gmakstutis/LightRAG
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Add presence penalty config option for smaller models

Browse source 
- Add OPENAI_LLM_PRESENCE_PENALTY setting
- Recommend 1.5 for Qwen3 <32B params
- Update max completion tokens comment
This commit is contained in:
yangdx 2025-08-21 11:35:23 +08:00
parent 0dd245e847
commit 5d34007f2c
1 changed files with 3 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

4
env.example
View file

@ -149,7 +149,9 @@ LLM_BINDING_API_KEY=your_api_key
### OpenAI Specific Parameters ### OpenAI Specific Parameters
# OPENAI_LLM_TEMPERATURE=1.0 # OPENAI_LLM_TEMPERATURE=1.0
# OPENAI_LLM_REASONING_EFFORT=low # OPENAI_LLM_REASONING_EFFORT=low
### Set the maximum number of completion tokens if your LLM generates repetitive or unconstrained output ### For models like Qwen3 with fewer than 32B param, it is recommended to set the presence penalty to 1.5
# OPENAI_LLM_PRESENCE_PENALTY=1.5
### If the presence penalty still can not stop the model from generates repetitive or unconstrained output
# OPENAI_LLM_MAX_COMPLETION_TOKENS=16384 # OPENAI_LLM_MAX_COMPLETION_TOKENS=16384
### use the following command to see all support options for openai and azure_openai ### use the following command to see all support options for openai and azure_openai
### lightrag-server --llm-binding openai --help ### lightrag-server --llm-binding openai --help