Use a different model to guess pip and respond

If we finetune a model it will learn to always respond in python
surrounded with backticks so it won't be able to produce pip packages
well. We need to have a different model.

src/rawdog/config.py

@@ -10,6 +10,7 @@
     "llm_api_key": None,
     "llm_base_url": None,
     "llm_model": "gpt-4-turbo-preview",
+    "pip_model": None
     "llm_custom_provider": None,
     "llm_temperature": 1.0,
     "retries": 2,
@@ -21,6 +22,7 @@
 setting_descriptions = {
     "retries": "If the script fails, retry this many times before giving up.",
     "leash": "Print the script before executing and prompt for confirmation.",
+    "pip_model": "The model to use to get package name from import name.",
 }
 
 

src/rawdog/llm_client.py

@@ -46,7 +46,14 @@ def add_message(self, role: str, content: str):
 
     def get_python_package(self, import_name: str):
         base_url = self.config.get("llm_base_url")
-        model = self.config.get("llm_model")
+        model = self.config.get("pip_model")
+        llm_model = self.config.get("llm_model")
+        if model is None:
+            if "ft:" in llm_model or "rawdog" in llm_model or "abante" in llm_model:
+                model = "gpt-3.5-turbo"
+            else:
+                model = llm_model
+
         custom_llm_provider = self.config.get("llm_custom_provider")
 
         messages = [