[Bug] Fix Regex for identifying OpenAI models (#778)

Addresses the issue with the model naming convention used by Azure's OpenAI API. Completion models are recognised strictly through the `-instruct` suffix on their names.

Additionally, the deployment path check has been changed to handle scenarios where ```pathlib.parts``` does not include the deployment path, ensuring more robust path validation during Azure deployments.

Closes #761

guidance/models/_azure_openai.py

@@ -1,5 +1,4 @@
 import pathlib
-import re
 
 from typing import Type
 from urllib.parse import parse_qs, urlparse
@@ -11,8 +10,7 @@
 from ._openai import (
     OpenAIChatEngine,
     OpenAICompletionEngine,
-    OpenAIInstructEngine,
-    chat_model_pattern,
+    OpenAIInstructEngine
 )
 
 try:
@@ -73,18 +71,15 @@ def __init__(
         if api_key is None and azure_ad_token_provider is None:
             raise ValueError("Please provide either api_key or azure_ad_token_provider")
 
-        parsed_url = urlparse(azure_endpoint)
-
         # if we are called directly (as opposed to through super()) then we convert ourselves to
         # a more specific subclass if possible
         if self.__class__ is AzureOpenAI:
             # Default to a completion model
-            found_subclass: Type[AzureOpenAI] = AzureOpenAICompletion
-            # Now see if we should be using a chat model
-            if parsed_url.path.endswith("/chat/completions"):
-                found_subclass = AzureOpenAIChat
-            elif re.match(chat_model_pattern, model):
-                found_subclass = AzureOpenAIChat
+            found_subclass: Type[AzureOpenAI] = (
+                AzureOpenAICompletion
+                if model.endswith("-instruct") 
+                else AzureOpenAIChat
+            )
 
             # convert to any found subclass
             self.__class__ = found_subclass
@@ -101,9 +96,14 @@ def __init__(
                 **kwargs,
             )
             return
+
+        parsed_url = urlparse(azure_endpoint)
 
         if azure_deployment is None:
-            azure_deployment = pathlib.Path(parsed_url.path).parts[3]
+            parts = pathlib.Path(parsed_url.path).parts
+            if len(parts) > 2:
+                azure_deployment = parts[3]
+
         parsed_query = parse_qs(parsed_url.query)
         api_version = (
             version
@@ -159,4 +159,4 @@ def get_role_end(self, name):
         else:
             raise ValueError(
                 f"The OpenAIInstruct model does not know about the {name} role type!"
-            )
+            )

guidance/models/_openai.py

@@ -1,5 +1,4 @@
 import os
-import re
 import typing
 
 import diskcache as dc
@@ -18,8 +17,6 @@
 except ImportError:
     client_class = None
 
-chat_model_pattern = r"^(ft:)?(gpt-3\.5-turbo|gpt-4)(?:(?!-instruct$)(-\w+)+)?(:[\w-]+(?:[:\w-]+)*)?(::\w+)?$"
-
 
 class OpenAIEngine(GrammarlessEngine):
     def __init__(
@@ -92,19 +89,16 @@ def __init__(
 
         # if we are called directly (as opposed to through super()) then we convert ourselves to a more specific subclass if possible
         if self.__class__ is OpenAI:
-            found_subclass = None
-
-            # chat
-            if re.match(chat_model_pattern, model):
-                found_subclass = OpenAIChat
 
             # instruct
             # elif "instruct" in model: # All current OpenAI instruct models behave as Completion models.
             #     found_subclass = OpenAIInstruct
-
-            # regular completion
-            else:
-                found_subclass = OpenAICompletion
+
+            found_subclass: typing.Type[OpenAI] = (
+                OpenAICompletion
+                if model.endswith("-instruct") 
+                else OpenAIChat
+            )
 
             # convert to any found subclass
             self.__class__ = found_subclass