Merge pull request #30 from LlmKira/chatgpt2

Chatgpt Api 魔法兼容实现

lab/test.py

@@ -8,6 +8,7 @@
 import time
 from typing import List
 
+from llm_kira.client.llms import ChatGptParam
 from llm_kira.radio.anchor import SearchCraw
 from loguru import logger
 from llm_kira import radio
@@ -68,19 +69,6 @@ async def completion():
     conversation_id=12094,  # random.randint(1, 10000000),
 )
 
-llm = llm_kira.client.llms.OpenAi(
-    profile=conversation,
-    api_key=openaiApiKey,
-    token_limit=4000,
-    auto_penalty=False,
-    call_func=None,
-)
-
-mem = receiver.MemoryManager(profile=conversation)
-chat_client = receiver.ChatBot(profile=conversation,
-                               llm_model=llm
-                               )
-
 
 async def mood_hook():
     _think = ThinkEngine(profile=conversation)
@@ -90,8 +78,84 @@ async def mood_hook():
     print(_think.build_status(rank=5))
 
 
+async def chatGpt():
+    llm = llm_kira.client.llms.ChatGpt(
+        profile=conversation,
+        api_key=openaiApiKey,
+        token_limit=4000,
+        auto_penalty=False,
+        call_func=None,
+    )
+
+    mem = llm_kira.client.MemoryManager(profile=conversation)
+    chat_client = llm_kira.client.ChatBot(
+        profile=conversation,
+        llm_model=llm
+    )
+    promptManager = llm_kira.creator.engine.PromptEngine(
+        reverse_prompt_buffer=False,
+        profile=conversation,
+        connect_words="\n",
+        memory_manger=mem,
+        llm_model=llm,
+        description="晚上了，这里是河边",
+        reference_ratio=0.5,
+        forget_words=["忘掉对话"],
+        optimizer=Optimizer.SinglePoint,
+    )
+    # 大型数据对抗测试
+    # promptManager.insert_prompt(prompt=PromptItem(start="Neko", text=random_string(8000)))
+    # promptManager.insert_prompt(prompt=PromptItem(start="Neko", text=random_string(500)))
+
+    # 多 prompt 对抗测试
+    testPrompt = input("TestPrompt:")
+    promptManager.insert_prompt(prompt=PromptItem(start="Neko", text="喵喵喵"))
+    promptManager.insert_interaction(Interaction(single=True, ask=PromptItem(start="alice", text="MewMewMewMew")))
+    _result = await promptManager.build_skeleton(query=testPrompt,
+                                                 llm_task="Summary Text" if len(
+                                                     testPrompt) > 20 else None,
+                                                 skeleton=random.choice([SearchCraw(
+                                                     deacon=["https://www.bing.com/search?q={}&form=QBLH"])])
+                                                 )
+    _index = 1
+    for item in _result:
+        logger.trace(item.content)
+        item.ask.start = f"[{_index}]"
+        promptManager.insert_knowledge(knowledge=item)
+        _index += 1
+    promptManager.insert_knowledge(Interaction(single=True, ask=PromptItem(start="alice", text="MewMewMewMew")))
+    # 测试
+    promptManager.insert_prompt(prompt=PromptItem(start=conversation.start_name, text=testPrompt))
+    response = await chat_client.predict(
+        prompt=promptManager,
+        llm_param=ChatGptParam(model_name="gpt-3.5-turbo", temperature=0.8, presence_penalty=0.1, n=1, best_of=1),
+        predict_tokens=1000,
+    )
+    print(f"id {response.conversation_id}")
+    print(f"ask {response.ask}")
+    print(f"reply {response.reply}")
+    print(f"usage:{response.llm.usage}")
+    print(f"raw:{response.llm.raw}")
+    print(f"---{response.llm.time}---")
+    promptManager.clean(clean_prompt=True, clean_knowledge=False, clean_memory=False)
+    return "End"
+
+
 async def chat():
-    promptManager = llm_kira.creator.PromptEngine(
+    llm = llm_kira.client.llms.OpenAi(
+        profile=conversation,
+        api_key=openaiApiKey,
+        token_limit=4000,
+        auto_penalty=False,
+        call_func=None,
+    )
+
+    mem = llm_kira.client.MemoryManager(profile=conversation)
+    chat_client = llm_kira.client.ChatBot(
+        profile=conversation,
+        llm_model=llm
+    )
+    promptManager = llm_kira.creator.engine.PromptEngine(
         reverse_prompt_buffer=False,
         profile=conversation,
         connect_words="\n",
@@ -181,7 +245,13 @@ async def Sentiment():
 async def Sim():
     # response = llm_kira.utils.chat.Utils.edit_similarity(pre="4552", aft="1224")
     # print(response)
-    response = llm_kira.utils.chat.Sim.cosion_similarity(pre="", aft="你是不是啊")
+    test1 = """
+    早苗（さなえ）
+耕种水稻时刚刚种植的幼苗
+原型是守矢早苗（もりやさなえ，生于1945年），守矢家第七十八代当主，是实际存在的人物。
+守矢家是洩矢神的子孙，现任诹访神社下社神长官。洩矢神的祭祀司守矢家代代口传的祭神秘法。那个秘传是一脉相承的，在半夜没有火光的祈祷殿之中秘密传授。但是随着时代的变迁，世袭神官制度在明治五年被取消了。到明治六年，家传之宝（包括：印（印文「卖神祝印」）与镜、太刀等）从诹访大社上社被移走家里只残留下用佐奈伎铃（在大御立座祭神中所使用的）祭祀御左口神的方法。在明治时代，守矢实久（第七十六代当主）被取消了神长官一职，可惜当时口传秘法已失，实久只告诉了守矢真幸（第七十七代当主，实久之弟，诹访大社的祢宜宫司）剩下的部分。到守矢早苗（第七十八代当主，真幸之孙，平成18年(注)3月末从校长(注)[6]一职退下之后，一直致力于环境保护的演讲）这一代，已经不再继承代代相传的已消失的秘法了。到现在，再也没有人知道守矢祭神秘法了。
+    """
+    response = llm_kira.utils.chat.Sim.cosion_similarity(pre=test1, aft="守矢家第七十八代当主")
     print(response)
 
 
@@ -230,7 +300,8 @@ async def Web():
 # asyncio.run(completion())
 # asyncio.run(mood_hook())
 # asyncio.run(Web())
-asyncio.run(chat())
+# asyncio.run(chat())
+asyncio.run(chatGpt())
 # asyncio.run(Moderation())
 # asyncio.run(Sentiment())
 # asyncio.run(KeyParse())

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "llm_kira"
-version = "0.6.50"
+version = "0.7.0"
 description = "chatbot client for llm"
 authors = ["sudoskys <[email protected]>"]
 maintainers = [

src/llm_kira/client/Optimizer.py

@@ -179,7 +179,10 @@ def run(self) -> List[Interaction]:
         knowledge = Scorer.build_weight(self.knowledge)
         _knowledge_token_limit = int(self.token_limit * self.reference_ratio)
         _interaction_token_limit = self.token_limit - _knowledge_token_limit
-        _returner = [Interaction(single=True, ask=PromptItem(start="*", text=self.desc))]
+
+        # Desc
+        _returner = [Interaction(single=True, ask=PromptItem(start="system", text=self.desc))]
+
         _old_prompt = interaction[:1]
         # Desc
         if self.tokenizer(self.desc) > self.token_limit:

src/llm_kira/client/__init__.py

@@ -6,5 +6,5 @@
 
 from .llms import openai
 from .enhance import Support
-from .agent import Conversation,MemoryManager
+from .agent import Conversation, MemoryManager
 from .anchor import ChatBot

src/llm_kira/client/agent.py

@@ -76,3 +76,4 @@ def read_context(self) -> List[Interaction]:
     def save_context(self, message: List[Interaction], override: bool = True):
         self._DataManager.save(interaction_flow=message, override=override)
         return message
+

src/llm_kira/client/anchor.py

@@ -10,12 +10,10 @@
 # from loguru import logger
 from .llms.base import LlmBaseParam
 from .llms.openai import LlmBase
-from .types import LlmReturn, Interaction, PromptItem
+from .types import LlmReturn, PromptItem
 from ..creator.engine import PromptEngine
-from ..error import LLMException
 
-# Utils
-from ..utils.chat import Sim
+from ..error import LLMException
 
 # Completion
 from .types import ChatBotReturn
@@ -37,40 +35,17 @@ def __init__(self,
         if llm_model is None:
             raise LLMException("Whats your llm model?")
 
-    def __person(self, prompt, prompt_list):
-        _person_list = [f"{self.profile.start_name}:",
-                        f"{self.profile.restart_name}:",
-                        f"{self.profile.start_name}：",
-                        f"{self.profile.restart_name}：",
-                        ]
-        for item in prompt_list:
-            if item.ask.connect_words.strip() in [":", "："]:
-                _person_list.append(f"{item.ask.start}{item.ask.connect_words}")
-        _person_list = self.__rank_name(prompt=prompt.prompt, users=_person_list)
-        return _person_list
-
-    @staticmethod
-    def __rank_name(prompt: str, users: List[str]):
-        __temp = {}
-        for item in users:
-            __temp[item] = 0
-        users = list(__temp.keys())
-        _ranked = list(sorted(users, key=lambda i: Sim.cosion_similarity(pre=str(prompt), aft=str(i)), reverse=True))
-        return _ranked
-
     async def predict(self,
                       prompt: PromptEngine,
                       predict_tokens: Union[int] = 100,
                       llm_param: LlmBaseParam = None,
                       parse_reply: Callable[[list], str] = None,
-                      rank_name: bool = True,
                       ) -> ChatBotReturn:
         """
         :param prompt: PromptEngine
         :param predict_tokens: 预测 Token 位
         :param llm_param: 大语言模型参数
         :param parse_reply: Callable[[list], str] 覆写解析方法
-        :param rank_name: 自动排序停止词减少第三人称的冲突出现
         """
         self.prompt = prompt
         # ReWrite
@@ -79,28 +54,17 @@ async def predict(self,
         if predict_tokens > self.llm.get_token_limit():
             # Or Auto Cut?
             raise LLMException("Why your predict token > set token limit?")
-        _llm_result_limit = self.llm.get_token_limit() - predict_tokens
-        _llm_result_limit = _llm_result_limit if _llm_result_limit > 0 else 1
-        # Get
-        _prompt_index, _prompt = self.prompt.build_prompt(predict_tokens=predict_tokens)
-        _prompt_list = []
-        _person_list = None if not rank_name else self.__person(prompt=_prompt_index, prompt_list=_prompt)
-
-        # Prompt 构建
-        for item in _prompt:
-            _prompt_list.extend(item.content)
 
-        prompt_build = "\n".join(_prompt_list) + f"\n{self.profile.restart_name}:"
-        prompt_build = self.llm.resize_sentence(prompt_build, token=_llm_result_limit)
-        # ODO
-        # logger.trace(prompt_build)
+        # Get Question Index
+        _prompt_index = self.prompt.prompt
         # Get
         llm_result: LlmReturn = await self.llm.run(
-            prompt=prompt_build,
+            prompt=prompt,
             predict_tokens=predict_tokens,
-            llm_param=llm_param,
-            stop_words=_person_list
+            llm_param=llm_param
         )
+        prompt.clean(clean_prompt=True)
+
         self.prompt.build_interaction(
             ask=_prompt_index,
             response=PromptItem(

src/llm_kira/client/llms/base.py

@@ -8,10 +8,8 @@
 import os
 from abc import abstractmethod, ABC
 from typing import Union, Optional, Callable, Any, Dict, Tuple, Mapping, List
-
 from loguru import logger
 from pydantic import BaseModel
-
 from ..types import LlmReturn
 
 
@@ -87,7 +85,7 @@ def parse_usage(response) -> Optional[int]:
 
     @abstractmethod
     async def run(self,
-                  prompt: str,
+                  prompt: Any,
                   validate: Union[List[str], None] = None,
                   predict_tokens: int = 500,
                   llm_param: LlmBaseParam = None,