weiweiw
/
Langchain-Chatchat
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

解决faiss相似度阈值不在0-1之间的问题

This commit is contained in:
hzg0601 2023-12-06 21:57:59 +08:00
parent 4c2fda7200
commit ad0b133ac8
2 changed files with 5 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
configs/kb_config.py.example
View File

@ -21,8 +21,10 @@ OVERLAP_SIZE = 50
# 知识库匹配向量数量 # 知识库匹配向量数量
VECTOR_SEARCH_TOP_K = 3 VECTOR_SEARCH_TOP_K = 3
# 知识库匹配相关度阈值，取值范围在0-1之间，SCORE越小，相关度越高，取到1相当于不筛选，建议设置在0.5左右 # 知识库匹配的距离阈值，取值范围在0-1之间，SCORE越小，距离越小从而相关度越高，
SCORE_THRESHOLD = 1 # 取到1相当于不筛选，实测bge-large的距离得分大部分在0.01-0.7之间，
# 相似文本的得分最高在0.55左右，因此建议针对bge设置得分为0.6
SCORE_THRESHOLD = 0.6
# 默认搜索引擎。可选：bing, duckduckgo, metaphor # 默认搜索引擎。可选：bing, duckduckgo, metaphor
DEFAULT_SEARCH_ENGINE = "duckduckgo" DEFAULT_SEARCH_ENGINE = "duckduckgo"

2
startup.py
View File

@ -128,7 +128,7 @@ def create_model_worker_app(log_level: str = "INFO", **kwargs) -> FastAPI:
args.conv_template = None args.conv_template = None
args.limit_worker_concurrency = 5 args.limit_worker_concurrency = 5
args.no_register = False args.no_register = False
args.num_gpus = 4 # vllm worker的切分是tensor并行，这里填写显卡的数量 args.num_gpus = 1 # vllm worker的切分是tensor并行，这里填写显卡的数量
args.engine_use_ray = False args.engine_use_ray = False
args.disable_log_requests = False args.disable_log_requests = False