Add agent feedback policy

src/sherpa_ai/actions/base.py

@@ -40,6 +40,9 @@ def __str__(self):
 
         return json.dumps(tool_desc, indent=4)
 
+    def __repr__(self):
+        return self.__str__()
+
 
 class BaseRetrievalAction(BaseAction, ABC):
     resources: list[ActionResource] = Field(default_factory=list)

src/sherpa_ai/memory/belief.py

@@ -68,9 +68,6 @@ def get_context(self, token_counter: Callable[[str], int], max_tokens=4000):
                 if token_counter(context) > max_tokens:
                     break
 
-        if context == "":
-            exit(1)
-
         return context
 
     def get_internal_history(

src/sherpa_ai/output_parsers/citation_validation.py

@@ -7,10 +7,9 @@
 from sherpa_ai.output_parsers.base import BaseOutputProcessor
 from sherpa_ai.output_parsers.validation_result import ValidationResult
 
-
 # download the punkt tokenizer. This is necessary for the sent_tokenize in NLTK.
 # The download will only happen once and the result will be cached.
-nltk.download("punkt")
+nltk.download("punkt_tab")
 
 
 class CitationValidation(BaseOutputProcessor):

src/sherpa_ai/policies/agent_feedback_policy.py

@@ -1,14 +1,23 @@
+from loguru import logger
+
 from sherpa_ai.agents.base import BaseAgent
 from sherpa_ai.events import Event, EventType
 from sherpa_ai.memory import Belief
 from sherpa_ai.policies import ReactPolicy
 
 AGENT_FEEDBACK_DESCRIPTION = """You are an intelligent assistant helping the user to complete their task. You have the following task to complete:
+{task}
+
+Context of you work is as follows:
+{context}
+
+You have the following options to continue completing the this task:
 
 {options}
 
-Use polite and engaging language, and ask the user what they want you to help them next.  Not need to greet and keep it short and concise. 
+Use polite and engaging language, and ask the user what they want you to help them next. Not need to greet and keep it short and concise. 
 
+If you think the context is not enough to choose an action. Ask the user for more information.
 """
 
 
@@ -26,12 +35,18 @@ class Config:
 
     def select_action(self, belief: Belief, **kwargs):
         actions = belief.actions
+
+        task = belief.current_task.content
+        context = belief.get_context(self.llm.get_num_tokens)
         options = "\n".join(
             [f"{i+1}. {action.name}" for i, action in enumerate(actions)]
         )
 
-        agent_feedback_prompt = self.agent_feedback_description.format(options=options)
+        agent_feedback_prompt = self.agent_feedback_description.format(
+            task=task, context=context, options=options
+        )
         question = self.llm.predict(agent_feedback_prompt)
+        logger.info(f"Question to the user: {question}")
         self.agent.shared_memory.add_event(Event(EventType.task, "Agent", question))
         result = self.agent.run()
 

src/tests/data/test_feedback_policy_in_qa_test_feedback_policy_in_qa_complete_0.jsonl

@@ -0,0 +1,7 @@
+{"input": [{"text": "You are an intelligent assistant helping the user to complete their task. You have the following task to complete:\\nWhat is the biggest?, Jupyter, Mars or Earth?\\n\\nContext of you work is as follows:\\nWhat is the biggest?, Jupyter, Mars or Earth?\\n\\n\\nYou have the following options to continue completing the this task:\\n\\n1. Google Search\\n\\nUse polite and engaging language, and ask the user what they want you to help them next. Not need to greet and keep it short and concise. \\n\\nIf you think the context is not enough to choose an action. Ask the user for more information.\\n", "agent": "human"}], "output": "To clarify, are you asking about the size of Jupiter, Mars, and Earth in terms of diameter, volume, or mass? Let me know how you'd like to proceed!", "llm_name": "gpt-4o-mini"}
+{"input": [{"text": "\\n\\n\\n\\n**Task Description**: What is the biggest?, Jupyter, Mars or Earth?\\n\\n**Possible Actions**:\\n{\\n    \"name\": \"Google Search\",\\n    \"args\": {\\n        \"query\": \"string\"\\n    },\\n    \"usage\": \"Get answers from Google Search\"\\n}\\n\\n**Task Context**:\\nWhat is the biggest?, Jupyter, Mars or Earth?\\nGoogle search the question\\n\\n\\n**History of Previous Actions**:\\n\\n\\nYou should only respond in JSON format as described below without any extra text.\\nResponse Format:\\n{\\n    \"command\": {\\n        \"name\": \"tool/command name you choose\",\\n        \"args\": {\\n            \"arg name\": \"value\"\\n        }\\n    }\\n}\\nEnsure the response can be parsed by Python json.loads\\n\\nFollow the described format strictly.\\n\\n", "agent": "human"}], "output": "{\\n    \"command\": {\\n        \"name\": \"Google Search\",\\n        \"args\": {\\n            \"query\": \"What is the biggest planet Jupiter Mars or Earth?\"\\n        }\\n    }\\n}", "llm_name": "gpt-4o-mini"}
+{"input": [{"text": "You are an intelligent assistant helping the user to complete their task. You have the following task to complete:\\nWhat is the biggest?, Jupyter, Mars or Earth?\\n\\nContext of you work is as follows:\\nWhat is the biggest?, Jupyter, Mars or Earth?\\nGoogle search the question\\n\\n\\nYou have the following options to continue completing the this task:\\n\\n1. Google Search\\n\\nUse polite and engaging language, and ask the user what they want you to help them next. Not need to greet and keep it short and concise. \\n\\nIf you think the context is not enough to choose an action. Ask the user for more information.\\n", "agent": "human"}], "output": "Would you like me to perform a Google search to find out which is the biggest: Jupiter, Mars, or Earth?", "llm_name": "gpt-4o-mini"}
+{"input": [{"text": "\\n\\n\\n\\n**Task Description**: What is the biggest?, Jupyter, Mars or Earth?\\n\\n**Possible Actions**:\\n{\\n    \"name\": \"Google Search\",\\n    \"args\": {\\n        \"query\": \"string\"\\n    },\\n    \"usage\": \"Get answers from Google Search\"\\n}\\n\\n**Task Context**:\\nWhat is the biggest?, Jupyter, Mars or Earth?\\nGoogle search the question\\nGoogle search the question\\n\\n\\n**History of Previous Actions**:\\nGoogle Search{'query': 'What is the biggest planet Jupiter Mars or Earth?'}\\nJupiter is the biggest planet\\n\\nYou should only respond in JSON format as described below without any extra text.\\nResponse Format:\\n{\\n    \"command\": {\\n        \"name\": \"tool/command name you choose\",\\n        \"args\": {\\n            \"arg name\": \"value\"\\n        }\\n    }\\n}\\nEnsure the response can be parsed by Python json.loads\\n\\nFollow the described format strictly.\\n\\n", "agent": "human"}], "output": "{\\n    \"command\": {\\n        \"name\": \"Google Search\",\\n        \"args\": {\\n            \"query\": \"What is the biggest planet Jupiter Mars or Earth?\"\\n        }\\n    }\\n}", "llm_name": "gpt-4o-mini"}
+{"input": [{"text": "You are an intelligent assistant helping the user to complete their task. You have the following task to complete:\\nWhat is the biggest?, Jupyter, Mars or Earth?\\n\\nContext of you work is as follows:\\nWhat is the biggest?, Jupyter, Mars or Earth?\\nGoogle search the question\\nGoogle search the question\\n\\n\\nYou have the following options to continue completing the this task:\\n\\n1. Google Search\\n\\nUse polite and engaging language, and ask the user what they want you to help them next. Not need to greet and keep it short and concise. \\n\\nIf you think the context is not enough to choose an action. Ask the user for more information.\\n", "agent": "human"}], "output": "Would you like me to perform a Google search to find out which is the biggest: Jupiter, Mars, or Earth?", "llm_name": "gpt-4o-mini"}
+{"input": [{"text": "\\n\\n\\n\\n**Task Description**: What is the biggest?, Jupyter, Mars or Earth?\\n\\n**Possible Actions**:\\n{\\n    \"name\": \"Google Search\",\\n    \"args\": {\\n        \"query\": \"string\"\\n    },\\n    \"usage\": \"Get answers from Google Search\"\\n}\\n\\n**Task Context**:\\nWhat is the biggest?, Jupyter, Mars or Earth?\\nGoogle search the question\\nGoogle search the question\\nGoogle search the question\\n\\n\\n**History of Previous Actions**:\\nGoogle Search{'query': 'What is the biggest planet Jupiter Mars or Earth?'}\\nJupiter is the biggest planet\\nGoogle Search{'query': 'What is the biggest planet Jupiter Mars or Earth?'}\\nJupiter is the biggest planet\\n\\nYou should only respond in JSON format as described below without any extra text.\\nResponse Format:\\n{\\n    \"command\": {\\n        \"name\": \"tool/command name you choose\",\\n        \"args\": {\\n            \"arg name\": \"value\"\\n        }\\n    }\\n}\\nEnsure the response can be parsed by Python json.loads\\n\\nFollow the described format strictly.\\n\\n", "agent": "human"}], "output": "{\\n    \"command\": {\\n        \"name\": \"Google Search\",\\n        \"args\": {\\n            \"query\": \"What is the biggest planet Jupiter Mars or Earth?\"\\n        }\\n    }\\n}", "llm_name": "gpt-4o-mini"}
+{"input": [{"text": "You are a **question answering assistant** who solves user questions and offers a detailed solution.\\n\\nYour name is QA Agent.\\n\\nContext: What is the biggest?, Jupyter, Mars or Earth?\\nGoogle search the question\\nGoogle search the question\\nGoogle search the question\\n\\n\\nAction - Result History:\\nGoogle Search{'query': 'What is the biggest planet Jupiter Mars or Earth?'}\\nJupiter is the biggest planet\\nGoogle Search{'query': 'What is the biggest planet Jupiter Mars or Earth?'}\\nJupiter is the biggest planet\\nGoogle Search{'query': 'What is the biggest planet Jupiter Mars or Earth?'}\\nJupiter is the biggest planet\\n\\nGiven the context and the action-result history, please complete the task mentioned. Include any links you used from the context and history in the result.\\nTask: What is the biggest?, Jupyter, Mars or Earth?\\nResult:\\n", "agent": "human"}], "output": "The biggest among Jupiter, Mars, and Earth is **Jupiter**. \\n\\nJupiter is the largest planet in our solar system, significantly larger than both Mars and Earth. To provide some context:\\n\\n- **Jupiter** has a diameter of about 86,881 miles (139,822 kilometers) and is primarily composed of gas, making it a gas giant.\\n- **Earth** has a diameter of about 7,918 miles (12,742 kilometers).\\n- **Mars** has a diameter of about 4,221 miles (6,779 kilometers).\\n\\nThus, in terms of size, Jupiter is the clear winner. \\n\\nFor further information, you can refer to the following links:\\n- [NASA - Jupiter](https://solarsystem.nasa.gov/planets/jupiter/overview/)\\n- [NASA - Mars](https://solarsystem.nasa.gov/planets/mars/overview/)\\n- [NASA - Earth](https://solarsystem.nasa.gov/planets/earth/overview/) \\n\\nIf you have any more questions or need further clarification, feel free to ask!", "llm_name": "gpt-4o-mini"}

src/tests/data/test_feedback_policy_in_qa_test_feedback_policy_in_qa_incomplete_0.jsonl

@@ -0,0 +1,7 @@
+{"input": [{"text": "You are an intelligent assistant helping the user to complete their task. You have the following task to complete:\\nWhat is the biggest ?\\n\\nContext of you work is as follows:\\nWhat is the biggest ?\\n\\n\\nYou have the following options to continue completing the this task:\\n\\n1. Google Search\\n\\nUse polite and engaging language, and ask the user what they want you to help them next. Not need to greet and keep it short and concise. \\n\\nIf you think the context is not enough to choose an action. Ask the user for more information.\\n", "agent": "human"}], "output": "Could you please provide more details about what you're looking for? What specific topic or item do you want to know the biggest of?", "llm_name": "gpt-4o-mini"}
+{"input": [{"text": "\\n\\n\\n\\n**Task Description**: What is the biggest ?\\n\\n**Possible Actions**:\\n{\\n    \"name\": \"Google Search\",\\n    \"args\": {\\n        \"query\": \"string\"\\n    },\\n    \"usage\": \"Get answers from Google Search\"\\n}\\n\\n**Task Context**:\\nWhat is the biggest ?\\nWhich planet is biggest, Earth, Mars or Jupiter?\\n\\n\\n**History of Previous Actions**:\\n\\n\\nYou should only respond in JSON format as described below without any extra text.\\nResponse Format:\\n{\\n    \"command\": {\\n        \"name\": \"tool/command name you choose\",\\n        \"args\": {\\n            \"arg name\": \"value\"\\n        }\\n    }\\n}\\nEnsure the response can be parsed by Python json.loads\\n\\nFollow the described format strictly.\\n\\n", "agent": "human"}], "output": "{\\n    \"command\": {\\n        \"name\": \"Google Search\",\\n        \"args\": {\\n            \"query\": \"Which planet is the biggest: Earth, Mars, or Jupiter?\"\\n        }\\n    }\\n}", "llm_name": "gpt-4o-mini"}
+{"input": [{"text": "You are an intelligent assistant helping the user to complete their task. You have the following task to complete:\\nWhat is the biggest ?\\n\\nContext of you work is as follows:\\nWhat is the biggest ?\\nWhich planet is biggest, Earth, Mars or Jupiter?\\n\\n\\nYou have the following options to continue completing the this task:\\n\\n1. Google Search\\n\\nUse polite and engaging language, and ask the user what they want you to help them next. Not need to greet and keep it short and concise. \\n\\nIf you think the context is not enough to choose an action. Ask the user for more information.\\n", "agent": "human"}], "output": "Jupiter is the biggest planet among Earth, Mars, and itself. Would you like to know more about the planets or anything else?", "llm_name": "gpt-4o-mini"}
+{"input": [{"text": "\\n\\n\\n\\n**Task Description**: What is the biggest ?\\n\\n**Possible Actions**:\\n{\\n    \"name\": \"Google Search\",\\n    \"args\": {\\n        \"query\": \"string\"\\n    },\\n    \"usage\": \"Get answers from Google Search\"\\n}\\n\\n**Task Context**:\\nWhat is the biggest ?\\nWhich planet is biggest, Earth, Mars or Jupiter?\\nWhich planet is biggest, Earth, Mars or Jupiter?\\n\\n\\n**History of Previous Actions**:\\nGoogle Search{'query': 'Which planet is the biggest: Earth, Mars, or Jupiter?'}\\nJupiter is the biggest planet\\n\\nYou should only respond in JSON format as described below without any extra text.\\nResponse Format:\\n{\\n    \"command\": {\\n        \"name\": \"tool/command name you choose\",\\n        \"args\": {\\n            \"arg name\": \"value\"\\n        }\\n    }\\n}\\nEnsure the response can be parsed by Python json.loads\\n\\nFollow the described format strictly.\\n\\n", "agent": "human"}], "output": "{\\n    \"command\": {\\n        \"name\": \"Google Search\",\\n        \"args\": {\\n            \"query\": \"Which planet is the biggest: Earth, Mars, or Jupiter?\"\\n        }\\n    }\\n}", "llm_name": "gpt-4o-mini"}
+{"input": [{"text": "You are an intelligent assistant helping the user to complete their task. You have the following task to complete:\\nWhat is the biggest ?\\n\\nContext of you work is as follows:\\nWhat is the biggest ?\\nWhich planet is biggest, Earth, Mars or Jupiter?\\nWhich planet is biggest, Earth, Mars or Jupiter?\\n\\n\\nYou have the following options to continue completing the this task:\\n\\n1. Google Search\\n\\nUse polite and engaging language, and ask the user what they want you to help them next. Not need to greet and keep it short and concise. \\n\\nIf you think the context is not enough to choose an action. Ask the user for more information.\\n", "agent": "human"}], "output": "Jupiter is the biggest planet among Earth, Mars, and Jupiter. How can I assist you further?", "llm_name": "gpt-4o-mini"}
+{"input": [{"text": "\\n\\n\\n\\n**Task Description**: What is the biggest ?\\n\\n**Possible Actions**:\\n{\\n    \"name\": \"Google Search\",\\n    \"args\": {\\n        \"query\": \"string\"\\n    },\\n    \"usage\": \"Get answers from Google Search\"\\n}\\n\\n**Task Context**:\\nWhat is the biggest ?\\nWhich planet is biggest, Earth, Mars or Jupiter?\\nWhich planet is biggest, Earth, Mars or Jupiter?\\nWhich planet is biggest, Earth, Mars or Jupiter?\\n\\n\\n**History of Previous Actions**:\\nGoogle Search{'query': 'Which planet is the biggest: Earth, Mars, or Jupiter?'}\\nJupiter is the biggest planet\\nGoogle Search{'query': 'Which planet is the biggest: Earth, Mars, or Jupiter?'}\\nJupiter is the biggest planet\\n\\nYou should only respond in JSON format as described below without any extra text.\\nResponse Format:\\n{\\n    \"command\": {\\n        \"name\": \"tool/command name you choose\",\\n        \"args\": {\\n            \"arg name\": \"value\"\\n        }\\n    }\\n}\\nEnsure the response can be parsed by Python json.loads\\n\\nFollow the described format strictly.\\n\\n", "agent": "human"}], "output": "{\\n    \"command\": {\\n        \"name\": \"Google Search\",\\n        \"args\": {\\n            \"query\": \"Which planet is the biggest: Earth, Mars, or Jupiter?\"\\n        }\\n    }\\n}", "llm_name": "gpt-4o-mini"}
+{"input": [{"text": "You are a **question answering assistant** who solves user questions and offers a detailed solution.\\n\\nYour name is QA Agent.\\n\\nContext: What is the biggest ?\\nWhich planet is biggest, Earth, Mars or Jupiter?\\nWhich planet is biggest, Earth, Mars or Jupiter?\\nWhich planet is biggest, Earth, Mars or Jupiter?\\n\\n\\nAction - Result History:\\nGoogle Search{'query': 'Which planet is the biggest: Earth, Mars, or Jupiter?'}\\nJupiter is the biggest planet\\nGoogle Search{'query': 'Which planet is the biggest: Earth, Mars, or Jupiter?'}\\nJupiter is the biggest planet\\nGoogle Search{'query': 'Which planet is the biggest: Earth, Mars, or Jupiter?'}\\nJupiter is the biggest planet\\n\\nGiven the context and the action-result history, please complete the task mentioned. Include any links you used from the context and history in the result.\\nTask: What is the biggest ?\\nResult:\\n", "agent": "human"}], "output": "The biggest planet among Earth, Mars, and Jupiter is Jupiter. \\n\\nJupiter is the largest planet in our solar system, significantly larger than both Earth and Mars. To provide some context:\\n\\n- **Jupiter**: It has a diameter of about 86,881 miles (139,822 kilometers) and is primarily composed of gas, making it a gas giant.\\n- **Earth**: The diameter of Earth is about 7,917.5 miles (12,742 kilometers).\\n- **Mars**: Mars has a diameter of about 4,212 miles (6,779 kilometers).\\n\\nIn summary, Jupiter is the largest of the three planets mentioned. \\n\\nFor further information, you can refer to resources like NASA's planetary fact sheets or educational websites about the solar system.", "llm_name": "gpt-4o-mini"}