fix: use max_completion_tokens for OpenAI models on OCI GenAI

OpenAI models reject 'max_tokens' and require 'max_completion_tokens'.
Also skip 'top_k' for OpenAI as it is unsupported. Detection is based
on the provider field in the model catalog.

backend/app.py

@@ -868,15 +868,22 @@ def _call_genai(gc: dict, message: str, history: list = None) -> str:
                 chat_history.append(entry)
             chat_request.chat_history = chat_history
     else:
-        # ── Generic format (Meta Llama, Google, xAI) - exact chat_demo.py pattern ──
+        # ── Generic format (Meta Llama, Google, xAI, OpenAI) ──
         chat_request = oci.generative_ai_inference.models.GenericChatRequest()
         chat_request.api_format = oci.generative_ai_inference.models.BaseChatRequest.API_FORMAT_GENERIC
-        chat_request.max_tokens = int(gc.get("max_tokens", 6000))
+        provider = model_info.get("provider", "")
+        is_openai = provider == "openai"
+        # OpenAI models use max_completion_tokens instead of max_tokens
+        # and do not support top_k
+        if is_openai:
+            chat_request.max_completion_tokens = int(gc.get("max_tokens", 6000))
+        else:
+            chat_request.max_tokens = int(gc.get("max_tokens", 6000))
+            chat_request.top_k = int(gc.get("top_k", 1))
         chat_request.temperature = float(gc.get("temperature", 1))
         chat_request.frequency_penalty = float(gc.get("frequency_penalty", 0))
         chat_request.presence_penalty = float(gc.get("presence_penalty", 0))
         chat_request.top_p = float(gc.get("top_p", 0.95))
-        chat_request.top_k = int(gc.get("top_k", 1))
 
         messages = []
         if history: