Supply prompt prefix as a list of messages

motleycrew/agents/langchain/langchain.py

@@ -26,6 +26,7 @@ def __init__(
         tools: Sequence[MotleySupportedTool] | None = None,
         output_handler: MotleySupportedTool | None = None,
         chat_history: bool | GetSessionHistoryCallable = True,
+        input_as_messages: bool = False,
         verbose: bool = False,
     ):
         """
@@ -65,6 +66,8 @@ def __init__(
                 See :class:`langchain_core.runnables.history.RunnableWithMessageHistory`
                 for more details.
 
+            input_as_messages: Whether the agent expects a list of messages as input instead of a single string.
+
             verbose: Whether to log verbose output.
         """
         super().__init__(
@@ -85,6 +88,8 @@ def __init__(
         else:
             self.get_session_history_callable = chat_history
 
+        self.input_as_messages = input_as_messages
+
     def materialize(self):
         """Materialize the agent and wrap it in RunnableWithMessageHistory if needed."""
         if self.is_materialized:
@@ -141,7 +146,7 @@ def invoke(
         config: Optional[RunnableConfig] = None,
         **kwargs: Any,
     ) -> Any:
-        prompt = self.prepare_for_invocation(input=input)
+        prompt = self.prepare_for_invocation(input=input, prompt_as_messages=self.input_as_messages)
 
         config = add_default_callbacks_to_langchain_config(config)
         if self.get_session_history_callable:

motleycrew/agents/langchain/tool_calling_react.py

@@ -1,16 +1,18 @@
 from __future__ import annotations
 
-from typing import Sequence, Optional
+from typing import Sequence, Optional, Callable
 
 from langchain.agents import AgentExecutor
 from langchain.agents.format_scratchpad.tools import format_to_tool_messages
 from langchain.agents.output_parsers.tools import ToolsAgentOutputParser
 from langchain_core.language_models import BaseChatModel
 from langchain_core.prompts.chat import ChatPromptTemplate
-from langchain_core.runnables import Runnable, RunnablePassthrough
+from langchain_core.runnables import Runnable, RunnablePassthrough, RunnableLambda
 from langchain_core.runnables.history import GetSessionHistoryCallable
 from langchain_core.tools import BaseTool
 
+from motleycrew.common.utils import print_passthrough
+
 try:
     from langchain_anthropic import ChatAnthropic
 except ImportError:
@@ -21,7 +23,7 @@
     ToolCallingReActPromptsForOpenAI,
     ToolCallingReActPromptsForAnthropic,
 )
-from motleycrew.common import LLMFramework
+from motleycrew.common import LLMFramework, Defaults
 from motleycrew.common import MotleySupportedTool
 from motleycrew.common.llms import init_llm
 from motleycrew.tools import MotleyTool
@@ -63,6 +65,7 @@ def create_tool_calling_react_agent(
     tools: Sequence[BaseTool],
     prompt: ChatPromptTemplate,
     output_handler: BaseTool | None = None,
+    intermediate_steps_processor: Callable | None = None,
 ) -> Runnable:
     prompt = prompt.partial(
         tools=render_text_description(list(tools)),
@@ -76,12 +79,18 @@ def create_tool_calling_react_agent(
 
     llm_with_tools = llm.bind_tools(tools=tools_for_llm)
 
+    if not intermediate_steps_processor:
+        intermediate_steps_processor = lambda x: x
+
     agent = (
         RunnablePassthrough.assign(
-            agent_scratchpad=lambda x: format_to_tool_messages(x["intermediate_steps"]),
+            agent_scratchpad=lambda x: format_to_tool_messages(
+                intermediate_steps_processor(x["intermediate_steps"])
+            ),
             additional_notes=lambda x: x.get("additional_notes") or [],
         )
         | prompt
+        | RunnableLambda(print_passthrough)
         | llm_with_tools
         | ToolsAgentOutputParser()
     )
@@ -108,6 +117,8 @@ def __init__(
         handle_parsing_errors: bool = True,
         handle_tool_errors: bool = True,
         llm: BaseChatModel | None = None,
+        max_iterations: int | None = Defaults.DEFAULT_REACT_AGENT_MAX_ITERATIONS,
+        intermediate_steps_processor: Callable | None = None,
         verbose: bool = False,
     ):
         """
@@ -128,6 +139,9 @@ def __init__(
             handle_tool_errors: Whether to handle tool errors.
                 If True, `handle_tool_error` and `handle_validation_error` in all tools
                 are set to True.
+            max_iterations: The maximum number of agent iterations.
+            intermediate_steps_processor: Function that modifies the intermediate steps array
+                in some way before each agent iteration.
             llm: Language model to use.
             verbose: Whether to log verbose output.
 
@@ -162,6 +176,7 @@ def agent_factory(
                 tools=tools_for_langchain,
                 prompt=prompt,
                 output_handler=output_handler_for_langchain,
+                intermediate_steps_processor=intermediate_steps_processor,
             )
 
             if output_handler_for_langchain:
@@ -177,6 +192,7 @@ def agent_factory(
                 tools=tools_for_langchain,
                 handle_parsing_errors=handle_parsing_errors,
                 verbose=verbose,
+                max_iterations=max_iterations,
             )
             return agent_executor
 
@@ -188,5 +204,6 @@ def agent_factory(
             tools=tools,
             output_handler=output_handler,
             chat_history=chat_history,
+            input_as_messages=True,
             verbose=verbose,
         )

motleycrew/agents/langchain/tool_calling_react_prompts.py

@@ -25,7 +25,7 @@ def __init__(self):
                 MessagesPlaceholder(variable_name="chat_history", optional=True),
                 ("system", self.main_instruction),
                 MessagesPlaceholder(variable_name="example_messages", optional=True),
-                ("user", "{input}"),
+                MessagesPlaceholder(variable_name="input"),
                 MessagesPlaceholder(variable_name="agent_scratchpad"),
                 MessagesPlaceholder(variable_name="additional_notes", optional=True),
             ]
@@ -64,14 +64,14 @@ class ToolCallingReActPromptsForOpenAI(ToolCallingReActPrompts):
 Begin!
 """
 
-    output_instruction_with_output_handler = """
+    output_instruction_without_output_handler = """
 If you have sufficient information to answer the question, your reply must look like
 ```
 Final Answer: [the final answer to the original input question]
 ```
 but without the backticks."""
 
-    output_instruction_without_output_handler = """
+    output_instruction_with_output_handler = """
 If you have sufficient information to answer the question, you must call the output handler tool.
 
 NEVER return the final answer directly, but always do it by CALLING this tool:

motleycrew/agents/parent.py

@@ -10,6 +10,7 @@
     Union,
 )
 
+from langchain_core.messages import BaseMessage
 from langchain_core.prompts.chat import ChatPromptTemplate, HumanMessage, SystemMessage
 from langchain_core.runnables import RunnableConfig
 from langchain_core.tools import StructuredTool
@@ -108,13 +109,14 @@ def __str__(self):
         return self.__repr__()
 
     def compose_prompt(
-        self, input_dict: dict, prompt: ChatPromptTemplate | str
-    ) -> Union[str, ChatPromptTemplate]:
+        self, input_dict: dict, prompt: ChatPromptTemplate | str, as_messages: bool = False
+    ) -> Union[str, list[BaseMessage]]:
         """Compose the agent's prompt from the prompt prefix and the provided prompt.
 
         Args:
             input_dict: The input dictionary to the agent.
             prompt: The prompt to be added to the agent's prompt.
+            as_messages: Whether the prompt should be returned as a Langchain messages list instead of a single string.
 
         Returns:
             The composed prompt.
@@ -145,6 +147,9 @@ def compose_prompt(
             else:
                 raise ValueError("Prompt must be a string or a ChatPromptTemplate")
 
+        if as_messages:
+            return prompt_messages
+
         # TODO: pass the unformatted messages list to agents that can handle it
         prompt = "\n\n".join([m.content for m in prompt_messages]) + "\n"
         return prompt
@@ -237,13 +242,15 @@ def materialize(self):
         else:
             self._agent = self.agent_factory(tools=self.tools)
 
-    def prepare_for_invocation(self, input: dict) -> str:
+    def prepare_for_invocation(self, input: dict, prompt_as_messages: bool = False) -> str:
         """Prepare the agent for invocation by materializing it and composing the prompt.
 
         Should be called in the beginning of the agent's invoke method.
 
         Args:
             input: the input to the agent
+            prompt_as_messages: Whether the prompt should be returned as a Langchain messages list
+                instead of a single string.
 
         Returns:
             str: the composed prompt
@@ -254,7 +261,7 @@ def prepare_for_invocation(self, input: dict) -> str:
             self.output_handler.agent = self
             self.output_handler.agent_input = input
 
-        prompt = self.compose_prompt(input, input.get("prompt"))
+        prompt = self.compose_prompt(input, input.get("prompt"), as_messages=prompt_as_messages)
         return prompt
 
     def add_tools(self, tools: Sequence[MotleySupportedTool]):

motleycrew/common/defaults.py

@@ -5,6 +5,7 @@
 class Defaults:
     """Default values for various settings."""
 
+    DEFAULT_REACT_AGENT_MAX_ITERATIONS = 15
     DEFAULT_LLM_FAMILY = LLMFamily.OPENAI
     DEFAULT_LLM_NAME = "gpt-4o"
     DEFAULT_LLM_TEMPERATURE = 0.0

motleycrew/common/utils.py

@@ -1,8 +1,10 @@
 """Various helpers and utility functions used throughout the project."""
+
+import hashlib
 import sys
 from typing import Optional, Sequence
-import hashlib
 from urllib.parse import urlparse
+
 from langchain_core.messages import BaseMessage
 
 from motleycrew.common.exceptions import ModuleNotInstalled
@@ -47,9 +49,8 @@ def generate_hex_hash(data: str, length: Optional[int] = None):
 def print_passthrough(x):
     """A helper function useful for debugging LCEL chains. It just returns the input value.
 
-    You can put a breakpoint in this function to debug the chain.
+    You can put a breakpoint in this function to debug a chain.
     """
-
     return x