Merge pull request #763 from TransformerOptimus/json_schema_changes

moving json structure to json schema structure in agent

superagi/agent/output_parser.py

@@ -2,6 +2,7 @@
 from abc import ABC, abstractmethod
 from typing import Dict, NamedTuple, List
 import re
+import ast
 import json5
 from superagi.helper.json_cleaner import JsonCleaner
 from superagi.lib.logger import logger
@@ -22,7 +23,21 @@ class BaseOutputParser(ABC):
     def parse(self, text: str) -> AgentGPTAction:
         """Return AgentGPTAction"""
 
-
+class AgentSchemaOutputParser(BaseOutputParser):
+    def parse(self, response: str) -> AgentGPTAction:
+        if response.startswith("```") and response.endswith("```"):
+            response = "```".join(response.split("```")[1:-1])
+        # OpenAI returns `str(content_dict)`, literal_eval reverses this
+        try:
+            logger.debug("AgentSchemaOutputParser: ", response)
+            response_obj = ast.literal_eval(response)
+            return AgentGPTAction(
+                name=response_obj['tool']['name'],
+                args=response_obj['tool']['args'],
+            )
+        except BaseException as e:
+            logger.info(f"AgentSchemaOutputParser: Error parsing JSON respons {e}")
+            return {}
 
 class AgentOutputParser(BaseOutputParser):
     def parse(self, text: str) -> AgentGPTAction:
@@ -95,4 +110,4 @@ def parse_tasks(self, text: str) -> AgentTasks:
             # If the command is null or incomplete, return an erroneous tool
             return AgentTasks(
                 error=f"Incomplete tool args: {parsed}",
-            )
+            )

superagi/agent/prompts/analyse_task.txt

@@ -15,12 +15,34 @@ Ensure next action tool is picked from the below tool list.
 TOOLS:
 {tools}
 
-RESPONSE FORMAT:
+Respond with only valid JSON conforming to the following schema:
 {
-    "thoughts": {
-        "reasoning": "short reasoning"
-    },
-    "tool": {"name": "tool name", "args": {"arg name": "arg value(escape in case of string)"}}
-}
-
-Your answer must be something that JSON.parse() can read, and nothing else.
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "type": "object",
+    "properties": {
+        "thoughts": {
+            "type": "object",
+            "properties": {
+                "reasoning": {
+                    "type": "string",
+                    "description": "short reasoning",
+                }
+            },
+            "required": ["reasoning"]
+        },
+        "tool": {
+            "type": "object",
+            "properties": {
+                "name": {
+                    "type": "string",
+                    "description": "tool name",
+                },
+                "args": {
+                    "type": "object",
+                    "description": "tool arguments",
+                }
+            },
+            "required": ["name", "args"]
+        }
+    }
+}

superagi/agent/prompts/initialize_tasks.txt

@@ -5,7 +5,7 @@ GOALS:
 
 {task_instructions}
 
-Construct a sequence of actions, not exceeding 4 steps, to achieve this goal.
+Construct a sequence of actions, not exceeding 3 steps, to achieve this goal.
 
 Submit your response as a formatted ARRAY of strings, suitable for utilization with JSON.parse().
 

superagi/agent/prompts/superagi.txt

@@ -20,9 +20,54 @@ PERFORMANCE EVALUATION:
 4. Reflect on past decisions and strategies to refine your approach.
 5. Every tool has a cost, so be smart and efficient.
 
-I should only respond in JSON format as described below.
-Response Format:
-{response_format}
-
-
-Ensure the response can be parsed by Python json.loads.
+Respond with only valid JSON conforming to the following schema:
+{
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "type": "object",
+    "properties": {
+        "thoughts": {
+            "type": "object",
+            "properties": {
+                "text": {
+                    "type": "string",
+                    "description": "thought"
+                },
+                "reasoning": {
+                    "type": "string",
+                    "description": "short reasoning"
+                },
+                "plan": {
+                    "type": "string",
+                    "description": "- short bulleted\n- list that conveys\n- long-term plan"
+                },
+                "criticism": {
+                    "type": "string",
+                    "description": "constructive self-criticism"
+                },
+                "speak": {
+                    "type": "string",
+                    "description": "thoughts summary to say to user"
+                }
+            },
+            "required": ["text", "reasoning", "plan", "criticism", "speak"],
+            "additionalProperties": false
+        },
+        "tool": {
+            "type": "object",
+            "properties": {
+                "name": {
+                    "type": "string",
+                    "description": "tool name"
+                },
+                "args": {
+                    "type": "object",
+                    "description": "tool arguments"
+                }
+            },
+            "required": ["name", "args"],
+            "additionalProperties": false
+        }
+    },
+    "required": ["thoughts", "tool"],
+    "additionalProperties": false
+}

superagi/agent/super_agi.py

@@ -4,37 +4,32 @@
 from __future__ import annotations
 
 import time
-from typing import Any, Dict
+from typing import Any
 from typing import Tuple
+import json
 import numpy as np
-from halo import Halo
 from pydantic import ValidationError
 from pydantic.types import List
-from sqlalchemy import desc, asc
-from sqlalchemy.exc import SQLAlchemyError
+from sqlalchemy import asc
 from sqlalchemy.orm import sessionmaker
 
 from superagi.agent.agent_prompt_builder import AgentPromptBuilder
-from superagi.agent.output_parser import BaseOutputParser, AgentOutputParser
+from superagi.agent.output_parser import BaseOutputParser, AgentOutputParser, AgentSchemaOutputParser
 from superagi.agent.task_queue import TaskQueue
+from superagi.apm.event_handler import EventHandler
 from superagi.helper.token_counter import TokenCounter
+from superagi.lib.logger import logger
 from superagi.llms.base_llm import BaseLlm
-from superagi.models.agent_config import AgentConfiguration
+from superagi.models.agent import Agent
 from superagi.models.agent_execution import AgentExecution
 # from superagi.models.types.agent_with_config import AgentWithConfig
 from superagi.models.agent_execution_feed import AgentExecutionFeed
 from superagi.models.agent_execution_permission import AgentExecutionPermission
 from superagi.models.agent_workflow_step import AgentWorkflowStep
 from superagi.models.db import connect_db
 from superagi.tools.base_tool import BaseTool
-from superagi.types.common import BaseMessage, HumanMessage, AIMessage, SystemMessage
+from superagi.types.common import BaseMessage
 from superagi.vector_store.base import VectorStore
-from superagi.models.agent import Agent
-from superagi.models.resource import Resource
-from superagi.config.config import get_config
-from superagi.apm.event_handler import EventHandler
-import os
-from superagi.lib.logger import logger
 
 FINISH = "finish"
 WRITE_FILE = "Write File"
@@ -59,7 +54,7 @@ def __init__(self,
                  tools: List[BaseTool],
                  agent_config: Any,
                  agent_execution_config: Any,
-                 output_parser: BaseOutputParser = AgentOutputParser(),
+                 output_parser: BaseOutputParser = AgentSchemaOutputParser(),
                  ):
         self.ai_name = ai_name
         self.ai_role = ai_role
@@ -70,8 +65,6 @@ def __init__(self,
         self.tools = tools
         self.agent_config = agent_config
         self.agent_execution_config = agent_execution_config
-        # Init Log
-        # print("\033[92m\033[1m" + "\nWelcome to SuperAGI - The future of AGI" + "\033[0m\033[0m")
 
     @classmethod
     def from_llm_and_tools(
@@ -144,9 +137,8 @@ def execute(self, workflow_step: AgentWorkflowStep):
             #                                           agent_id=self.agent_config["agent_id"], feed=template_step.prompt,
             #                                           role="user")
 
-        logger.info(messages)
+        logger.debug("Prompt messages:", messages)
         if len(agent_feeds) <= 0:
-            logger.info(prompt)
             for message in messages:
                 agent_execution_feed = AgentExecutionFeed(agent_execution_id=self.agent_config["agent_execution_id"],
                                                           agent_id=self.agent_config["agent_id"],
@@ -160,7 +152,7 @@ def execute(self, workflow_step: AgentWorkflowStep):
         current_calls = current_calls + 1
         total_tokens = current_tokens + TokenCounter.count_message_tokens(response, self.llm.get_model())
         self.update_agent_execution_tokens(current_calls, total_tokens)
-        logger.info("Response:", response)
+        logger.debug("Response:", response)
         if 'content' not in response or response['content'] is None:
             raise RuntimeError(f"Failed to get response from llm")
         assistant_reply = response['content']
@@ -175,7 +167,8 @@ def execute(self, workflow_step: AgentWorkflowStep):
             tool_response = self.handle_tool_response(assistant_reply)
 
             agent_execution_feed = AgentExecutionFeed(agent_execution_id=self.agent_config["agent_execution_id"],
-                                                      agent_id=self.agent_config["agent_id"], feed=assistant_reply,
+                                                      agent_id=self.agent_config["agent_id"],
+                                                      feed=assistant_reply,
                                                       role="assistant")
             session.add(agent_execution_feed)
             tool_response_feed = AgentExecutionFeed(agent_execution_id=self.agent_config["agent_execution_id"],