Feature: Added Self Feedback (#3013)

* Feature: Added Self Feedback

* minor fix: complied to flake8

* Add: Self Feedback To Usage.md

* Add: role/goal allignment

* Added: warning to usage.md

* fix: Formatted with black

---------

Co-authored-by: Richard Beales <rich@richbeales.net>

autogpt/agent/agent.py

@@ -1,3 +1,4 @@
+import yaml
 from colorama import Fore, Style
 
 from autogpt.app import execute_command, get_command
@@ -5,6 +6,7 @@ from autogpt.chat import chat_with_ai, create_chat_message
 from autogpt.config import Config
 from autogpt.json_utils.json_fix_llm import fix_json_using_multiple_techniques
 from autogpt.json_utils.utilities import validate_json
+from autogpt.llm_utils import create_chat_completion
 from autogpt.logs import logger, print_assistant_thoughts
 from autogpt.speech import say_text
 from autogpt.spinner import Spinner
@@ -129,8 +131,8 @@ class Agent:
                     f"ARGUMENTS = {Fore.CYAN}{arguments}{Style.RESET_ALL}",
                 )
                 print(
-                    "Enter 'y' to authorise command, 'y -N' to run N continuous "
-                    "commands, 'n' to exit program, or enter feedback for "
+                    "Enter 'y' to authorise command, 'y -N' to run N continuous commands, 's' to run self-feedback commands"
+                    "'n' to exit program, or enter feedback for "
                     f"{self.ai_name}...",
                     flush=True,
                 )
@@ -141,6 +143,24 @@ class Agent:
                     if console_input.lower().strip() == "y":
                         user_input = "GENERATE NEXT COMMAND JSON"
                         break
+                    elif console_input.lower().strip() == "s":
+                        logger.typewriter_log(
+                            "-=-=-=-=-=-=-= THOUGHTS, REASONING, PLAN AND CRITICISM WILL NOW BE VERIFIED BY AGENT -=-=-=-=-=-=-=",
+                            Fore.GREEN,
+                            "",
+                        )
+                        thoughts = assistant_reply_json.get("thoughts", {})
+                        self_feedback_resp = self.get_self_feedback(thoughts)
+                        logger.typewriter_log(
+                            f"SELF FEEDBACK: {self_feedback_resp}",
+                            Fore.YELLOW,
+                            "",
+                        )
+                        if self_feedback_resp[0].lower().strip() == "y":
+                            user_input = "GENERATE NEXT COMMAND JSON"
+                        else:
+                            user_input = self_feedback_resp
+                        break
                     elif console_input.lower().strip() == "":
                         print("Invalid input format.")
                         continue
@@ -245,3 +265,33 @@ class Agent:
                         self.workspace.get_path(command_args[pathlike])
                     )
         return command_args
+
+    @staticmethod
+    def get_self_feedback(thoughts: dict) -> str:
+        """Generates a feedback response based on the provided thoughts dictionary.
+        This method takes in a dictionary of thoughts containing keys such as 'reasoning',
+        'plan', 'thoughts', and 'criticism'. It combines these elements into a single
+        feedback message and uses the create_chat_completion() function to generate a
+        response based on the input message.
+        Args:
+            thoughts (dict): A dictionary containing thought elements like reasoning,
+            plan, thoughts, and criticism.
+        Returns:
+            str: A feedback response generated using the provided thoughts dictionary.
+        """
+
+        with open("ai_settings.yaml", "r") as yaml_file:
+            parsed_yaml = yaml.safe_load(yaml_file)
+            ai_role = parsed_yaml["ai_role"]
+
+        feedback_prompt = f"Below is a message from an AI agent with the role of {ai_role}. Please review the provided Thought, Reasoning, Plan, and Criticism. If these elements accurately contribute to the successful execution of the assumed role, respond with the letter 'Y' followed by a space, and then explain why it is effective. If the provided information is not suitable for achieving the role's objectives, please provide one or more sentences addressing the issue and suggesting a resolution."
+        reasoning = thoughts.get("reasoning", "")
+        plan = thoughts.get("plan", "")
+        thought = thoughts.get("thoughts", "")
+        criticism = thoughts.get("criticism", "")
+        feedback_thoughts = thought + reasoning + plan + criticism
+        feedback_response = create_chat_completion(
+            [{"role": "user", "content": feedback_prompt + feedback_thoughts}],
+            "gpt-3.5-turbo",
+        )  # * This hardcodes the model to use GPT3.5. should be an argument
+        return feedback_response

docs/usage.md

@@ -71,6 +71,10 @@ Use at your own risk.
 
 2. To exit the program, press Ctrl + C
 
+### ♻️ Self-Feedback Mode ⚠️
+
+Running Self-Feedback will **INCREASE** token use and thus cost more. This feature enables the agent to provide self-feedback by verifying its own actions and checking if they align with its current goals. If not, it will provide better feedback for the next loop. To enable this feature for the current loop, input `S` into the input field.
+
 ### GPT3.5 ONLY Mode
 
 If you don't have access to the GPT4 api, this mode will allow you to use Auto-GPT!