From 21b079d751d93a7bc581f7db81d0039b75cd350f Mon Sep 17 00:00:00 2001 From: zR <2448370773@qq.com> Date: Tue, 14 Nov 2023 17:45:22 +0800 Subject: [PATCH] =?UTF-8?q?GLM3=E6=94=AF=E6=8C=81=E4=BC=A0=E5=85=A5?= =?UTF-8?q?=E6=8F=90=E7=A4=BA=E8=AF=8D=20(#2058)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 支持传入提示词的Agent ChatGLM3-6B --- configs/prompt_config.py.example | 204 ++++++++++++--------- server/agent/custom_agent/ChatGLM3Agent.py | 103 ++++------- server/agent/custom_template.py | 1 - server/chat/agent_chat.py | 5 +- server/utils.py | 2 +- 5 files changed, 155 insertions(+), 160 deletions(-) diff --git a/configs/prompt_config.py.example b/configs/prompt_config.py.example index 52ff73e..710195a 100644 --- a/configs/prompt_config.py.example +++ b/configs/prompt_config.py.example @@ -17,96 +17,118 @@ # - input: 用户输入内容 # - agent_scratchpad: Agent的思维记录 -PROMPT_TEMPLATES = { - "completion": { - "default": "{input}" - }, +PROMPT_TEMPLATES = {} - "llm_chat": { - "default": "{{ input }}", - - "py": - """ - 你是一个聪明的代码助手，请你给我写出简单的py代码。 \n - {{ input }} - """ - , - }, - - "knowledge_base_chat": { - "default": - """ - <指令>根据已知信息，简洁和专业的来回答问题。如果无法从中得到答案，请说 “根据已知信息无法回答该问题”，不允许在答案中添加编造成分，答案请使用中文。 - <已知信息>{{ context }}、 - <问题>{{ question }} - """, - "text": - """ - <指令>根据已知信息，简洁和专业的来回答问题。如果无法从中得到答案，请说 “根据已知信息无法回答该问题”，答案请使用中文。 - <已知信息>{{ context }}、 - <问题>{{ question }} - """, - "Empty": # 搜不到内容的时候调用，此时没有已知信息，这个Empty可以更改，但不能删除，会影响程序使用 - """ - <指令>请根据用户的问题，进行简洁明了的回答 - <问题>{{ question }} - """, - }, - - "search_engine_chat": { - "default": - """ - <指令>这是我搜索到的互联网信息，请你根据这些信息进行提取并有调理，简洁的回答问题。如果无法从中得到答案，请说 “无法搜索到能回答问题的内容”。 - <已知信息>{{ context }}、 - <问题>{{ question }} - """, - "search": - """ - <指令>根据已知信息，简洁和专业的来回答问题。如果无法从中得到答案，请说 “根据已知信息无法回答该问题”，答案请使用中文。 - <已知信息>{{ context }}、 - <问题>{{ question }} - """, - "Empty": # 搜不到内容的时候调用，此时没有已知信息，这个Empty可以更改，但不能删除，会影响程序使用 - """ - <指令>请根据用户的问题，进行简洁明了的回答 - <问题>{{ question }} - """, - }, - - "agent_chat": { - "default": - """ - Answer the following questions as best you can. If it is in order, you can use some tools appropriately.You have access to the following tools: - - {tools} - - Please note that the "知识库查询工具" is information about the "西交利物浦大学" ,and if a question is asked about it, you must answer with the knowledge base， - Please note that the "天气查询工具" can only be used once since Question begin. - - Use the following format: - Question: the input question you must answer1 - Thought: you should always think about what to do and what tools to use. - Action: the action to take, should be one of [{tool_names}] - Action Input: the input to the action - Observation: the result of the action - ... (this Thought/Action/Action Input/Observation can be repeated zero or more times) - Thought: I now know the final answer - Final Answer: the final answer to the original input question - - - Begin! - history: - {history} - Question: {input} - Thought: {agent_scratchpad} - """, - - "ChatGLM3": # ChatGLM3必须用官方的提示词，没有修改空间，目前参数都不会传入进去 - """ - history: - {history} - Question: {input} - Thought: {agent_scratchpad} - """, - }, +PROMPT_TEMPLATES["llm_chat"] = { +"default": "{{ input }}", +"py": +""" +你是一个聪明的代码助手，请你给我写出简单的py代码。 \n +{{ input }} +""", +} + +PROMPT_TEMPLATES["knowledge_base_chat"] = { +"default": +""" +<指令>根据已知信息，简洁和专业的来回答问题。如果无法从中得到答案，请说 “根据已知信息无法回答该问题”，不允许在答案中添加编造成分，答案请使用中文。 +<已知信息>{{ context }}、 +<问题>{{ question }} +""", +"text": +""" +<指令>根据已知信息，简洁和专业的来回答问题。如果无法从中得到答案，请说 “根据已知信息无法回答该问题”，答案请使用中文。 +<已知信息>{{ context }}、 +<问题>{{ question }} +""", +"Empty": # 搜不到知识库的时候使用 +""" +请你回答我的问题: +{{ question }} +\n +""", +} +PROMPT_TEMPLATES["search_engine_chat"] = { +"default": +""" +<指令>这是我搜索到的互联网信息，请你根据这些信息进行提取并有调理，简洁的回答问题。如果无法从中得到答案，请说 “无法搜索到能回答问题的内容”。 +<已知信息>{{ context }} +<问题>{{ question }} +""", +"search": +""" +<指令>根据已知信息，简洁和专业的来回答问题。如果无法从中得到答案，请说 “根据已知信息无法回答该问题”，答案请使用中文。 +<已知信息>{{ context }}、 +<问题>{{ question }} +""", +} +PROMPT_TEMPLATES["agent_chat"] = { +"default": +""" +Answer the following questions as best you can. If it is in order, you can use some tools appropriately.You have access to the following tools: + +{tools} + +Please note that the "知识库查询工具" is information about the "西交利物浦大学" ,and if a question is asked about it, you must answer with the knowledge base， +Please note that the "天气查询工具" can only be used once since Question begin. + +Use the following format: +Question: the input question you must answer1 +Thought: you should always think about what to do and what tools to use. +Action: the action to take, should be one of [{tool_names}] +Action Input: the input to the action +Observation: the result of the action +... (this Thought/Action/Action Input/Observation can be repeated zero or more times) +Thought: I now know the final answer +Final Answer: the final answer to the original input question +Begin! + +history: {history} + +Question: {input} + +Thought: {agent_scratchpad} +""", + +"ChatGLM3": +""" +You can answer using the tools, or answer directly using your knowledge without using the tools.Respond to the human as helpfully and accurately as possible. +You have access to the following tools: +{tools} +Use a json blob to specify a tool by providing an action key (tool name) and an action_input key (tool input). +Valid "action" values: "Final Answer" or [{tool_names}] +Provide only ONE action per $JSON_BLOB, as shown: + +``` +{{{{ + "action": $TOOL_NAME, + "action_input": $INPUT +}}}} +``` + +Follow this format: + +Question: input question to answer +Thought: consider previous and subsequent steps +Action: +``` +$JSON_BLOB +``` +Observation: action result +... (repeat Thought/Action/Observation N times) +Thought: I know what to respond +Action: +``` +{{{{ + "action": "Final Answer", + "action_input": "Final response to human" +}}}} +Begin! Reminder to ALWAYS respond with a valid json blob of a single action. Use tools if necessary. Respond directly if appropriate. Format is Action:```$JSON_BLOB```then Observation:. + +history: {history} + +Question: {input} + +Thought: {agent_scratchpad} +""", } diff --git a/server/agent/custom_agent/ChatGLM3Agent.py b/server/agent/custom_agent/ChatGLM3Agent.py index b10b781..92ddc28 100644 --- a/server/agent/custom_agent/ChatGLM3Agent.py +++ b/server/agent/custom_agent/ChatGLM3Agent.py @@ -1,11 +1,11 @@ """ This file is a modified version for ChatGLM3-6B the original ChatGLM3Agent.py file from the langchain repo. - """ from __future__ import annotations import yaml from langchain.agents.structured_chat.output_parser import StructuredChatOutputParser +from langchain.memory import ConversationBufferWindowMemory from typing import Any, List, Sequence, Tuple, Optional, Union import os from langchain.agents.agent import Agent @@ -13,7 +13,7 @@ from langchain.chains.llm import LLMChain from langchain.prompts.chat import ( ChatPromptTemplate, HumanMessagePromptTemplate, - SystemMessagePromptTemplate, + SystemMessagePromptTemplate, MessagesPlaceholder, ) import json import logging @@ -26,45 +26,6 @@ from langchain.callbacks.base import BaseCallbackManager from langchain.schema.language_model import BaseLanguageModel from langchain.tools.base import BaseTool -PREFIX = """ -You can answer using the tools, or answer directly using your knowledge without using the tools. -Respond to the human as helpfully and accurately as possible. -You have access to the following tools: -""" -FORMAT_INSTRUCTIONS = """Use a json blob to specify a tool by providing an action key (tool name) and an action_input key (tool input). - -Valid "action" values: "Final Answer" or {tool_names} - -Provide only ONE action per $JSON_BLOB, as shown: - -``` -{{{{ - "action": $TOOL_NAME, - "action_input": $INPUT -}}}} -``` - -Follow this format: - -Question: input question to answer -Thought: consider previous and subsequent steps -Action: -``` -$JSON_BLOB -``` -Observation: action result -... (repeat Thought/Action/Observation N times) -Thought: I know what to respond -Action: -``` -{{{{ - "action": "Final Answer", - "action_input": "Final response to human" -}}}} -```""" -SUFFIX = """Begin! Reminder to ALWAYS respond with a valid json blob of a single action. Use tools if necessary. Respond directly if appropriate. Format is Action:```$JSON_BLOB```then Observation:. -Thought:""" - HUMAN_MESSAGE_TEMPLATE = "{input}\n\n{agent_scratchpad}" logger = logging.getLogger(__name__) @@ -77,9 +38,6 @@ class StructuredChatOutputParserWithRetries(AgentOutputParser): output_fixing_parser: Optional[OutputFixingParser] = None """The output fixing parser to use.""" - def get_format_instructions(self) -> str: - return FORMAT_INSTRUCTIONS - def parse(self, text: str) -> Union[AgentAction, AgentFinish]: special_tokens = ["Action:", "<|observation|>"] first_index = min([text.find(token) if token in text else len(text) for token in special_tokens]) @@ -112,6 +70,7 @@ Action: return parsed_obj except Exception as e: raise OutputParserException(f"Could not parse LLM output: {text}") from e + @property def _type(self) -> str: return "structured_chat_ChatGLM3_6b_with_retries" @@ -168,47 +127,57 @@ class StructuredGLM3ChatAgent(Agent): def create_prompt( cls, tools: Sequence[BaseTool], - prefix: str = PREFIX, - suffix: str = SUFFIX, - human_message_template: str = HUMAN_MESSAGE_TEMPLATE, - format_instructions: str = FORMAT_INSTRUCTIONS, + prompt: str = None, input_variables: Optional[List[str]] = None, memory_prompts: Optional[List[BasePromptTemplate]] = None, ) -> BasePromptTemplate: def tool_config_from_file(tool_name, directory="server/agent/tools/"): - """search tool yaml and return json format""" + """search tool yaml and return simplified json format""" file_path = os.path.join(directory, f"{tool_name.lower()}.yaml") try: with open(file_path, 'r', encoding='utf-8') as file: - return yaml.safe_load(file) + tool_config = yaml.safe_load(file) + # Simplify the structure if needed + simplified_config = { + "name": tool_config.get("name", ""), + "description": tool_config.get("description", ""), + "parameters": tool_config.get("parameters", {}) + } + return simplified_config except FileNotFoundError: - print(f"File not found: {file_path}") + logger.error(f"File not found: {file_path}") return None except Exception as e: - print(f"An error occurred while reading {file_path}: {e}") + logger.error(f"An error occurred while reading {file_path}: {e}") return None tools_json = [] - tool_names = "" + tool_names = [] for tool in tools: tool_config = tool_config_from_file(tool.name) if tool_config: tools_json.append(tool_config) - tool_names.join(tool.name + ", ") + tool_names.append(tool.name) + # Format the tools for output formatted_tools = "\n".join([ - json.dumps(tool, ensure_ascii=False).replace("\"", "\\\"").replace("{", "{{").replace("}", "}}") + f"{tool['name']}: {tool['description']}, args: {tool['parameters']}" for tool in tools_json ]) - format_instructions = format_instructions.format(tool_names=tool_names) - template = "\n\n".join([prefix, formatted_tools, format_instructions, suffix]) + formatted_tools = formatted_tools.replace("'", "\\'").replace("{", "{{").replace("}", "}}") + + template = prompt.format(tool_names=tool_names, + tools=formatted_tools, + history="{history}", + input="{input}", + agent_scratchpad="{agent_scratchpad}") + if input_variables is None: input_variables = ["input", "agent_scratchpad"] _memory_prompts = memory_prompts or [] messages = [ SystemMessagePromptTemplate.from_template(template), *_memory_prompts, - HumanMessagePromptTemplate.from_template(human_message_template), ] return ChatPromptTemplate(input_variables=input_variables, messages=messages) @@ -217,12 +186,10 @@ class StructuredGLM3ChatAgent(Agent): cls, llm: BaseLanguageModel, tools: Sequence[BaseTool], + prompt: str = None, callback_manager: Optional[BaseCallbackManager] = None, output_parser: Optional[AgentOutputParser] = None, - prefix: str = PREFIX, - suffix: str = SUFFIX, human_message_template: str = HUMAN_MESSAGE_TEMPLATE, - format_instructions: str = FORMAT_INSTRUCTIONS, input_variables: Optional[List[str]] = None, memory_prompts: Optional[List[BasePromptTemplate]] = None, **kwargs: Any, @@ -231,10 +198,7 @@ class StructuredGLM3ChatAgent(Agent): cls._validate_tools(tools) prompt = cls.create_prompt( tools, - prefix=prefix, - suffix=suffix, - human_message_template=human_message_template, - format_instructions=format_instructions, + prompt=prompt, input_variables=input_variables, memory_prompts=memory_prompts, ) @@ -260,7 +224,9 @@ class StructuredGLM3ChatAgent(Agent): def initialize_glm3_agent( tools: Sequence[BaseTool], llm: BaseLanguageModel, + prompt: str = None, callback_manager: Optional[BaseCallbackManager] = None, + memory: Optional[ConversationBufferWindowMemory] = None, agent_kwargs: Optional[dict] = None, *, tags: Optional[Sequence[str]] = None, @@ -269,12 +235,17 @@ def initialize_glm3_agent( tags_ = list(tags) if tags else [] agent_kwargs = agent_kwargs or {} agent_obj = StructuredGLM3ChatAgent.from_llm_and_tools( - llm, tools, callback_manager=callback_manager, **agent_kwargs + llm=llm, + tools=tools, + prompt=prompt, + callback_manager=callback_manager, **agent_kwargs ) return AgentExecutor.from_agent_and_tools( agent=agent_obj, tools=tools, callback_manager=callback_manager, + memory=memory, tags=tags_, **kwargs, ) + diff --git a/server/agent/custom_template.py b/server/agent/custom_template.py index b1f259d..1431d2a 100644 --- a/server/agent/custom_template.py +++ b/server/agent/custom_template.py @@ -21,7 +21,6 @@ class CustomPromptTemplate(StringPromptTemplate): kwargs["tool_names"] = ", ".join([tool.name for tool in self.tools]) return self.template.format(**kwargs) - class CustomOutputParser(AgentOutputParser): begin: bool = False def __init__(self): diff --git a/server/chat/agent_chat.py b/server/chat/agent_chat.py index b094caf..8ae38db 100644 --- a/server/chat/agent_chat.py +++ b/server/chat/agent_chat.py @@ -91,8 +91,11 @@ async def agent_chat(query: str = Body(..., description="用户输入", examples llm=model, tools=tools, callback_manager=None, - verbose=True, + # Langchain Prompt is not constructed directly here, it is constructed inside the GLM3 agent. + prompt=prompt_template, + input_variables=["input", "intermediate_steps", "history"], memory=memory, + verbose=True, ) else: agent = LLMSingleActionAgent( diff --git a/server/utils.py b/server/utils.py index 9368c58..27187b5 100644 --- a/server/utils.py +++ b/server/utils.py @@ -9,7 +9,7 @@ from configs import (LLM_MODELS, LLM_DEVICE, EMBEDDING_DEVICE, FSCHAT_MODEL_WORKERS, HTTPX_DEFAULT_TIMEOUT) import os from concurrent.futures import ThreadPoolExecutor, as_completed -from langchain.chat_models import ChatOpenAI, AzureChatOpenAI, ChatAnthropic +from langchain.chat_models import ChatOpenAI from langchain.llms import OpenAI, AzureOpenAI, Anthropic import httpx from typing import Literal, Optional, Callable, Generator, Dict, Any, Awaitable, Union