basic example of meta prompter

Former-commit-id: e2d6a42402

swarms/agents/meta_prompter.py

@@ -0,0 +1,143 @@
+from langchain.chains import LLMChain
+from langchain.prompts import PromptTemplate
+from langchain.memory import ConversationBufferWindowMemory
+
+class MetaPrompterAgent:
+    """
+    Meta Prompting Agent
+    The Meta Prompting Agent has 1 purpose: to create better prompts for an agent.
+
+    The meta prompting agent would be used in this flow: 
+    user task -> MetaPrompterAgent -> Agent
+
+    Args:
+        llm (BaseLanguageModel): Language Model
+        max_iters (int, optional): Maximum number of iterations. Defaults to 3.
+        max_meta_iters (int, optional): Maximum number of meta iterations. Defaults to 5.
+        failed_phrase (str, optional): Phrase to indicate failure. Defaults to "task failed".
+        success_phrase (str, optional): Phrase to indicate success. Defaults to "task succeeded".
+        instructions (str, optional): Instructions to be used in the meta prompt. Defaults to "None".
+        template (str, optional): Template to be used in the meta prompt. Defaults to None.
+        memory (ConversationBufferWindowMemory, optional): Memory to be used in the meta prompt. Defaults to None.
+        meta_template (str, optional): Template to be used in the meta prompt. Defaults to None.
+        human_input (bool, optional): Whether to use human input. Defaults to False.
+    
+    Returns:
+        str: Response from the agent
+
+    Usage:
+    --------------
+    agent = MetaPrompterAgent(llm=OpenAI, instructions="Please tell me the weather in New York")
+    agent.run("What is the weather in New York?")    
+    """
+    def __init__(
+        self,
+        llm,
+        max_iters: int = 3,
+        max_meta_iters: int = 5,
+        failed_phrase: str = "task failed",
+        success_phrase: str = "task succeeded",
+        instructions: str = "None",
+        template: str = None,
+        memory = None,
+        meta_template: str = None,
+        human_input: bool = False,
+    ):
+        self.llm = llm
+        self.max_iters = max_iters
+        self.max_meta_iters = max_meta_iters
+        self.failed_phrase = failed_phrase
+        self.success_phrase = success_phrase
+        self.instructions = instructions
+        self.template = template
+        self.memory = memory 
+        self.meta_template = meta_template
+        self.human_input = human_input
+
+        if memory is None:
+            memory = ConversationBufferWindowMemory()
+            memory.ai_prefix = "Assistant:"
+
+        self.template = f"""
+        Instructions: {self.instructions}
+        {{{memory.memory_key}}}
+        Human: {{human_input}}
+        Assistant:
+        """
+
+        prompt = PromptTemplate(
+            input_variables=["history", "human_input"], template=template
+        )
+
+        self.chain = LLMChain(
+            llm=self.llm(),
+            prompt=prompt,
+            verbose=True,
+            memory=ConversationBufferWindowMemory(),
+        )
+
+    def get_chat_history(self, chain_memory):
+        """Get Chat History from the memory"""
+        memory_key = chain_memory.memory_key
+        chat_history = chain_memory.load_memory_variables(memory_key)[memory_key]
+        return chat_history
+
+    def get_new_instructions(self, meta_output):
+        """Get New Instructions from the meta_output"""
+        delimiter = "Instructions: "
+        new_instructions = meta_output[meta_output.find(delimiter) + len(delimiter):]
+        return new_instructions
+    
+    def run(self, task: str):
+        """
+        Run the MetaPrompterAgent
+
+        Args:
+            task (str): The task to be completed
+
+        Returns:
+            str: The response from the agent
+        """
+        key_phrases = [
+            self.success_phrase,
+            self.failed_phrase
+        ]
+
+        for i in range(self.max_meta_iters):
+            print(f"[Epsisode: {i+1}/{self.max_meta_iters}]")
+
+            chain = self.chain(
+                memory=None
+            )
+
+            output = chain.predict(human_input=task)
+            
+            for j in range(self.max_iters):
+                print(f"(Step {j+1}/{self.max_iters})")
+                print(f"Assistant: {output}")
+                print(f"Human: ")
+
+
+
+                if self.human_input:
+                    human_input = input()
+                
+                if any(
+                    phrase in human_input.lower() for phrase in key_phrases
+                ):
+                    break
+
+                output = chain.predict(human_input.lower)
+            
+            if self.success_phrase in human_input.lower():
+                print(f"You succeed! Thanks for using!")
+                return
+            
+            meta_chain = self.initialize_meta_chain()
+            meta_output = meta_chain.predict(chat_history=self.get_chat_history(chain.memory))
+            print(f"Feedback: {meta_output}")
+            self.instructions = self.get_new_instructions(meta_output)
+            print(f"New Instruction: {self.instructions}")
+            print("\n" + "#" * 80 + "\n")
+
+