klbr/khoj
1
0
Fork 0
mirror of https://github.com/khoaliber/khoj.git synced 2026-03-09 13:25:11 +00:00
Code Issues Packages Projects Releases Wiki Activity

Support deepseek v3.1 via official deepseek api

Browse Source 
The new deepseek-chat is powered by deepseek v3.1, which is a hybrid
reasoning model unlike it's predecessor, deepseek v3.
This commit is contained in:
Debanjum
2025-08-22 22:13:15 -07:00
parent 40488b3b68
commit 464c1546b7
1 changed files with 7 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
src/khoj/processor/conversation/openai/utils.py
View File

@@ -126,7 +126,7 @@ def completion_with_backoff(
if model_name.startswith("grok-4"): if model_name.startswith("grok-4"):
# Grok-4 models do not support reasoning_effort parameter # Grok-4 models do not support reasoning_effort parameter
model_kwargs.pop("reasoning_effort", None) model_kwargs.pop("reasoning_effort", None)
elif model_name.startswith("deepseek-reasoner"): elif model_name.startswith("deepseek-reasoner") or model_name.startswith("deepseek-chat"):
stream_processor = in_stream_thought_processor stream_processor = in_stream_thought_processor
# Two successive messages cannot be from the same role. Should merge any back-to-back messages from the same role. # Two successive messages cannot be from the same role. Should merge any back-to-back messages from the same role.
# The first message should always be a user message (except system message). # The first message should always be a user message (except system message).
@@ -325,8 +325,12 @@ async def chat_completion_with_backoff(
# Grok-4 models do not support reasoning_effort parameter # Grok-4 models do not support reasoning_effort parameter
if not model_name.startswith("grok-4"): if not model_name.startswith("grok-4"):
model_kwargs["reasoning_effort"] = reasoning_effort model_kwargs["reasoning_effort"] = reasoning_effort
elif model_name.startswith("deepseek-reasoner") or "deepseek-r1" in model_name: elif (
# Official Deepseek reasoner model and some inference APIs like vLLM return structured thinking output. model_name.startswith("deepseek-chat")
or model_name.startswith("deepseek-reasoner")
or "deepseek-r1" in model_name
):
# Official Deepseek models and some inference APIs like vLLM return structured thinking output.
# Others like DeepInfra return it in response stream. # Others like DeepInfra return it in response stream.
# Using the instream thought processor handles both cases, structured thoughts and in response thoughts. # Using the instream thought processor handles both cases, structured thoughts and in response thoughts.
stream_processor = ain_stream_thought_processor stream_processor = ain_stream_thought_processor