Added templates, improved --local

Author: KillianLucas

.gitignore

@@ -1,3 +1,5 @@
+llama.log
+
 # Byte-compiled / optimized / DLL files
 __pycache__/
 *.py[cod]

interpreter/core/core.py

@@ -67,13 +67,17 @@ def __init__(
         llm=None,
         system_message=default_system_message,
         custom_instructions="",
+        user_message_template="{content}",
+        code_output_template="Code output: {content}\n\nWhat does this output mean / what's next (if anything, or are we done)?",
+        empty_code_output_template="The code above was executed on my machine. It produced no text output. what's next (if anything, or are we done?)",
+        code_output_sender="user",
         computer=None,
         sync_computer=False,
         import_computer_api=False,
         skills_path=None,
         import_skills=False,
         multi_line=False,
-        contribute_conversation=False
+        contribute_conversation=False,
     ):
         # State
         self.messages = [] if messages is None else messages
@@ -124,6 +128,10 @@ def __init__(
         # These are LLM related
         self.system_message = system_message
         self.custom_instructions = custom_instructions
+        self.user_message_template = user_message_template
+        self.code_output_template = code_output_template
+        self.empty_code_output_template = empty_code_output_template
+        self.code_output_sender = code_output_sender
 
     def server(self, *args, **kwargs):
         server(self, *args, **kwargs)
@@ -140,7 +148,9 @@ def anonymous_telemetry(self) -> bool:
 
     @property
     def will_contribute(self):
-        overrides = self.offline or not self.conversation_history or self.disable_telemetry
+        overrides = (
+            self.offline or not self.conversation_history or self.disable_telemetry
+        )
         return self.contribute_conversation and not overrides
 
     def chat(self, message=None, display=True, stream=False, blocking=True):
@@ -245,7 +255,9 @@ def _streaming_chat(self, message=None, display=True):
                 # If it's the first message, set the conversation name
                 if not self.conversation_filename:
                     first_few_words_list = self.messages[0]["content"][:25].split(" ")
-                    if len(first_few_words_list) >= 2:  # for languages like English with blank between words
+                    if (
+                        len(first_few_words_list) >= 2
+                    ):  # for languages like English with blank between words
                         first_few_words = "_".join(first_few_words_list[:-1])
                     else:  # for languages like Chinese without blank between words
                         first_few_words = self.messages[0]["content"][:15]

interpreter/core/llm/llm.py

@@ -28,11 +28,13 @@ def __init__(self, interpreter):
         self.model = "gpt-4-turbo"
         self.temperature = 0
 
-        self.supports_vision = None # Will try to auto-detect
-        self.vision_renderer = self.interpreter.computer.vision.query # Will only use if supports_vision is False
+        self.supports_vision = None  # Will try to auto-detect
+        self.vision_renderer = (
+            self.interpreter.computer.vision.query
+        )  # Will only use if supports_vision is False
 
-        self.supports_functions = None # Will try to auto-detect
-        self.execution_instructions = "To execute code on the user's machine, write a markdown code block. Specify the language after the ```. You will receive the output. Use any programming language." # If supports_functions is False, this will be added to the system message
+        self.supports_functions = None  # Will try to auto-detect
+        self.execution_instructions = "To execute code on the user's machine, write a markdown code block. Specify the language after the ```. You will receive the output. Use any programming language."  # If supports_functions is False, this will be added to the system message
 
         # Optional settings
         self.context_window = None
@@ -81,7 +83,7 @@ def run(self, messages):
                     self.supports_vision = False
             except:
                 self.supports_vision = False
-            
+
         # Trim image messages if they're there
         image_messages = [msg for msg in messages if msg["type"] == "image"]
         if self.supports_vision:
@@ -103,15 +105,19 @@ def run(self, messages):
         elif self.supports_vision == False and self.vision_renderer:
             for img_msg in image_messages:
                 if img_msg["format"] != "description":
-                        img_msg["content"] = "Imagine I have just shown you an image with this description: " + self.vision_renderer(lmc=img_msg)
-                        img_msg["format"] = "description"
+                    img_msg["content"] = (
+                        "Imagine I have just shown you an image with this description: "
+                        + self.vision_renderer(lmc=img_msg)
+                    )
+                    img_msg["format"] = "description"
 
         # Convert to OpenAI messages format
         messages = convert_to_openai_messages(
             messages,
             function_calling=self.supports_functions,
             vision=self.supports_vision,
             shrink_images=self.interpreter.shrink_images,
+            interpreter=self.interpreter,
         )
 
         system_message = messages[0]["content"]

interpreter/core/llm/utils/convert_to_openai_messages.py

@@ -10,7 +10,7 @@ def convert_to_openai_messages(
     function_calling=True,
     vision=False,
     shrink_images=True,
-    code_output_sender="assistant",
+    interpreter=None,
 ):
     """
     Converts LMC messages into OpenAI messages
@@ -27,7 +27,7 @@ def convert_to_openai_messages(
     #                 message["type"] = "message"
     #                 message["content"] = "```" + message.get("format", "") + "\n" + message.get("content").strip("\n`") + "\n```"
     #         prev_message = message
-        
+
     #     messages = [message for message in messages if message.get("type") != "code"]
 
     for message in messages:
@@ -41,7 +41,15 @@ def convert_to_openai_messages(
             new_message["role"] = message[
                 "role"
             ]  # This should never be `computer`, right?
-            new_message["content"] = message["content"]
+
+            if (
+                message["role"] == "user" and message == messages[-1]
+            ):  # Only add the template for the last message?
+                new_message["content"] = interpreter.user_message_template.replace(
+                    "{content}", message["content"]
+                )
+            else:
+                new_message["content"] = message["content"]
 
         elif message["type"] == "code":
             new_message["role"] = "assistant"
@@ -79,19 +87,17 @@ def convert_to_openai_messages(
 
             else:
                 # This should be experimented with.
-                if code_output_sender == "user":
+                if interpreter.code_output_sender == "user":
                     if message["content"].strip() == "":
-                        content = "The code above was executed on my machine. It produced no text output. what's next (if anything, or are we done?)"
+                        content = interpreter.empty_code_output_template
                     else:
-                        content = (
-                            "Code output: "
-                            + message["content"]
-                            + "\n\nWhat does this output mean / what's next (if anything, or are we done)?"
+                        content = interpreter.code_output_template.replace(
+                            "{content}", message["content"]
                         )
 
                     new_message["role"] = "user"
                     new_message["content"] = content
-                elif code_output_sender == "assistant":
+                elif interpreter.code_output_sender == "assistant":
                     if "@@@SEND_MESSAGE_AS_USER@@@" in message["content"]:
                         new_message["role"] = "user"
                         new_message["content"] = message["content"].replace(
@@ -136,7 +142,9 @@ def convert_to_openai_messages(
                             # Convert the image back to base64
                             buffered = io.BytesIO()
                             img.save(buffered, format=extension)
-                            img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")
+                            img_str = base64.b64encode(buffered.getvalue()).decode(
+                                "utf-8"
+                            )
                             content = f"data:image/{extension};base64,{img_str}"
                         except:
                             # This should be non blocking. It's not required
@@ -149,7 +157,9 @@ def convert_to_openai_messages(
                     file_extension = image_path.split(".")[-1]
 
                     with open(image_path, "rb") as image_file:
-                        encoded_string = base64.b64encode(image_file.read()).decode("utf-8")
+                        encoded_string = base64.b64encode(image_file.read()).decode(
+                            "utf-8"
+                        )
 
                     content = f"data:image/{file_extension};base64,{encoded_string}"
                 else:
@@ -158,7 +168,9 @@ def convert_to_openai_messages(
                     if "format" not in message:
                         raise Exception("Format of the image is not specified.")
                     else:
-                        raise Exception(f"Unrecognized image format: {message['format']}")
+                        raise Exception(
+                            f"Unrecognized image format: {message['format']}"
+                        )
 
                 # Calculate the size of the original binary data in bytes
                 content_size_bytes = len(content) * 3 / 4
@@ -206,27 +218,24 @@ def convert_to_openai_messages(
                 elif current_role == message["role"]:
                     current_content.append(message["content"])
                 else:
-                    combined_messages.append({
-                        "role": current_role,
-                        "content": "\n".join(current_content)
-                    })
+                    combined_messages.append(
+                        {"role": current_role, "content": "\n".join(current_content)}
+                    )
                     current_role = message["role"]
                     current_content = [message["content"]]
             else:
                 if current_content:
-                    combined_messages.append({
-                        "role": current_role,
-                        "content": "\n".join(current_content)
-                    })
+                    combined_messages.append(
+                        {"role": current_role, "content": "\n".join(current_content)}
+                    )
                     current_content = []
                 combined_messages.append(message)
 
         # Add the last message
         if current_content:
-            combined_messages.append({
-                "role": current_role,
-                "content": " ".join(current_content)
-            })
+            combined_messages.append(
+                {"role": current_role, "content": " ".join(current_content)}
+            )
 
         new_messages = combined_messages
 

interpreter/terminal_interface/profiles/defaults/local.py

@@ -3,20 +3,19 @@
 import subprocess
 import sys
 import time
+
 import inquirer
 import psutil
 import wget
+
 from interpreter import interpreter
 
-def get_ram():
-    total_ram = psutil.virtual_memory().total / (
-        1024 * 1024 * 1024
-    )  # Convert bytes to GB
-    return total_ram
 
 def download_model(models_dir, models, interpreter):
     # Get RAM and disk information
-    total_ram = get_ram()
+    total_ram = psutil.virtual_memory().total / (
+        1024 * 1024 * 1024
+    )  # Convert bytes to GB
     free_disk_space = psutil.disk_usage("/").free / (
         1024 * 1024 * 1024
     )  # Convert bytes to GB
@@ -53,6 +52,18 @@ def download_model(models_dir, models, interpreter):
 
     try:
         model_list = [
+            {
+                "name": "Llama-3-8B-Instruct",
+                "file_name": " Meta-Llama-3-8B-Instruct.Q5_K_M.llamafile",
+                "size": 5.76,
+                "url": "https://huggingface.co/jartine/Meta-Llama-3-8B-Instruct-llamafile/resolve/main/Meta-Llama-3-8B-Instruct.Q5_K_M.llamafile?download=true",
+            },
+            {
+                "name": "Phi-3-mini",
+                "file_name": "Phi-3-mini-4k-instruct.Q5_K_M.llamafile",
+                "size": 2.84,
+                "url": "https://huggingface.co/jartine/Phi-3-mini-4k-instruct-llamafile/resolve/main/Phi-3-mini-4k-instruct.Q5_K_M.llamafile?download=true",
+            },
             {
                 "name": "TinyLlama-1.1B",
                 "file_name": "TinyLlama-1.1B-Chat-v1.0.Q5_K_M.llamafile",
@@ -71,12 +82,6 @@ def download_model(models_dir, models, interpreter):
                 "size": 1.96,
                 "url": "https://huggingface.co/jartine/phi-2-llamafile/resolve/main/phi-2.Q5_K_M.llamafile?download=true",
             },
-            {
-                "name": "Phi-3-mini",
-                "file_name": "Phi-3-mini-4k-instruct.Q5_K_M.llamafile",
-                "size": 2.84,
-                "url": "https://huggingface.co/jartine/Phi-3-mini-4k-instruct-llamafile/resolve/main/Phi-3-mini-4k-instruct.Q5_K_M.llamafile?download=true",
-            },
             {
                 "name": "LLaVA 1.5",
                 "file_name": "llava-v1.5-7b-q4.llamafile",
@@ -89,12 +94,6 @@ def download_model(models_dir, models, interpreter):
                 "size": 5.15,
                 "url": "https://huggingface.co/jartine/Mistral-7B-Instruct-v0.2-llamafile/resolve/main/mistral-7b-instruct-v0.2.Q5_K_M.llamafile?download=true",
             },
-            {
-                "name": "Llama-3-8B-Instruct",
-                "file_name": " Meta-Llama-3-8B-Instruct.Q5_K_M.llamafile",
-                "size": 5.76,
-                "url": "https://huggingface.co/jartine/Meta-Llama-3-8B-Instruct-llamafile/resolve/main/Meta-Llama-3-8B-Instruct.Q5_K_M.llamafile?download=true",
-            },
             {
                 "name": "WizardCoder-Python-13B",
                 "file_name": "wizardcoder-python-13b.llamafile",
@@ -136,7 +135,7 @@ def download_model(models_dir, models, interpreter):
                 )
             ]
             answers = inquirer.prompt(questions)
-            
+
             if answers == None:
                 exit()
 
@@ -373,7 +372,7 @@ def list_ollama_models():
             )
         ]
         answers = inquirer.prompt(questions)
-        
+
         if answers == None:
             exit()
 
@@ -410,7 +409,9 @@ def list_ollama_models():
     interpreter.llm.api_base = "http://localhost:8080/v1"
     interpreter.llm.supports_functions = False
 
-user_ram = get_ram()
+user_ram = total_ram = psutil.virtual_memory().total / (
+    1024 * 1024 * 1024
+)  # Convert bytes to GB
 # Set context window and max tokens for all local models based on the users available RAM
 if user_ram and user_ram > 9:
     interpreter.llm.max_tokens = 1200
@@ -434,5 +435,29 @@ def list_ollama_models():
 Once you have accomplished the task, ask the user if they are happy with the result and wait for their response. It is very important to get feedback from the user. 
 The user will tell you the next task after you ask them.
 """
+
+interpreter.system_message = """You are an AI assistant that writes markdown code snippets to answer the user's request. You speak very concisely and quickly, you say nothing irrelevant to the user's request. For example:
+
+User: Open the chrome app.
+Assistant: On it. 
+```python
+import webbrowser
+webbrowser.open('https://chrome.google.com')
+```
+User: The code you ran produced no output. Was this expected, or are we finished?
+Assistant: No further action is required; the provided snippet opens Chrome.
+
+Now, your turn:
+"""
+
+interpreter.user_message_template = "{content} Please send me some code that would be able to answer my question, in the form of ```python\n... the code ...\n``` or ```shell\n... the code ...\n```"
+interpreter.code_output_template = "I executed that code. This was the ouput: {content}\n\nWhat does this output mean / what's next (if anything, or are we done)?"
+interpreter.empty_code_output_template = "The code above was executed on my machine. It produced no text output. what's next (if anything, or are we done?)"
+interpreter.code_output_sender = "user"
+interpreter.max_output = 500
+interpreter.llm.context_window = 8000
+interpreter.force_task_completion = False
+# interpreter.user_message_template = "{content}. If my question must be solved by running code on my computer, send me code to run enclosed in ```python or ```shell. Otherwise, don't send code and answer like a chatbot. Be concise, don't include anything unnecessary. Don't use placeholders, I can't edit code."
+
 # Set offline for all local models
 interpreter.offline = True