feat: customized tool-use

2024-09-08 19:41:00 +08:00
parent 7f4f96f77b
commit b5cb5eb969
5 changed files with 129 additions and 114 deletions
@@ -1,4 +1,4 @@
-import time
+import time, json
 import re, os
 import asyncio
 import traceback
@@ -16,6 +16,8 @@ from logging import Logger
 from nakuru.entities.components import Image
 from util.agent.func_call import FuncCall
 import util.agent.web_searcher as web_searcher
+from openai._exceptions import *
+from openai.types.chat.chat_completion_message_tool_call import Function

 logger: Logger = LogManager.GetLogger(log_name='astrbot')

@@ -186,31 +188,82 @@ class MessageHandler():
                image_url = comp.url if comp.url else comp.file
                break
        
-        web_search = self.context.web_search
-        if not web_search and msg_plain.startswith("ws"):
-            # leverage web search feature
-            web_search = True
-            msg_plain = msg_plain.removeprefix("ws").strip()
-        
+        # web_search = self.context.web_search
+        # if not web_search and msg_plain.startswith("ws"):
+        #     # leverage web search feature
+        #     web_search = True
+        #     msg_plain = msg_plain.removeprefix("ws").strip()
        try:
-            if web_search:
-                llm_result = await web_searcher.web_search(msg_plain, provider, message.session_id, official_fc=True)
+            if not self.llm_tools.empty():
+                # tools-use
+                tool_use_flag = True
+                llm_result = await provider.text_chat(
+                    prompt=msg_plain, 
+                    session_id=message.session_id, 
+                    tools=self.llm_tools.get_func()
+                )
+                
+                if isinstance(llm_result, Function):
+                    logger.debug(f"function-calling: {llm_result}")
+                    func_obj = None
+                    for i in self.llm_tools.func_list:
+                        if i["name"] == llm_result.name:
+                            func_obj = i["func_obj"]
+                            break
+                    if not func_obj:
+                        return MessageResult("AstrBot Function-calling 异常：未找到请求的函数调用。")
+                    try:
+                        args = json.loads(llm_result.arguments)
+                        function_invoked_ret = await func_obj(**args)
+                        has_func = True
+                    except BaseException as e:
+                        traceback.print_exc()
+                        return MessageResult("AstrBot Function-calling 异常：" + str(e))
+                else:
+                    return MessageResult(llm_result)
+        
            else:
+                # normal chat
+                tool_use_flag = False
                llm_result = await provider.text_chat(
                    prompt=msg_plain, 
                    session_id=message.session_id, 
                    image_url=image_url
                )
+        except BadRequestError as e:
+            if tool_use_flag:
+                # seems like the model don't support function-calling
+                logger.error(f"error: {e}. Using local function-calling implementation")
+                
+                try:
+                    # use local function-calling implementation
+                    args = {
+                        'question': llm_result,
+                        'func_definition': self.llm_tools.func_dump(),
+                    }
+                    _, has_func = await self.llm_tools.func_call(**args)
+                    
+                    if not has_func:
+                        # normal chat
+                        llm_result = await provider.text_chat(
+                            prompt=msg_plain, 
+                            session_id=message.session_id, 
+                            image_url=image_url
+                        )
+                except BaseException as e:
+                    logger.error(traceback.format_exc())
+                    return CommandResult("AstrBot Function-calling 异常：" + str(e))
+
        except BaseException as e:
            logger.error(traceback.format_exc())
            logger.error(f"LLM 调用失败。")
            return MessageResult("AstrBot 请求 LLM 资源失败：" + str(e))
-        
-        # concatenate the reply prefix
+
+        # concatenate reply prefix
        if self.reply_prefix:
            llm_result = self.reply_prefix + llm_result
        
-        # mask the unsafe content
+        # mask unsafe content
        llm_result = self.content_safety_helper.filter_content(llm_result)
        check = self.content_safety_helper.baidu_check(llm_result)
        if not check:
@@ -9,6 +9,7 @@ from type.config import VERSION
 from SparkleLogging.utils.core import LogManager
 from logging import Logger
 from nakuru.entities.components import Image
+from util.agent.web_searcher import search_from_bing, fetch_website_content

 logger: Logger = LogManager.GetLogger(log_name='astrbot')

@@ -212,6 +213,23 @@ class InternalCommandHandler:
            )
        elif l[1] == 'on':
            context.web_search = True
+            context.register_llm_tool("web_search", [{
+                "type": "string",
+                "name": "keyword",
+                "description": "搜索关键词"
+            }],
+                "通过搜索引擎搜索。如果问题需要获取近期、实时的消息，在网页上搜索(如天气、新闻或任何需要通过网页获取信息的问题)，则调用此函数；如果没有，不要调用此函数。",
+                search_from_bing
+            )
+            context.register_llm_tool("fetch_website_content", [{
+                "type": "string",
+                "name": "url",
+                "description": "要获取内容的网页链接"
+            }],
+                "获取网页的内容。如果问题带有合法的网页链接并且用户有需求了解网页内容(例如: `帮我总结一下 https://github.com 的内容`), 就调用此函数。如果没有，不要调用此函数。",
+                fetch_website_content
+            )
+            
            return CommandResult(
                hit=True,
                success=True,
@@ -219,6 +237,9 @@ class InternalCommandHandler:
            )
        elif l[1] == 'off':
            context.web_search = False
+            context.unregister_llm_tool("web_search")
+            context.unregister_llm_tool("fetch_website_content")
+            
            return CommandResult(
                hit=True,
                success=True,
@@ -110,6 +110,12 @@ class Context:
        '''
        self.message_handler.llm_tools.add_func(tool_name, params, desc, func)
    
+    def unregister_llm_tool(self, tool_name: str):
+        '''
+        删除一个函数调用工具。
+        '''
+        self.message_handler.llm_tools.remove_func(tool_name)
+     
    def find_platform(self, platform_name: str) -> RegisteredPlatform:
        for platform in self.platforms:
            if platform_name == platform.platform_name:
@@ -131,4 +137,9 @@ class Context:
        platform_name, message_type, id = l
        platform = self.find_platform(platform_name)
        await platform.platform_instance.send_msg_new(MessageType(message_type), id, message)
-        
+    
+    def get_current_llm_provider(self) -> Provider:
+        '''
+        获取当前的 LLM Provider。
+        '''
+        return self.message_handler.provider
@@ -23,6 +23,9 @@ class FuncCall():
    def __init__(self, provider: Provider) -> None:
        self.func_list = []
        self.provider = provider
+        
+    def empty(self) -> bool:
+        return len(self.func_list) == 0

    def add_func(self, name: str, func_args: list, desc: str, func_obj: callable) -> None:
        '''
@@ -34,7 +37,7 @@ class FuncCall():
        @param func_obj: 处理函数
        '''
        params = {
-            "type": "object",  # hardcore here
+            "type": "object",  # hard-coded here
            "properties": {}
        }
        for param in func_args:
@@ -42,14 +45,23 @@ class FuncCall():
                "type": param['type'],
                "description": param['description']
            }
-        self._func = {
+        _func = {
            "name": name,
            "parameters": params,
            "description": desc,
            "func_obj": func_obj,
        }
-        self.func_list.append(self._func)
-
+        self.func_list.append(_func)
+        
+    def remove_func(self, name: str) -> None:
+        '''
+        删除一个函数调用工具。
+        '''
+        for i, f in enumerate(self.func_list):
+            if f["name"] == name:
+                self.func_list.pop(i)
+                break
+    
    def func_dump(self) -> str:
        _l = []
        for f in self.func_list:
@@ -16,6 +16,8 @@ from util.websearch.google import Google
 from model.provider.provider import Provider
 from SparkleLogging.utils.core import LogManager
 from logging import Logger
+from type.types import Context
+from type.message_event import AstrMessageEvent

 logger: Logger = LogManager.GetLogger(log_name='astrbot')

@@ -31,24 +33,7 @@ def tidy_text(text: str) -> str:
    '''
    return text.strip().replace("\n", " ").replace("\r", " ").replace("  ", " ")

-# def special_fetch_zhihu(link: str) -> str:
-#     '''
-#     function-calling 函数, 用于获取知乎文章的内容
-#     '''
-#     response = requests.get(link, headers=HEADERS)
-#     response.encoding = "utf-8"
-#     soup = BeautifulSoup(response.text, "html.parser")
-
-#     if "zhuanlan.zhihu.com" in link:
-#         r = soup.find(class_="Post-RichTextContainer")
-#     else:
-#         r = soup.find(class_="List-item").find(class_="RichContent-inner")
-#     if r is None:
-#         print("debug: zhihu none")
-#         raise Exception("zhihu none")
-#     return tidy_text(r.text)
-
-async def search_from_bing(keyword: str) -> str:
+async def search_from_bing(context: Context, ame: AstrMessageEvent, keyword: str) -> str:
    '''
    tools, 从 bing 搜索引擎搜索
    '''
@@ -84,10 +69,11 @@ async def search_from_bing(keyword: str) -> str:
        site_result = site_result[:600] + "..." if len(site_result) > 600 else site_result
        ret += f"{idx}. {i.title} \n{i.snippet}\n{site_result}\n\n"
        idx += 1
-    return ret
+
+    return await summarize(context, ame, ret)


-async def fetch_website_content(url):
+async def fetch_website_content(context: Context, ame: AstrMessageEvent, url: str):
    header = HEADERS
    header.update({'User-Agent': random.choice(USER_AGENTS)})
    async with aiohttp.ClientSession() as session:
@@ -97,81 +83,13 @@ async def fetch_website_content(url):
            ret = doc.summary(html_partial=True)
            soup = BeautifulSoup(ret, 'html.parser')
            ret = tidy_text(soup.get_text())
-            return ret
-
-
-async def web_search(prompt: str, provider: Provider, session_id: str, official_fc: bool=False):
-    '''
-    @param official_fc: 使用官方 function-calling
-    '''
-    new_func_call = FuncCall(provider)
-
-    new_func_call.add_func("web_search", [{
-        "type": "string",
-        "name": "keyword",
-        "description": "搜索关键词"
-    }],
-        "通过搜索引擎搜索。如果问题需要获取近期、实时的消息，在网页上搜索(如天气、新闻或任何需要通过网页获取信息的问题)，则调用此函数；如果没有，不要调用此函数。",
-        search_from_bing
-    )
-    new_func_call.add_func("fetch_website_content", [{
-        "type": "string",
-        "name": "url",
-        "description": "要获取内容的网页链接"
-    }],
-        "获取网页的内容。如果问题带有合法的网页链接并且用户有需求了解网页内容(例如: `帮我总结一下 https://github.com 的内容`), 就调用此函数。如果没有，不要调用此函数。",
-        fetch_website_content
-    )
+            return await summarize(context, ame, ret)
    
-    has_func = False
-    function_invoked_ret = ""
-    if official_fc:
-        # we use official function-calling
-        try:
-            result = await provider.text_chat(prompt=prompt, session_id=session_id, tools=new_func_call.get_func())
-        except BadRequestError as e:
-            # seems dont support function-calling
-            logger.error(f"error: {e}. Try to use local function-calling implementation")
-            return await web_search(prompt, provider, session_id, official_fc=False)
-        if isinstance(result, Function):
-            logger.debug(f"function-calling: {result}")
-            func_obj = None
-            for i in new_func_call.func_list:
-                if i["name"] == result.name:
-                    func_obj = i["func_obj"]
-                    break
-            if not func_obj:
-                return await provider.text_chat(prompt=prompt, session_id=session_id, ) + "\n(网页搜索失败, 此为默认回复)"
-            try:
-                args = json.loads(result.arguments)
-                function_invoked_ret = await func_obj(**args)
-                has_func = True
-            except BaseException as e:
-                traceback.print_exc()
-                return await provider.text_chat(prompt=prompt, session_id=session_id, ) + "\n(网页搜索失败, 此为默认回复)"
-        else:
-            return result
-    else:
-        # we use our own function-calling
-        try:
-            args = {
-                'question': prompt,
-                'func_definition': new_func_call.func_dump(),
-            }
-            function_invoked_ret, has_func = await new_func_call.func_call(**args)
-            
-            if not has_func:
-                return await provider.text_chat(prompt, session_id)
-            
-        except BaseException as e:
-            logger.error(traceback.format_exc())
-            return await provider.text_chat(prompt, session_id) + "(网页搜索失败, 此为默认回复)"
-
-    if has_func:
-        await provider.forget(session_id=session_id)
-        summary_prompt = f"""
+async def summarize(context: Context, ame: AstrMessageEvent, text: str):
+                
+    summary_prompt = f"""
 你是一个专业且高效的助手，你的任务是
-1. 根据下面的相关材料对用户的问题 `{prompt}` 进行总结;
+1. 根据下面的相关材料对用户的问题 `{ame.message_str}` 进行总结;
 2. 简单地发表你对这个问题的看法。

 # 例子
@@ -183,7 +101,7 @@ async def web_search(prompt: str, provider: Provider, session_id: str, official_
 2. 请**直接输出总结**，不要输出多余的内容和提示语。

 # 相关材料
-{function_invoked_ret}"""
-        ret = await provider.text_chat(prompt=summary_prompt, session_id=session_id)
-        return ret
-    return function_invoked_ret
+{text}"""
+    
+    provider = context.get_current_llm_provider()
+    return await provider.text_chat(prompt=summary_prompt, session_id=ame.session_id)