GLM3支持传入提示词 (#2058)

支持传入提示词的Agent ChatGLM3-6B

configs/prompt_config.py.example

@@ -17,23 +17,18 @@
 #   - input: 用户输入内容
 #   - agent_scratchpad: Agent的思维记录
 
-PROMPT_TEMPLATES = {
+PROMPT_TEMPLATES = {}
-    "completion": {
-        "default": "{input}"
-    },
 
-    "llm_chat": {
+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": {
+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。 </指令>
@@ -46,18 +41,18 @@ PROMPT_TEMPLATES = {
 <已知信息>{{ context }}</已知信息>、
 <问题>{{ question }}</问题>
 """,
-        "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
+"Empty":  # 搜不到知识库的时候使用
 """
-        <指令>请根据用户的问题，进行简洁明了的回答</指令>
+请你回答我的问题:
-        <问题>{{ question }}</问题>
+{{ question }}
+\n
 """,
-    },
+}
-
+PROMPT_TEMPLATES["search_engine_chat"] = {
-    "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”。 </指令>
-            <已知信息>{{ context }}</已知信息>、
+<已知信息>{{ context }}</已知信息>
 <问题>{{ question }}</问题>
 """,
 "search":
@@ -66,14 +61,8 @@ PROMPT_TEMPLATES = {
 <已知信息>{{ context }}</已知信息>、
 <问题>{{ question }}</问题>
 """,
-        "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
+}
-            """
+PROMPT_TEMPLATES["agent_chat"] = {
-        <指令>请根据用户的问题，进行简洁明了的回答</指令>
-        <问题>{{ question }}</问题>
-        """,
-    },
-
-    "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:
@@ -92,21 +81,54 @@ PROMPT_TEMPLATES = {
 ... (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}
+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
+"ChatGLM3":
 """
-        history:
+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.
-        {history}
+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}
 """,
-    },
 }

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,
+            prompt: str = None,
-            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,
     ) -> 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)
+        formatted_tools = formatted_tools.replace("'", "\\'").replace("{", "{{").replace("}", "}}")
-        template = "\n\n".join([prefix, formatted_tools, format_instructions, suffix])
+
+        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,
+            prompt=prompt,
-            suffix=suffix,
-            human_message_template=human_message_template,
-            format_instructions=format_instructions,
             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,
     )
+

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):

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(

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