Make the json_parser more robust

For some reason the bot keeps prefacing its JSON. This fixes it for now.
2025-12-17 22:14:28 +01:00 · 2023-04-02 18:50:51 -05:00
parent a47da497b5
commit 3e587bc7fb
5 changed files with 160 additions and 58 deletions
--- a/scripts/ai_functions.py
+++ b/scripts/ai_functions.py
@@ -3,46 +3,23 @@ import json
 import openai
 import dirtyjson
 from config import Config
-
+from call_ai_function import call_ai_function
+from json_parser import fix_and_parse_json
 cfg = Config()

-# This is a magic function that can do anything with no-code. See
-# https://github.com/Torantulino/AI-Functions for more info.
-def call_ai_function(function, args, description, model=cfg.smart_llm_model):
-    # For each arg, if any are None, convert to "None":
-    args = [str(arg) if arg is not None else "None" for arg in args]
-    # parse args to comma seperated string
-    args = ", ".join(args)
-    messages = [
-        {
-            "role": "system",
-            "content": f"You are now the following python function: ```# {description}\n{function}```\n\nOnly respond with your `return` value.",
-        },
-        {"role": "user", "content": args},
-    ]
-
-    response = openai.ChatCompletion.create(
-        model=model, messages=messages, temperature=0
-    )
-
-    return response.choices[0].message["content"]
-
-
 # Evaluating code

-
 def evaluate_code(code: str) -> List[str]:
    function_string = "def analyze_code(code: str) -> List[str]:"
    args = [code]
    description_string = """Analyzes the given code and returns a list of suggestions for improvements."""

    result_string = call_ai_function(function_string, args, description_string)
-    return json.loads(result_string)
+    return fix_and_parse_json.loads(result_string)


 # Improving code

-
 def improve_code(suggestions: List[str], code: str) -> str:
    function_string = (
        "def generate_improved_code(suggestions: List[str], code: str) -> str:"
@@ -68,28 +45,3 @@ def write_tests(code: str, focus: List[str]) -> str:
    return result_string


-# TODO: Make debug a global config var
-def fix_json(json_str: str, schema:str = None, debug=True) -> str:
-    # Try to fix the JSON using gpt:
-    function_string = "def fix_json(json_str: str, schema:str=None) -> str:"
-    args = [json_str, schema]
-    description_string = """Fixes the provided JSON string to make it parseable. If the schema is provided, the JSON will be made to look like the schema, otherwise it will be made to look like a valid JSON object."""
-
-    result_string = call_ai_function(
-        function_string, args, description_string, model=cfg.fast_llm_model
-    )
-    if debug:
-        print("------------ JSON FIX ATTEMPT ---------------")
-        print(f"Original JSON: {json_str}")
-        print(f"Fixed JSON: {result_string}")
-        print("----------- END OF FIX ATTEMPT ----------------")
-    try:
-        return dirtyjson.loads(result_string)
-    except:
-        # Log the exception:
-        print("Failed to fix JSON")
-        # Get the call stack:
-        import traceback
-        call_stack = traceback.format_exc()
-        print(call_stack)
-        return {}
--- a/scripts/call_ai_function.py
+++ b/scripts/call_ai_function.py
@@ -0,0 +1,27 @@
+from typing import List, Optional
+import json
+import openai
+import dirtyjson
+from config import Config
+cfg = Config()
+
+# This is a magic function that can do anything with no-code. See
+# https://github.com/Torantulino/AI-Functions for more info.
+def call_ai_function(function, args, description, model=cfg.smart_llm_model):
+    # For each arg, if any are None, convert to "None":
+    args = [str(arg) if arg is not None else "None" for arg in args]
+    # parse args to comma seperated string
+    args = ", ".join(args)
+    messages = [
+        {
+            "role": "system",
+            "content": f"You are now the following python function: ```# {description}\n{function}```\n\nOnly respond with your `return` value.",
+        },
+        {"role": "user", "content": args},
+    ]
+
+    response = openai.ChatCompletion.create(
+        model=model, messages=messages, temperature=0
+    )
+
+    return response.choices[0].message["content"]
--- a/scripts/commands.py
+++ b/scripts/commands.py
@@ -8,12 +8,13 @@ from config import Config
 import ai_functions as ai
 from file_operations import read_file, write_to_file, append_to_file, delete_file
 from execute_code import execute_python_file
+from json_parser import fix_and_parse_json
 cfg = Config()


 def get_command(response):
    try:
-        response_json = json.loads(response)
+        response_json = fix_and_parse_json(response)
        command = response_json["command"]
        command_name = command["name"]
        arguments = command["args"]
--- a/scripts/json_parser.py
+++ b/scripts/json_parser.py
@@ -1,12 +1,53 @@
 import dirtyjson
-from ai_functions import fix_json
+from call_ai_function import call_ai_function
+from config import Config
+cfg = Config()

 def fix_and_parse_json(json_str: str, try_to_fix_with_gpt: bool = True):
    try:
        return dirtyjson.loads(json_str)
+    except Exception as e:
+        # Let's do something manually - sometimes GPT responds with something BEFORE the braces:
+        # "I'm sorry, I don't understand. Please try again."{"text": "I'm sorry, I don't understand. Please try again.", "confidence": 0.0}
+        # So let's try to find the first brace and then parse the rest of the string
+        try:
+          brace_index = json_str.index("{")
+          json_str = json_str[brace_index:]
+          last_brace_index = json_str.rindex("}")
+          json_str = json_str[:last_brace_index+1]
+          return dirtyjson.loads(json_str)
        except Exception as e:
          if try_to_fix_with_gpt:
            # Now try to fix this up using the ai_functions
            return fix_json(json_str, None, True)
          else:
            raise e
+        
+# TODO: Make debug a global config var
+def fix_json(json_str: str, schema:str = None, debug=True) -> str:
+    # Try to fix the JSON using gpt:
+    function_string = "def fix_json(json_str: str, schema:str=None) -> str:"
+    args = [json_str, schema]
+    description_string = """Fixes the provided JSON string to make it parseable. If the schema is provided, the JSON will be made to look like the schema, otherwise it will be made to look like a valid JSON object."""
+
+    # If it doesn't already start with a "`", add one:
+    if not json_str.startswith("`"):
+      json_str = "```json\n" + json_str + "\n```"
+    result_string = call_ai_function(
+        function_string, args, description_string, model=cfg.fast_llm_model
+    )
+    if debug:
+        print("------------ JSON FIX ATTEMPT ---------------")
+        print(f"Original JSON: {json_str}")
+        print(f"Fixed JSON: {result_string}")
+        print("----------- END OF FIX ATTEMPT ----------------")
+    try:
+        return dirtyjson.loads(result_string)
+    except:
+        # Log the exception:
+        print("Failed to fix JSON")
+        # Get the call stack:
+        import traceback
+        call_stack = traceback.format_exc()
+        print(call_stack)
+        return {}
--- a/tests/json_tests.py
+++ b/tests/json_tests.py
@@ -29,6 +29,87 @@ class TestParseJson(unittest.TestCase):
        with self.assertRaises(Exception):
            fix_and_parse_json(json_str, try_to_fix_with_gpt=False)

+    def test_invalid_json_leading_sentence_with_gpt(self):
+        # Test that a REALLY invalid JSON string raises an error when try_to_fix_with_gpt is False
+        json_str = """I suggest we start by browsing the repository to find any issues that we can fix.
+
+{
+    "command": {
+        "name": "browse_website",
+        "args":{
+            "url": "https://github.com/Torantulino/Auto-GPT"
+        }
+    },
+    "thoughts":
+    {
+        "text": "I suggest we start browsing the repository to find any issues that we can fix.",
+        "reasoning": "Browsing the repository will give us an idea of the current state of the codebase and identify any issues that we can address to improve the repo.",
+        "plan": "- Look through the repository to find any issues.\n- Investigate any issues to determine what needs to be fixed\n- Identify possible solutions to fix the issues\n- Open Pull Requests with fixes",
+        "criticism": "I should be careful while browsing so as not to accidentally introduce any new bugs or issues.",
+        "speak": "I will start browsing the repository to find any issues we can fix."
+    }
+}"""
+        good_obj = {
+          "command": {
+              "name": "browse_website",
+              "args":{
+                  "url": "https://github.com/Torantulino/Auto-GPT"
+              }
+          },
+          "thoughts":
+          {
+              "text": "I suggest we start browsing the repository to find any issues that we can fix.",
+              "reasoning": "Browsing the repository will give us an idea of the current state of the codebase and identify any issues that we can address to improve the repo.",
+              "plan": "- Look through the repository to find any issues.\n- Investigate any issues to determine what needs to be fixed\n- Identify possible solutions to fix the issues\n- Open Pull Requests with fixes",
+              "criticism": "I should be careful while browsing so as not to accidentally introduce any new bugs or issues.",
+              "speak": "I will start browsing the repository to find any issues we can fix."
+          }
+      }
+        # Assert that this raises an exception:
+        self.assertEqual(fix_and_parse_json(json_str, try_to_fix_with_gpt=False), good_obj)
+
+
+
+    def test_invalid_json_leading_sentence_with_gpt(self):
+        # Test that a REALLY invalid JSON string raises an error when try_to_fix_with_gpt is False
+        json_str = """I will first need to browse the repository (https://github.com/Torantulino/Auto-GPT) and identify any potential bugs that need fixing. I will use the "browse_website" command for this.
+
+{
+    "command": {
+        "name": "browse_website",
+        "args":{
+            "url": "https://github.com/Torantulino/Auto-GPT"
+        }
+    },
+    "thoughts":
+    {
+        "text": "Browsing the repository to identify potential bugs",
+        "reasoning": "Before fixing bugs, I need to identify what needs fixing. I will use the 'browse_website' command to analyze the repository.",
+        "plan": "- Analyze the repository for potential bugs and areas of improvement",
+        "criticism": "I need to ensure I am thorough and pay attention to detail while browsing the repository.",
+        "speak": "I am browsing the repository to identify potential bugs."
+    }
+}"""
+        good_obj = {
+    "command": {
+        "name": "browse_website",
+        "args":{
+            "url": "https://github.com/Torantulino/Auto-GPT"
+        }
+    },
+    "thoughts":
+    {
+        "text": "Browsing the repository to identify potential bugs",
+        "reasoning": "Before fixing bugs, I need to identify what needs fixing. I will use the 'browse_website' command to analyze the repository.",
+        "plan": "- Analyze the repository for potential bugs and areas of improvement",
+        "criticism": "I need to ensure I am thorough and pay attention to detail while browsing the repository.",
+        "speak": "I am browsing the repository to identify potential bugs."
+    }
+}
+        # Assert that this raises an exception:
+        self.assertEqual(fix_and_parse_json(json_str, try_to_fix_with_gpt=False), good_obj)
+
+

 if __name__ == '__main__':
    unittest.main()