history, msg="检测到有缺陷的非OpenAI官方接口，建议选择更稳定的接口。")
+                    break
+                # 其他情况，直接返回报错
+                chatbot, history = handle_error(inputs, llm_kwargs, chatbot, history, chunk_decoded, error_msg, api_key)
+                yield from update_ui(chatbot=chatbot, history=history, msg="非OpenAI官方接口返回了错误:" + chunk.decode()) # 刷新界面
+                return
+            
+            # 提前读取一些信息 （用于判断异常）
+            chunk_decoded, chunkjson, has_choices, choice_valid, has_content, has_role = decode_chunk(chunk)
+
+            if is_head_of_the_stream and (r'"object":"error"' not in chunk_decoded) and (r"content" not in chunk_decoded):
+                # 数据流的第一帧不携带content
+                is_head_of_the_stream = False; continue
+            
+            if chunk:
+                try:
+                    if has_choices and not choice_valid:
+                        # 一些垃圾第三方接口的出现这样的错误
+                        continue
+                    # 前者是API2D的结束条件，后者是OPENAI的结束条件
+                    if ('data: [DONE]' in chunk_decoded) or (len(chunkjson['choices'][0]["delta"]) == 0):
+                        # 判定为数据流的结束，gpt_replying_buffer也写完了
+                        lastmsg = chatbot[-1][-1] + f"<br/><br/>{llm_kwargs['llm_model']}调用结束，该模型不具备上下文对话能力，如需追问，请及时切换模型。"
+                        yield from update_ui_lastest_msg(lastmsg, chatbot, history, delay=1)
+                        logging.info(f'[response] {gpt_replying_buffer}')
+                        break
+                    # 处理数据流的主体
+                    status_text = f"finish_reason: {chunkjson['choices'][0].get('finish_reason', 'null')}"
+                    # 如果这里抛出异常，一般是文本过长，详情见get_full_error的输出
+                    if has_content:
+                        # 正常情况
+                        gpt_replying_buffer = gpt_replying_buffer + chunkjson['choices'][0]["delta"]["content"]
+                    elif has_role:
+                        # 一些第三方接口的出现这样的错误，兼容一下吧
+                        continue
+                    else:
+                        # 一些垃圾第三方接口的出现这样的错误
+                        gpt_replying_buffer = gpt_replying_buffer + chunkjson['choices'][0]["delta"]["content"]
+
+                    history[-1] = gpt_replying_buffer
+                    chatbot[-1] = (history[-2], history[-1])
+                    yield from update_ui(chatbot=chatbot, history=history, msg=status_text) # 刷新界面
+                except Exception as e:
+                    yield from update_ui(chatbot=chatbot, history=history, msg="Json解析不合常规") # 刷新界面
+                    chunk = get_full_error(chunk, stream_response)
+                    chunk_decoded = chunk.decode()
+                    error_msg = chunk_decoded
+                    chatbot, history = handle_error(inputs, llm_kwargs, chatbot, history, chunk_decoded, error_msg, api_key)
+                    yield from update_ui(chatbot=chatbot, history=history, msg="Json异常" + error_msg) # 刷新界面
+                    print(error_msg)
+                    return
+
+def handle_error(inputs, llm_kwargs, chatbot, history, chunk_decoded, error_msg, api_key=""):
+    from .bridge_all import model_info
+    openai_website = ' 请登录OpenAI查看详情 https://platform.openai.com/signup'
+    if "reduce the length" in error_msg:
+        if len(history) >= 2: history[-1] = ""; history[-2] = "" # 清除当前溢出的输入：history[-2] 是本次输入, history[-1] 是本次输出
+        history = clip_history(inputs=inputs, history=history, tokenizer=model_info[llm_kwargs['llm_model']]['tokenizer'], 
+                                               max_token_limit=(model_info[llm_kwargs['llm_model']]['max_token'])) # history至少释放二分之一
+        chatbot[-1] = (chatbot[-1][0], "[Local Message] Reduce the length. 本次输入过长, 或历史数据过长. 历史缓存数据已部分释放, 您可以请再次尝试. (若再次失败则更可能是因为输入过长.)")
+    elif "does not exist" in error_msg:
+        chatbot[-1] = (chatbot[-1][0], f"[Local Message] Model {llm_kwargs['llm_model']} does not exist. 模型不存在, 或者您没有获得体验资格.")
+    elif "Incorrect API key" in error_msg:
+        chatbot[-1] = (chatbot[-1][0], "[Local Message] Incorrect API key. OpenAI以提供了不正确的API_KEY为由, 拒绝服务. " + openai_website); report_invalid_key(api_key)
+    elif "exceeded your current quota" in error_msg:
+        chatbot[-1] = (chatbot[-1][0], "[Local Message] You exceeded your current quota. OpenAI以账户额度不足为由, 拒绝服务." + openai_website); report_invalid_key(api_key)
+    elif "account is not active" in error_msg:
+        chatbot[-1] = (chatbot[-1][0], "[Local Message] Your account is not active. OpenAI以账户失效为由, 拒绝服务." + openai_website); report_invalid_key(api_key)
+    elif "associated with a deactivated account" in error_msg:
+        chatbot[-1] = (chatbot[-1][0], "[Local Message] You are associated with a deactivated account. OpenAI以账户失效为由, 拒绝服务." + openai_website); report_invalid_key(api_key)
+    elif "API key has been deactivated" in error_msg:
+        chatbot[-1] = (chatbot[-1][0], "[Local Message] API key has been deactivated. OpenAI以账户失效为由, 拒绝服务." + openai_website); report_invalid_key(api_key)
+    elif "bad forward key" in error_msg:
+        chatbot[-1] = (chatbot[-1][0], "[Local Message] Bad forward key. API2D账户额度不足.")
+    elif "Not enough point" in error_msg:
+        chatbot[-1] = (chatbot[-1][0], "[Local Message] Not enough point. API2D账户点数不足.")
+    else:
+        from toolbox import regular_txt_to_markdown
+        tb_str = '```\n' + trimmed_format_exc() + '```'
+        chatbot[-1] = (chatbot[-1][0], f"[Local Message] 异常 \n\n{tb_str} \n\n{regular_txt_to_markdown(chunk_decoded)}")
+    return chatbot, history
+
+# Function to encode the image
+def encode_image(image_path):
+    with open(image_path, "rb") as image_file:
+        return base64.b64encode(image_file.read()).decode('utf-8')
+
+def generate_payload(inputs, llm_kwargs, history, system_prompt, image_paths):
+    """
+    整合所有信息，选择LLM模型，生成http请求，为发送请求做准备
+    """
+    if not is_any_api_key(llm_kwargs['api_key']):
+        raise AssertionError("你提供了错误的API_KEY。\n\n1. 临时解决方案：直接在输入区键入api_key，然后回车提交。\n\n2. 长效解决方案：在config.py中配置。")
+
+    api_key = select_api_key(llm_kwargs['api_key'], llm_kwargs['llm_model'])
+
+    headers = {
+        "Content-Type": "application/json",
+        "Authorization": f"Bearer {api_key}"
+    }
+    if API_ORG.startswith('org-'): headers.update({"OpenAI-Organization": API_ORG})
+    if llm_kwargs['llm_model'].startswith('azure-'): 
+        headers.update({"api-key": api_key})
+        if llm_kwargs['llm_model'] in AZURE_CFG_ARRAY.keys():
+            azure_api_key_unshared = AZURE_CFG_ARRAY[llm_kwargs['llm_model']]["AZURE_API_KEY"]
+            headers.update({"api-key": azure_api_key_unshared})
+
+    base64_images = []
+    for image_path in image_paths:
+        base64_images.append(encode_image(image_path))
+
+    messages = []
+    what_i_ask_now = {}
+    what_i_ask_now["role"] = "user"
+    what_i_ask_now["content"] = []
+    what_i_ask_now["content"].append({
+        "type": "text",
+        "text": inputs
+    })
+
+    for image_path, base64_image in zip(image_paths, base64_images):
+        what_i_ask_now["content"].append({
+            "type": "image_url",
+            "image_url": {
+                "url": f"data:image/jpeg;base64,{base64_image}"
+            }
+        })
+
+    messages.append(what_i_ask_now)
+    model = llm_kwargs['llm_model']
+    if llm_kwargs['llm_model'].startswith('api2d-'):
+        model = llm_kwargs['llm_model'][len('api2d-'):]
+
+    payload = {
+        "model": model,
+        "messages": messages, 
+        "temperature": llm_kwargs['temperature'],   # 1.0,
+        "top_p": llm_kwargs['top_p'],               # 1.0,
+        "n": 1,
+        "stream": True,
+        "max_tokens": get_max_token(llm_kwargs),
+        "presence_penalty": 0,
+        "frequency_penalty": 0,
+    }
+    try:
+        print(f" {llm_kwargs['llm_model']} : {inputs[:100]} ..........")
+    except:
+        print('输入中可能存在乱码。')
+    return headers, payload, api_key
+
+

toolbox.py

@@ -279,9 +279,12 @@ def text_divide_paragraph(text):
     
     if '```' in text:
         # careful input
-        return pre + text + suf
+        return text
+    elif '</div>' in text:
+        # careful input
+        return text
     else:
-        # wtf input
+        # whatever input
         lines = text.split("\n")
         for i, line in enumerate(lines):
             lines[i] = lines[i].replace(" ", "&nbsp;")