Add Support for Gemini 1.5 Pro & Gemini 1.5 Flash (#1926)

* Add Support for Gemini 1.5 Pro & 1.5 Flash.

* Update bridge_all.py

fix a spelling error in comments.

* Add Support for Gemini 1.5 Pro & Gemini 1.5 Flash

config.py

@@ -36,7 +36,7 @@
                     "gpt-4o", "gpt-4o-mini", "gpt-4-turbo", "gpt-4-turbo-2024-04-09",
                     "gpt-3.5-turbo-1106", "gpt-3.5-turbo-16k", "gpt-3.5-turbo", "azure-gpt-3.5",
                     "gpt-4", "gpt-4-32k", "azure-gpt-4", "glm-4", "glm-4v", "glm-3-turbo",
-                    "gemini-pro", "chatglm3"
+                    "gemini-1.5-pro", "chatglm3"
                     ]
 # --- --- --- ---
 # P.S. 其他可用的模型还包括
@@ -50,6 +50,7 @@
 #   "claude-3-haiku-20240307","claude-3-sonnet-20240229","claude-3-opus-20240229", "claude-2.1", "claude-instant-1.2",
 #   "moss", "llama2", "chatglm_onnx", "internlm", "jittorllms_pangualpha", "jittorllms_llama",
 #   "deepseek-chat" ,"deepseek-coder",
+#   "gemini-1.5-flash",
 #   "yi-34b-chat-0205","yi-34b-chat-200k","yi-large","yi-medium","yi-spark","yi-large-turbo","yi-large-preview",
 # ]
 # --- --- --- ---

request_llms/bridge_all.py

@@ -407,22 +407,46 @@ def decode(self, *args, **kwargs):
         "tokenizer": tokenizer_gpt35,
         "token_cnt": get_token_num_gpt35,
     },
+    # Gemini
+    # Note: now gemini-pro is an alias of gemini-1.0-pro.
+    # Warning: gemini-pro-vision has been deprecated.
+    # Support for gemini-pro-vision has been removed.
     "gemini-pro": {
         "fn_with_ui": genai_ui,
         "fn_without_ui": genai_noui,
         "endpoint": gemini_endpoint,
+        "has_multimodal_capacity": False,
         "max_token": 1024 * 32,
         "tokenizer": tokenizer_gpt35,
         "token_cnt": get_token_num_gpt35,
     },
-    "gemini-pro-vision": {
+    "gemini-1.0-pro": {
         "fn_with_ui": genai_ui,
         "fn_without_ui": genai_noui,
         "endpoint": gemini_endpoint,
+        "has_multimodal_capacity": False,
         "max_token": 1024 * 32,
         "tokenizer": tokenizer_gpt35,
         "token_cnt": get_token_num_gpt35,
     },
+    "gemini-1.5-pro": {
+        "fn_with_ui": genai_ui,
+        "fn_without_ui": genai_noui,
+        "endpoint": gemini_endpoint,
+        "has_multimodal_capacity": True,
+        "max_token": 1024 * 204800,
+        "tokenizer": tokenizer_gpt35,
+        "token_cnt": get_token_num_gpt35,
+    },
+    "gemini-1.5-flash": {
+        "fn_with_ui": genai_ui,
+        "fn_without_ui": genai_noui,
+        "endpoint": gemini_endpoint,
+        "has_multimodal_capacity": True,
+        "max_token": 1024 * 204800,
+        "tokenizer": tokenizer_gpt35,
+        "token_cnt": get_token_num_gpt35,
+    },
 
     # cohere
     "cohere-command-r-plus": {

request_llms/bridge_google_gemini.py

@@ -8,15 +8,15 @@
 import time
 from request_llms.com_google import GoogleChatInit
 from toolbox import ChatBotWithCookies
-from toolbox import get_conf, update_ui, update_ui_lastest_msg, have_any_recent_upload_image_files, trimmed_format_exc, log_chat
+from toolbox import get_conf, update_ui, update_ui_lastest_msg, have_any_recent_upload_image_files, trimmed_format_exc, log_chat, encode_image
 
 proxies, TIMEOUT_SECONDS, MAX_RETRY = get_conf('proxies', 'TIMEOUT_SECONDS', 'MAX_RETRY')
 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 predict_no_ui_long_connection(inputs, llm_kwargs, history=[], sys_prompt="", observe_window=None,
-                                  console_slience=False):
+def predict_no_ui_long_connection(inputs:str, llm_kwargs:dict, history:list=[], sys_prompt:str="", observe_window:list=[],
+                                  console_slience:bool=False):
     # 检查API_KEY
     if get_conf("GEMINI_API_KEY") == "":
         raise ValueError(f"请配置 GEMINI_API_KEY。")
@@ -44,9 +44,20 @@ def predict_no_ui_long_connection(inputs, llm_kwargs, history=[], sys_prompt="",
             raise RuntimeError(f'{gpt_replying_buffer} 对话错误')
     return gpt_replying_buffer
 
+def make_media_input(inputs, image_paths):
+    image_base64_array = []
+    for image_path in image_paths:
+        path = os.path.abspath(image_path)
+        inputs = inputs + f'<br/><br/><div align="center"><img src="file={path}"></div>'
+        base64 = encode_image(path)
+        image_base64_array.append(base64)
+    return inputs, image_base64_array
 
 def predict(inputs:str, llm_kwargs:dict, plugin_kwargs:dict, chatbot:ChatBotWithCookies,
             history:list=[], system_prompt:str='', stream:bool=True, additional_fn:str=None):
+
+    from .bridge_all import model_info
+
     # 检查API_KEY
     if get_conf("GEMINI_API_KEY") == "":
         yield from update_ui_lastest_msg(f"请配置 GEMINI_API_KEY。", chatbot=chatbot, history=history, delay=0)
@@ -57,26 +68,25 @@ def predict(inputs:str, llm_kwargs:dict, plugin_kwargs:dict, chatbot:ChatBotWith
         from core_functional import handle_core_functionality
         inputs, history = handle_core_functionality(additional_fn, inputs, history, chatbot)
 
-    if "vision" in llm_kwargs["llm_model"]:
-        have_recent_file, image_paths = have_any_recent_upload_image_files(chatbot)
-        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
-        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
-        if have_recent_file:
-            inputs = make_media_input(inputs, image_paths)
+    # multimodal capacity
+    # inspired by codes in bridge_chatgpt
+    has_multimodal_capacity = model_info[llm_kwargs['llm_model']].get('has_multimodal_capacity', False)
+    if has_multimodal_capacity:
+        has_recent_image_upload, image_paths = have_any_recent_upload_image_files(chatbot, pop=True)
+    else:
+        has_recent_image_upload, image_paths = False, []
+    if has_recent_image_upload:
+        inputs, image_base64_array = make_media_input(inputs, image_paths)
+    else:
+        inputs, image_base64_array = inputs, []
 
     chatbot.append((inputs, ""))
     yield from update_ui(chatbot=chatbot, history=history)
     genai = GoogleChatInit(llm_kwargs)
     retry = 0
     while True:
         try:
-            stream_response = genai.generate_chat(inputs, llm_kwargs, history, system_prompt)
+            stream_response = genai.generate_chat(inputs, llm_kwargs, history, system_prompt, image_base64_array, has_multimodal_capacity)
             break
         except Exception as e:
             retry += 1
@@ -112,7 +122,6 @@ def make_media_input(inputs, image_paths):
         yield from update_ui(chatbot=chatbot, history=history)
 
 
-
 if __name__ == '__main__':
     import sys
     llm_kwargs = {'llm_model': 'gemini-pro'}

request_llms/com_google.py

@@ -7,7 +7,7 @@
 import re
 import requests
 from typing import List, Dict, Tuple
-from toolbox import get_conf, encode_image, get_pictures_list, to_markdown_tabs
+from toolbox import get_conf, update_ui, encode_image, get_pictures_list, to_markdown_tabs
 
 proxies, TIMEOUT_SECONDS = get_conf("proxies", "TIMEOUT_SECONDS")
 
@@ -112,16 +112,24 @@ def html_local_img(__file, layout="left", max_width=None, max_height=None, md=Tr
     return a
 
 
+def reverse_base64_from_input(inputs):
+    pattern = re.compile(r'<br/><br/><div align="center"><img[^<>]+base64="([^"]+)"></div>')
+    base64_strings = pattern.findall(inputs)
+    return base64_strings
+
+def contain_base64(inputs):
+    base64_strings = reverse_base64_from_input(inputs)
+    return len(base64_strings) > 0
 
 class GoogleChatInit:
     def __init__(self, llm_kwargs):
         from .bridge_all import model_info
         endpoint = model_info[llm_kwargs['llm_model']]['endpoint']
         self.url_gemini = endpoint + "/%m:streamGenerateContent?key=%k"
 
-    def generate_chat(self, inputs, llm_kwargs, history, system_prompt):
+    def generate_chat(self, inputs, llm_kwargs, history, system_prompt, image_base64_array:list=[], has_multimodal_capacity:bool=False):
         headers, payload = self.generate_message_payload(
-            inputs, llm_kwargs, history, system_prompt
+            inputs, llm_kwargs, history, system_prompt, image_base64_array, has_multimodal_capacity
         )
         response = requests.post(
             url=self.url_gemini,
@@ -133,13 +141,16 @@ def generate_chat(self, inputs, llm_kwargs, history, system_prompt):
         )
         return response.iter_lines()
 
-    def __conversation_user(self, user_input, llm_kwargs):
+    def __conversation_user(self, user_input, llm_kwargs, enable_multimodal_capacity):
         what_i_have_asked = {"role": "user", "parts": []}
-        if "vision" not in self.url_gemini:
+        from .bridge_all import model_info
+
+        if enable_multimodal_capacity:
+            input_, encode_img = input_encode_handler(user_input, llm_kwargs=llm_kwargs)
+        else:
             input_ = user_input
             encode_img = []
-        else:
-            input_, encode_img = input_encode_handler(user_input, llm_kwargs=llm_kwargs)
+
         what_i_have_asked["parts"].append({"text": input_})
         if encode_img:
             for data in encode_img:
@@ -153,12 +164,12 @@ def __conversation_user(self, user_input, llm_kwargs):
                 )
         return what_i_have_asked
 
-    def __conversation_history(self, history, llm_kwargs):
+    def __conversation_history(self, history, llm_kwargs, enable_multimodal_capacity):
         messages = []
         conversation_cnt = len(history) // 2
         if conversation_cnt:
             for index in range(0, 2 * conversation_cnt, 2):
-                what_i_have_asked = self.__conversation_user(history[index], llm_kwargs)
+                what_i_have_asked = self.__conversation_user(history[index], llm_kwargs, enable_multimodal_capacity)
                 what_gpt_answer = {
                     "role": "model",
                     "parts": [{"text": history[index + 1]}],
@@ -168,7 +179,7 @@ def __conversation_history(self, history, llm_kwargs):
         return messages
 
     def generate_message_payload(
-        self, inputs, llm_kwargs, history, system_prompt
+        self, inputs, llm_kwargs, history, system_prompt, image_base64_array:list=[], has_multimodal_capacity:bool=False
     ) -> Tuple[Dict, Dict]:
         messages = [
             # {"role": "system", "parts": [{"text": system_prompt}]},  # gemini 不允许对话轮次为偶数，所以这个没有用，看后续支持吧。。。
@@ -179,21 +190,29 @@ def generate_message_payload(
             "%m", llm_kwargs["llm_model"]
         ).replace("%k", get_conf("GEMINI_API_KEY"))
         header = {"Content-Type": "application/json"}
-        if "vision" not in self.url_gemini:  # 不是vision 才处理history
+
+        if has_multimodal_capacity:
+            enable_multimodal_capacity = (len(image_base64_array) > 0) or any([contain_base64(h) for h in history])
+        else:
+            enable_multimodal_capacity = False
+
+        if not enable_multimodal_capacity:
             messages.extend(
-                self.__conversation_history(history, llm_kwargs)
+                self.__conversation_history(history, llm_kwargs, enable_multimodal_capacity)
             )  # 处理 history
-        messages.append(self.__conversation_user(inputs, llm_kwargs))  # 处理用户对话
+
+        messages.append(self.__conversation_user(inputs, llm_kwargs, enable_multimodal_capacity))  # 处理用户对话
         payload = {
             "contents": messages,
             "generationConfig": {
-                # "maxOutputTokens": 800,
+                # "maxOutputTokens": llm_kwargs.get("max_token", 1024),
                 "stopSequences": str(llm_kwargs.get("stop", "")).split(" "),
                 "temperature": llm_kwargs.get("temperature", 1),
                 "topP": llm_kwargs.get("top_p", 0.8),
                 "topK": 10,
             },
         }
+
         return header, payload