GustavoNogueira/A-Team-Security-Infra-Agent-Engineer
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: use max_completion_tokens for OpenAI models on OCI GenAI

Browse Source 
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.
This commit is contained in:
nogueiraguh
2026-03-03 20:02:22 -03:00
parent dcd2685671
commit 04e35dc681
1 changed files with 10 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
backend/app.py
View File

@@ -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
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: