[apps/Hal9] Multiple Tools loop in main, Add Arbitrary code execution…

… and other fixes (#470)

[apps/Hal9 Multiple Tools loop in main, Add Arbitrary code execution and other fixes]

apps/hal9/app.py

@@ -1,24 +1,22 @@
 from utils import generate_response, load_messages, insert_message, execute_function, save_messages, insert_tool_message, is_url, download_file, generate_text_embeddings_parquet
 from tools.calculator import solve_math_problem_description, solve_math_problem
-from tools.generic import answer_generic_question, answer_generic_question_description, answer_generic_future
+from tools.generic import answer_generic_question_description, answer_generic_question
 from tools.csv_agent import analyze_csv_description, analyze_csv
 from tools.image_agent import images_management_system, images_management_system_description, add_images_descriptions
 from tools.hal9 import answer_hal9_questions_description, answer_hal9_questions
 from tools.text_agent import analyze_text_file_description, analyze_text_file
-from tools.streamlit import streamlit_generator, streamlit_generator_description
-from tools.website import website_generator, website_generator_description
-from concurrent.futures import ThreadPoolExecutor
+from tools.other_tools import final_response_description, final_response
+from tools.python_execution import python_execution_description ,python_execution
 
 # load messages
 messages = load_messages()
 
 # load tools
-tools_descriptions = [answer_generic_question_description, solve_math_problem_description, analyze_csv_description, images_management_system_description, answer_hal9_questions_description, analyze_text_file_description, streamlit_generator_description, website_generator_description]
-tools_functions = [answer_generic_question, solve_math_problem, analyze_csv, images_management_system, answer_hal9_questions, analyze_text_file, streamlit_generator, website_generator]
+tools_descriptions = [python_execution_description, final_response_description, solve_math_problem_description, answer_generic_question_description, analyze_csv_description, images_management_system_description, answer_hal9_questions_description, analyze_text_file_description]
+tools_functions = [python_execution, final_response, solve_math_problem, answer_generic_question, analyze_csv, images_management_system, answer_hal9_questions, analyze_text_file]
 
 if len(messages) < 1:
-    messages = insert_message(messages, "system", "You are Hal9, a helpful and highly capable AI assistant. Your primary responsibility is to analyze user questions and select the most appropriate tool to provide precise, relevant, and actionable responses. Always prioritize using the right tool to ensure efficiency and clarity in your answers.")
-
+    messages = insert_message(messages, "system", "You are Hal9, a helpful and highly capable AI assistant. Your primary responsibility is to analyze user questions and select the most appropriate tool to provide precise, relevant, and actionable responses. Always prioritize using the right tool to ensure efficiency and clarity in your answers. If a tool is needed, follow these steps: 1. Identify the best tool for the task. 2. Execute the tool and process its response. 3. If the tool provides a valid result, return it to the user. 4. If the tool fails, do NOT retry with the same tool. Instead, explain the failure and suggest improvements in the prompt or alternative approaches.")
 user_input = input()
 if is_url(user_input):
     filename = user_input.split("/")[-1]
@@ -34,17 +32,16 @@
 else:
     user_input = user_input.replace("\f", "\n")
     messages = insert_message(messages, "user", user_input)
-
-    with ThreadPoolExecutor() as executor:
-        answer_generic_future = executor.submit(answer_generic_question, user_input)
-
-        response = executor.submit(generate_response, "openai", "gpt-4-turbo", messages, tools_descriptions, tool_choice="required", parallel_tool_calls=False)
-
-    response_result = response.result()
-
-    tool_result = execute_function(response_result, tools_functions)
-
-    if tool_result is not None:
-        insert_tool_message(messages, response_result, tool_result)
-
-save_messages(messages)
+    steps = 0
+    max_steps = 6
+    while steps < max_steps:
+        response = generate_response("openai", "gpt-4o-mini", messages, tools_descriptions, tool_choice = "required", parallel_tool_calls=False)
+        tool_result = execute_function(response, tools_functions)
+        insert_tool_message(messages, response, tool_result)
+        save_messages(messages, file_path="./.storage/.messages.json")
+        response_message = response.choices[0].message
+        tool_calls = getattr(response_message, 'tool_calls', None)
+        if tool_calls[0].function.name == "final_response":
+            break
+    if max_steps == steps:
+        print("Unable to generate a satisfactory response on time")

apps/hal9/tools/calculator.py

@@ -1,8 +1,4 @@
-
 def solve_math_problem(steps_explanation, code_solution):
-  print("Steps:\n")
-  print(steps_explanation)
-  print("\n\nPython Code:\n")
   exec(code_solution)
   return f"Steps:\n{steps_explanation}\n\n\nPython Code: {code_solution}"
 

apps/hal9/tools/csv_agent.py

@@ -201,7 +201,6 @@ def fix_python_code(csv_path, code):
         return f"An error has occurred again -> {last_line} ... Complete traceback: {tb}"
 
 def final_response(final_message):
-    print(final_message)
     return final_message
 
 ########################### Descriptions ########################## 

apps/hal9/tools/generic.py

@@ -1,26 +1,11 @@
 from groq import Groq
 from utils import load_messages, insert_message, save_messages
-from concurrent.futures import ThreadPoolExecutor
-
-answer_generic_future = None
 
 def answer_generic_question(user_input):
-    global answer_generic_future
-    if answer_generic_future is not None:
-        text_response=answer_generic_future.result()
-        print(text_response)
-        return text_response
-    else:
-        with ThreadPoolExecutor() as executor:
-            answer_generic_future = executor.submit(process_request, user_input)
-            text_response = answer_generic_future.result()
-            return text_response
-
-def process_request(user_input):
     messages = load_messages(file_path="./.storage/.generic_agent_messages.json")
     messages = insert_message(messages, "user", user_input)
     response = Groq().chat.completions.create(
-        model="llama3-70b-8192",
+        model="llama3-8b-8192",
         messages=messages,
         temperature=0,
         seed=1
@@ -41,7 +26,7 @@ def process_request(user_input):
             "properties": {
                 "user_input": {
                     "type": "string",
-                    "description": "Take the user input and pass the same string to the function",
+                    "description": "A generic question that requires to be answered",
                 },
             },
             "required": ["user_input"],

apps/hal9/tools/hal9.py

@@ -9,8 +9,7 @@ def answer_hal9_questions(user_input):
         temperature = 0,
         seed = 1,)
 
-    text_response = print(response.choices[0].message.content)
-    return text_response
+    return response.choices[0].message.content
 
 answer_hal9_questions_description = {
     "type": "function",

apps/hal9/tools/image_agent.py

@@ -80,7 +80,6 @@ def image_generator(prompt, filename):
 
     generated_description = add_images_descriptions(f".storage/{filename}")
 
-    print(generated_description)
     return f"The image generated is: {generated_description} \n\n Saved in path: '.storage/{filename}'"
 
 def image_analyzer(image_path, prompt):
@@ -100,7 +99,7 @@ def image_analyzer(image_path, prompt):
             }
         ],
     )
-    print(response.choices[0].message.content)
+
     return response.choices[0].message.content
 
 def edition_canny_model(image_path, modified_description, filename):
@@ -129,7 +128,7 @@ def edition_canny_model(image_path, modified_description, filename):
     shutil.copy(filename, f".storage/{filename}")
 
     generated_description = add_images_descriptions(f".storage/{filename}")
-    print(generated_description)
+
     return f"The image generated is: {generated_description} \n\n Saved in path: '.storage/{filename}'"
 
 def edition_depth_model(image_path, modified_description, filename):
@@ -157,7 +156,7 @@ def edition_depth_model(image_path, modified_description, filename):
     shutil.copy(filename, f".storage/{filename}")
 
     generated_description = add_images_descriptions(f".storage/{filename}")
-    print(generated_description)
+
     return f"The image generated is: {generated_description} \n\n Saved in path: '.storage/{filename}'"
 
 def generate_image_variation(image_path, filename):
@@ -187,7 +186,7 @@ def generate_image_variation(image_path, filename):
     shutil.copy(filename, f".storage/{filename}")
 
     generated_description = add_images_descriptions(f".storage/{filename}")
-    print(generated_description)
+
     return f"The image generated is: {generated_description} \n\n Saved in path: '.storage/{filename}'"
 
 ########################### Descriptions ########################## 

apps/hal9/tools/other_tools.py

@@ -0,0 +1,23 @@
+def final_response(final_message):
+    print(final_message)
+    return final_message
+
+final_response_description = { 
+    "type": "function",
+    "function": {
+        "name": "final_response",
+        "description": "This function is called when all necessary steeps has been completed through the tools, and all the requirements or the information requested by the user has been collected. It finalizes the process and delivers the results in a clear and concise message.",
+        "strict": True,
+        "parameters": {
+            "type": "object",
+            "properties": {
+                "final_message": {
+                    "type": "string",
+                    "description": "A clear and concise message that in simple terms mentions all the tools called to obtain the information neccesary. It explains how the information was gathered and whats the final insight.",
+                },
+            },
+            "required": ["final_message"],
+            "additionalProperties": False,
+        },
+    }
+}

apps/hal9/tools/python_execution.py

@@ -0,0 +1,92 @@
+from utils import generate_response, insert_message, extract_code_block
+import traceback
+from clients import openai_client
+import time
+import sys
+import io
+
+def fix_code(chat_input, error, complete_traceback, python_code):
+    stream = openai_client.chat.completions.create(
+      model =  "gpt-4-turbo",
+      messages = [
+        {"role": "user", "content": 
+f"""The following Python code needs to be fixed. It should fulfill this user request: '{chat_input}', return the fixed code as fenced code block with triple backticks (```) as ```python```
+
+### Error encountered:
+
+{error}
+
+### Code that needs fixing:
+
+{python_code}
+
+### Complete error traceback:
+
+{complete_traceback}
+"""
+    },]
+  )
+    return extract_code_block(stream.choices[0].message.content, "python")
+
+def debug_code(python_code):
+    try:
+        context = {}
+        timestamp = int(time.time())
+        with open(f"script_{timestamp}.txt", "w") as file:
+            file.write(python_code)
+        stdout_backup = sys.stdout
+        sys.stdout = io.StringIO()
+        exec(python_code, context)
+        output = sys.stdout.getvalue()
+        sys.stdout = stdout_backup
+        return f"Code Works properly: result of the execution: {output}", "", ""
+    except Exception as e:
+        tb = traceback.format_exc()
+        relevant_error_info = tb.splitlines()
+        last_line = relevant_error_info[-1]
+        complete_traceback="\n".join(relevant_error_info)
+        return "App fails to run", last_line, complete_traceback
+
+def python_execution(prompt):
+    # load messages
+    messages = []
+
+    if len(messages) < 1:
+        messages = insert_message(messages, "system", f"This Python generation system creates scripts from user prompts, including necessary imports; always include a print message indicating the process finished or the output. If a file is generated, it must be stored in './.storage/'. Return the code as a fenced block using triple backticks (```python```).")
+    messages = insert_message(messages, "user", f"Generates an app that fullfills this user request -> {prompt}")
+    model_response = generate_response("openai", "gpt-4-turbo", messages) 
+    response_content = model_response.choices[0].message.content
+    generated_code = extract_code_block(response_content, "python")
+    # Debug and fix the code if needed
+    max_tries = 3
+    tries = 0
+    while tries < max_tries:
+        result, error, complete_traceback = debug_code(generated_code)
+        if result.startswith("Code Works properly"):
+            messages = insert_message(messages, "assistant", generated_code)
+            return f"{result} ... This is the final code generated -> {generated_code}"
+        else:
+            generated_code = fix_code(prompt, error, complete_traceback, generated_code)
+        tries += 1
+
+    return f"Unable to generate an app that fulfills your request without errors. -> this was the final code obtained {generated_code} and the error found -> {error}"
+
+python_execution_description = {
+    "type": "function",
+    "function": {
+        "name": "python_execution",
+        "description": "Executes and debugs Python scripts to solve problems that cannot be addressed with other tools. It can run arbitrary Python code to generate files or perform complex calculations.",
+        "strict": True,
+        "parameters": {
+            "type": "object",
+            "properties": {
+                "prompt": {
+                    "type": "string",
+                    "description": "A detailed natural language description of the desired python code solution",
+                }
+            },
+            "required": ["prompt"],
+            "additionalProperties": False,
+        },
+    }
+}

apps/hal9/tools/streamlit.py

@@ -68,13 +68,11 @@ def streamlit_generator(prompt):
             save_python_code(streamlit_code)
             messages = insert_message(messages, "assistant", streamlit_code)
             save_messages(messages, file_path="./.storage/.streamlit_messages.json")
-            print(result)
             return result
         else:
             streamlit_code = fix_code(prompt, error, complete_traceback, streamlit_code)
         tries += 1
 
-    print("Unable to generate an app that fulfills your request without errors.")
     return "Unable to generate an app that fulfills your request without errors."
 
 streamlit_generator_description = {

apps/hal9/tools/text_agent.py

@@ -129,7 +129,6 @@ def random_pick_chunks(num_chunks, file_to_filter=None):
     return selected_chunks.to_dict(orient='records')
 
 def final_response(final_message):
-    print(final_message)
     return final_message
 
 ########################### Descriptions ########################## 

apps/hal9/tools/website.py

@@ -37,7 +37,6 @@ def website_generator(prompt):
   messages.append({"role": "user", "content": "briefly describe what was accomplished"})
   completion = client.chat.completions.create(model = "gpt-4", messages = messages)
   summary = h9.complete(completion, messages, show = False)
-  print(summary)
   return summary
 
 website_generator_description = {