fix: update tools and prompts for create_structured_chat_agent

src/codeinterpreterapi/agents/agents.py

@@ -83,6 +83,6 @@ def create_agent_and_executor_experimental(llm, tools, verbose, is_ja) -> AgentE
         print("create_agent_and_executor agent=", str(type(agent)))
 
         # agent_executor
-        agent_executor = load_agent_executor(llm, tools, verbose=verbose)
+        agent_executor = load_agent_executor(llm, tools, verbose=verbose, is_ja=is_ja)
 
         return agent_executor

src/codeinterpreterapi/agents/plan_and_execute/agent_executor.py

@@ -1,22 +1,13 @@
 from typing import List, Optional
 
 from langchain.agents.agent import AgentExecutor, AgentOutputParser
-from langchain.agents.structured_chat.base import StructuredChatAgent
+from langchain.agents.structured_chat.base import StructuredChatAgent, create_structured_chat_agent
 from langchain.tools import BaseTool
 from langchain_core.callbacks import BaseCallbackManager
 from langchain_core.language_models import BaseLanguageModel
 from langchain_experimental.plan_and_execute.executors.base import ChainExecutor
 
-from codeinterpreterapi.agents.plan_and_execute.prompts import (
-    FORMAT_INSTRUCTIONS,
-    FORMAT_INSTRUCTIONS_JA,
-    HUMAN_MESSAGE_TEMPLATE,
-    SUFFIX,
-    SUFFIX_JA,
-    TASK_PREFIX,
-    TOOLS_PREFIX,
-    TOOLS_PREFIX_JA,
-)
+from codeinterpreterapi.agents.plan_and_execute.prompts import create_structured_chat_agent_prompt
 
 
 def load_agent_executor(
@@ -40,34 +31,20 @@ def load_agent_executor(
     Returns:
         ChainExecutor
     """
-    input_variables = ["previous_steps", "current_step", "agent_scratchpad"]
-
-    # message_template
-    message_template = ""
-    if include_task_in_prompt:
-        input_variables.append("objective")
-        message_template += TASK_PREFIX
-    message_template += HUMAN_MESSAGE_TEMPLATE
-
-    # format_instructions, tools_prefix, suffix
-    if is_ja:
-        format_instructions = FORMAT_INSTRUCTIONS_JA
-        tools_prefix = TOOLS_PREFIX_JA
-        suffix = SUFFIX_JA
-    else:
-        format_instructions = FORMAT_INSTRUCTIONS
-        tools_prefix = TOOLS_PREFIX
-        suffix = SUFFIX
-    agent = StructuredChatAgent.from_llm_and_tools(
+    input_variables = ["previous_steps", "current_step", "agent_scratchpad", "tools", "tool_names"]
+    print("input_variables=", input_variables)
+    prompt = create_structured_chat_agent_prompt(is_ja)
+    print("prompt=", prompt.get_prompts())
+    agent = create_structured_chat_agent(
         llm=llm,
         tools=tools,
-        callback_manager=callback_manager,
-        output_parser=output_parser,
-        prefix=tools_prefix,
-        suffix=suffix,
-        human_message_template=message_template,
-        format_instructions=format_instructions,
-        input_variables=input_variables,
+        # callback_manager=callback_manager,
+        # output_parser=output_parser,
+        # prefix=tools_prefix,
+        # suffix=suffix,
+        prompt=prompt,
+        # format_instructions=format_instructions,
+        # input_variables=input_variables,
         # memory_prompts = memory_prompts,
     )
     agent_executor = AgentExecutor.from_agent_and_tools(agent=agent, tools=tools, verbose=verbose)

src/codeinterpreterapi/agents/plan_and_execute/prompts.py

@@ -1,88 +1,121 @@
 from textwrap import dedent
 
-from langchain_core.prompts import PromptTemplate
+from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder, PromptTemplate
 from langchain_experimental.tot.prompts import JSONListOutputParser
 
-HUMAN_MESSAGE_TEMPLATE = """Previous steps: {previous_steps}
+SYSTEM_MESSAGE_TEMPLATE = '''Respond to the human as helpfully and accurately as possible. You have access to the following tools:
 
-Current objective: {current_step}
+            {tools}
 
-{agent_scratchpad}"""
+            Use a json blob to specify a tool by providing an action key (tool name) and an action_input key (tool input).
 
-TASK_PREFIX = """{objective}
+            Valid "action" values: "Final Answer" or {tool_names}
 
-"""
+            Provide only ONE action per $JSON_BLOB, as shown:
 
-TOOLS_PREFIX = (
-    """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).
+            ```
+            {{
+              "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
+            '''
+
+HUMAN_MESSAGE_TEMPLATE = '''{input}
+
+            {agent_scratchpad}
+
+            (reminder to respond in a JSON blob no matter what)'''
+
+SYSTEM_MESSAGE_TEMPLATE_JA = '''初期質問にできる限り丁寧かつ正確に答えてください。以下のツールが利用可能です:
+    {tools}
+
+1つの $JSON_BLOB では常に単一のアクションで応答してください。
+action は (TOOL_NAME) 、action_input は (INPUT) を使ってください。
 Valid "action" values: "Final Answer" or {tool_names}
 
-Provide only ONE action per $JSON_BLOB, as shown:
+フォーマットは次のようになります。
 
-```
-{{{{
-  "action": $TOOL_NAME,
-  "action_input": $INPUT
-}}}}
-```
+            ```
+            {{
+              "action": $TOOL_NAME,
+              "action_input": $INPUT
+            }}
+            ```
 
-Follow this format:
+その後のシーケンスは以下の形式に従ってください:
 
-Question: input question to answer
-Thought: consider previous and subsequent steps
+Question: 初期質問に正しく答えるための質問
+Thought: 前後のステップを検討する
 Action:
+
 ```
 $JSON_BLOB
 ```
-Observation: action result
-... (repeat Thought/Action/Observation N times)
-Thought: I know what to respond
+
+Observation: アクションの結果
+
+(Thought/Action/Observation をN回繰り返す）
+
+Thought: 何を答えるべきかわかった
 Action:
+
 ```
-{{{{
-  "action": "Final Answer",
-  "action_input": "Final response to human"
-}}}}
-```"""
-SUFFIX = """Begin! Reminder to ALWAYS respond with a valid json blob of a single action.
-Use tools if necessary. Respond directly if appropriate.
-Format is like this.
-Action:```$JSON_BLOB```
-Observation:
-Thought:"""
-
-TOOLS_PREFIX_JA = """できる限り丁寧かつ正確に人間の質問に答えてください。以下のツールが利用可能です:"""
-
-FORMAT_INSTRUCTIONS_JA = """
-始めましょう！常に単一のアクションの有効なJSONブロブで応答することを忘れないでください。
-必要に応じてツールを使用してください。適切な場合は直接回答してください。
-フォーマットは次のようにしてください。
-Action:```$JSON_BLOB```
-Observation:
-Thought:
-"""
-SUFFIX_JA = """それでは始めましょう。常に単一のアクションの有効なJSON_BLOBで応答してください。
-必要に応じてツールを使用してください。適切な場合は直接回答してください。
-フォーマットは次のようにしてください。
-Action:```$JSON_BLOB```
-Observation:
-Thought:"""
-
-
-def get_tools_prefix_prompt(is_ja: bool) -> str:
-    """Get the prefix prompt for plan_and_execute."""
-    if is_ja:
-        return TOOLS_PREFIX_JA
-    else:
-        return TOOLS_PREFIX
+{{
+  "action": "最終回答",
+  "action_input": "初期質問への最終的な回答"
+}}
+```
+
+それでは始めましょう。常に単一のアクションの有効なJSONブロブで応答することを忘れないでください。
+必要に応じてツールを使用してください。
+適切な場合は直接回答してください。
+            '''
+
+HUMAN_MESSAGE_TEMPLATE_JA = '''{input}
+
+            {agent_scratchpad}
+
+            リマインダ： 何があっても $JSON_BLOB だけで応答するようにしてください。
+'''
 
 
-def get_plan_and_execute_prompt(is_ja: bool) -> str:
-    """Get the main prompt for plan_and_execute."""
+def create_structured_chat_agent_prompt(is_ja: bool = True):
     if is_ja:
-        return FORMAT_INSTRUCTIONS_JA
+        prompt = ChatPromptTemplate.from_messages(
+            [
+                ("system", SYSTEM_MESSAGE_TEMPLATE_JA),
+                MessagesPlaceholder("chat_history", optional=True),
+                ("human", HUMAN_MESSAGE_TEMPLATE_JA),
+            ]
+        )
     else:
-        return FORMAT_INSTRUCTIONS
+        prompt = ChatPromptTemplate.from_messages(
+            [
+                ("system", SYSTEM_MESSAGE_TEMPLATE),
+                MessagesPlaceholder("chat_history", optional=True),
+                ("human", HUMAN_MESSAGE_TEMPLATE),
+            ]
+        )
+    return prompt

src/codeinterpreterapi/session.py

@@ -75,8 +75,10 @@ def __init__(
         if self.is_local:
             run_handler_func = self._run_handler_local
             arun_handler_func = self._arun_handler_local
-        self.tools: list[BaseTool] = CodeInterpreterTools.get_all(additional_tools, run_handler_func, arun_handler_func)
         self.llm: BaseLanguageModel = llm or CodeInterpreterLlm.get_llm()
+        self.tools: list[BaseTool] = CodeInterpreterTools.get_all(
+            additional_tools, run_handler_func, arun_handler_func, self.llm
+        )
         self.log("self.llm=" + str(self.llm))
 
         self.callbacks = callbacks

src/codeinterpreterapi/tools/tools.py

@@ -1,14 +1,19 @@
-from langchain_community.tools.shell.tool import ShellTool
+from langchain_community.tools.shell.tool import BaseTool, ShellTool
 from langchain_community.tools.tavily_search import TavilySearchResults
-from langchain_core.tools import BaseTool, StructuredTool
+from langchain_core.language_models import LLM, BaseLanguageModel, BaseLLM
+from langchain_core.tools import StructuredTool, Tool
+from langchain_experimental.chat_models.llm_wrapper import ChatWrapper
+from langchain_experimental.llm_bash.base import LLMBashChain
 
 from codeinterpreterapi.config import settings
 from codeinterpreterapi.schema import CodeInput
 
 
 class CodeInterpreterTools:
     @staticmethod
-    def get_all(additional_tools: list[BaseTool], run_handler_func, arun_handler_func) -> list[BaseTool]:
+    def get_all(
+        additional_tools: list[BaseTool], llm: BaseLanguageModel, run_handler_func, arun_handler_func
+    ) -> list[BaseTool]:
         additional_tools = CodeInterpreterTools.get_python(additional_tools, run_handler_func, arun_handler_func)
         additional_tools = CodeInterpreterTools.get_shell(additional_tools)
         additional_tools = CodeInterpreterTools.get_web_search(additional_tools)
@@ -39,11 +44,37 @@ def get_python(additional_tools: list[BaseTool], run_handler_func, arun_handler_
 
     @staticmethod
     def get_shell(additional_tools: list[BaseTool]) -> list[BaseTool]:
-        tools = [ShellTool()]
+        """
+        ShellTool cause this error. Should not use this.
+        pydantic.v1.error_wrappers.ValidationError: 1 validation error for ShellInput
+             commands
+             field required (type=value_error.missing)
+        """
+        shell_tool = ShellTool()
+        shell_tool.description = shell_tool.description + f"args {shell_tool.args}".replace("{", "{{").replace(
+            "}", "}}"
+        )
+        tools = [shell_tool]
+        return additional_tools + tools
+
+    @staticmethod
+    def get_shell_v2(additional_tools: list[BaseTool], llm: BaseLanguageModel) -> list[BaseTool]:
+        """
+        ShellTool cause this error. Should not use this.
+        pydantic.v1.error_wrappers.ValidationError: 1 validation error for ShellInput
+             commands
+             field required (type=value_error.missing)
+        """
+        llm_runnable = ChatWrapper(llm=llm)
+        bash_chain = LLMBashChain.from_llm(llm=llm)
+        bash_tool = Tool(
+            name="Bash", func=bash_chain.invoke, description="Executes bash commands in a terminal environment."
+        )
+
+        tools = [bash_tool]
         return additional_tools + tools
 
     @staticmethod
     def get_web_search(additional_tools: list[BaseTool]) -> list[BaseTool]:
-        # TODO: use ShellInput
         tools = [TavilySearchResults(max_results=1)]
         return additional_tools + tools