Better LLM <> computer exchanges, added zsh, less annoying messages, significantly better "dumb" model performance

Author: KillianLucas

README.md

@@ -214,7 +214,7 @@ Once the server is running, you can begin your conversation with Open Interprete
 
 (When you run the command `interpreter --local`, the steps above will be displayed.)
 
-> **Note:** Local mode sets your `context_window` to 3000, and your `max_tokens` to 600. If your model has different requirements, set these parameters manually (see below).
+> **Note:** Local mode sets your `context_window` to 3000, and your `max_tokens` to 1000. If your model has different requirements, set these parameters manually (see below).
 
 #### Context Window, Max Tokens
 

interpreter/README.md

@@ -37,7 +37,7 @@
     {
         "name": "disable_procedures",
         "nickname": "dp",
-        "help_text": "disables procedures (RAG of some common OI use-cases). disable to shrink system message",
+        "help_text": "disables procedures (RAG of some common OI use-cases). disable to shrink system message. auto-disabled for non-OpenAI models",
         "type": bool,
     },
     {
@@ -187,7 +187,7 @@ def cli(interpreter):
             "openai/" + interpreter.model
         )  # This tells LiteLLM it's an OpenAI compatible server
         interpreter.api_base = "http://localhost:1234/v1"
-        interpreter.max_tokens = 600
+        interpreter.max_tokens = 1000
         interpreter.context_window = 3000
         interpreter.api_key = "0"
 

interpreter/cli/cli.py

@@ -10,6 +10,7 @@
     "python": Python,
     "bash": Shell,
     "shell": Shell,
+    "zsh": Shell,
     "javascript": JavaScript,
     "html": HTML,
     "applescript": AppleScript,

interpreter/code_interpreters/language_map.py

@@ -11,10 +11,14 @@ def convert_to_coding_llm(text_llm, debug_mode=False):
     def coding_llm(messages):
         # First, tell it how to run code.
 
-        # Old way:
-        # system_message += "\nTo execute code on the user's machine, write a markdown code block *with the language*, i.e:\n\n```python\nprint('Hi!')\n```\n\nYou will receive the output ('Hi!'). Use any language."
-
-        # New way:
+        # System message method:
+        assert messages[0]["role"] == "system"
+        messages[0][
+            "message"
+        ] += "\nTo 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."
+
+        # Gaslight method (DISABLED):
+        '''
         gaslight = None
         if messages[-1]["role"] == "user":
             # Last message came from the user.
@@ -42,6 +46,7 @@ def coding_llm(messages):
 
         if gaslight:
             messages.append({"role": "assistant", "message": gaslight})
+        '''
 
         # If it tried to use Jupyter, let it know.
         if "code" in messages[-1]:

interpreter/llm/convert_to_coding_llm.py

@@ -20,6 +20,9 @@ def setup_llm(interpreter):
         # Function-calling LLM
         coding_llm = setup_openai_coding_llm(interpreter)
     else:
+        # Disable procedures, which confuse most of these models
+        interpreter.disable_procedures = True
+
         # Non-function-calling LLM
         text_llm = setup_text_llm(interpreter)
         coding_llm = convert_to_coding_llm(text_llm, debug_mode=interpreter.debug_mode)

interpreter/llm/setup_llm.py

@@ -67,12 +67,13 @@ def coding_llm(messages):
                     max_tokens=interpreter.context_window,
                 )
             else:
-                display_markdown_message(
+                if len(messages) == 1:
+                    display_markdown_message(
+                        """
+                    **We were unable to determine the context window of this model.** Defaulting to 3000.
+                    If your model can handle more, run `interpreter --context_window {token limit}` or `interpreter.context_window = {token limit}`.
                     """
-                **We were unable to determine the context window of this model.** Defaulting to 3000.
-                If your model can handle more, run `interpreter --context_window {token limit}` or `interpreter.context_window = {token limit}`.
-                """
-                )
+                    )
                 messages = tt.trim(
                     messages=messages, system_message=system_message, max_tokens=3000
                 )

interpreter/llm/setup_openai_coding_llm.py

@@ -37,13 +37,14 @@ def base_llm(messages):
                     messages, system_message=system_message, model=interpreter.model
                 )
             except:
-                display_markdown_message(
+                if len(messages) == 1:
+                    display_markdown_message(
+                        """
+                    **We were unable to determine the context window of this model.** Defaulting to 3000.
+                    If your model can handle more, run `interpreter --context_window {token limit}` or `interpreter.context_window = {token limit}`.
+                    Also, please set max_tokens: `interpreter --max_tokens {max tokens per response}` or `interpreter.max_tokens = {max tokens per response}`
                     """
-                **We were unable to determine the context window of this model.** Defaulting to 3000.
-                If your model can handle more, run `interpreter --context_window {token limit}` or `interpreter.context_window = {token limit}`.
-                Also, please set max_tokens: `interpreter --max_tokens {max tokens per response}` or `interpreter.max_tokens = {max tokens per response}`
-                """
-                )
+                    )
                 messages = tt.trim(
                     messages, system_message=system_message, max_tokens=3000
                 )

interpreter/llm/setup_text_llm.py

@@ -43,18 +43,26 @@ def convert_to_openai_messages(messages, function_calling=True):
                 )
             else:
                 if message["output"] == "No output":
+                    content = "The code above was executed on my machine. It produced no output. Was that expected?"
+                else:
                     content = (
                         "Code output: "
                         + message["output"]
-                        + "\n\nThe output is obfuscated to me. What does it mean / how should we proceed if there's more to do?"
+                        + "\n\nWhat does this output mean / what's next (if anything)?"
                     )
-                else:
-                    content = "The code above was executed on my machine. It produced no output. Was that expected?"
+
                 new_messages.append(
                     {
                         "role": "user",
                         "content": content,
                     }
                 )
 
+    if not function_calling:
+        new_messages = [
+            msg
+            for msg in new_messages
+            if "content" in msg and msg["content"].strip() != ""
+        ]
+
     return new_messages