perf: adapt glm-4v-flash

astrbot/core/config/default.py

@@ -278,7 +278,7 @@ CONFIG_METADATA_2 = {
                     },
                     "zhipu": {
                         "id": "zhipu_default",
-                        "type": "openai_chat_completion",
+                        "type": "zhipu_chat_completion",
                         "enable": True,
                         "key": [],
                         "api_base": "https://open.bigmodel.cn/api/paas/v4/",

astrbot/core/provider/manager.py

@@ -29,6 +29,8 @@ class ProviderManager():
             match provider_cfg['type']:
                 case "openai_chat_completion":
                     from .sources.openai_source import ProviderOpenAIOfficial # noqa: F401
+                case "zhipu_chat_completion":
+                    from .sources.zhipu_source import ProviderZhipu # noqa: F401
                 case "llm_tuner":
                     logger.info("加载 LLM Tuner 工具 ...")
                     from .sources.llmtuner_source import LLMTunerModelLoader # noqa: F401

astrbot/core/provider/sources/openai_source.py

@@ -162,7 +162,12 @@ class ProviderOpenAIOfficial(Provider):
                 logger.warning(f"请求失败：{e}。上下文长度超过限制。尝试弹出最早的记录然后重试。")
                 self.pop_record(session_id)
             logger.warning(traceback.format_exc())
-        
+            
+        await self.save_history(contexts, new_record, session_id, llm_response)
+
+        return llm_response
+    
+    async def save_history(self, contexts: List, new_record: dict, session_id: str, llm_response: LLMResponse):
         if llm_response.role == "assistant" and session_id:
             # 文本回复
             if not contexts:
@@ -180,8 +185,6 @@ class ProviderOpenAIOfficial(Provider):
                 }]
             self.db_helper.update_llm_history(session_id, json.dumps(self.session_memory[session_id]), self.provider_config['type'])
         
-        return llm_response
-
     async def forget(self, session_id: str) -> bool:
         self.session_memory[session_id] = []
         return True

astrbot/core/provider/sources/zhipu_source.py

@@ -0,0 +1,73 @@
+import traceback
+from astrbot.core.db import BaseDatabase
+from astrbot import logger
+from astrbot.core.provider.func_tool_manager import FuncCall
+from typing import List
+from ..register import register_provider_adapter
+from astrbot.core.provider.entites import LLMResponse
+from .openai_source import ProviderOpenAIOfficial
+
+@register_provider_adapter("zhipu_chat_completion", "智浦 Chat Completion 提供商适配器")
+class ProviderZhipu(ProviderOpenAIOfficial):
+    def __init__(
+        self, 
+        provider_config: dict, 
+        provider_settings: dict,
+        db_helper: BaseDatabase, 
+        persistant_history = True
+    ) -> None:
+        super().__init__(provider_config, provider_settings, db_helper, persistant_history)
+    
+    async def text_chat(
+        self,
+        prompt: str,
+        session_id: str,
+        image_urls: List[str]=None,
+        func_tool: FuncCall=None,
+        contexts=None,
+        system_prompt=None,
+        **kwargs
+    ) -> LLMResponse: 
+        new_record = await self.assemble_context(prompt, image_urls)
+        context_query = []
+        
+        if not contexts:
+            context_query = [*self.session_memory[session_id], new_record]
+        else:
+            context_query = [*contexts, new_record]
+        
+        model_cfgs: dict = self.provider_config.get("model_config", {})
+        # glm-4v-flash 只支持一张图片
+        model: str = model_cfgs.get("model", "")
+        if model.lower() == 'glm-4v-flash' and image_urls and len(context_query) > 1:
+            logger.debug("glm-4v-flash 只支持一张图片，将只保留最后一张图片")
+            logger.debug(context_query)
+            new_context_query_ = []
+            for i in range(0, len(context_query) - 1, 2):
+                if isinstance(context_query[i].get("content", ""), list):
+                    continue
+                new_context_query_.append(context_query[i])
+                new_context_query_.append(context_query[i+1])
+            new_context_query_.append(context_query[-1]) # 保留最后一条记录
+            context_query = new_context_query_
+            logger.debug(context_query)
+            
+        if system_prompt:
+            context_query.insert(0, {"role": "system", "content": system_prompt})
+            
+        payloads = {
+            "messages": context_query,
+            **model_cfgs
+        }    
+        llm_response = None
+        try:
+            llm_response = await self._query(payloads, func_tool)
+        except Exception as e:
+            if "maximum context length" in str(e):
+                logger.warning(f"请求失败：{e}。上下文长度超过限制。尝试弹出最早的记录然后重试。")
+                self.pop_record(session_id)
+            logger.warning(traceback.format_exc())
+            
+        await self.save_history(contexts, new_record, session_id, llm_response)
+
+        return llm_response