From ad0b133ac878312497c9f514f9a2527fd0b85a45 Mon Sep 17 00:00:00 2001 From: hzg0601 Date: Wed, 6 Dec 2023 21:57:59 +0800 Subject: [PATCH] =?UTF-8?q?=E8=A7=A3=E5=86=B3faiss=E7=9B=B8=E4=BC=BC?= =?UTF-8?q?=E5=BA=A6=E9=98=88=E5=80=BC=E4=B8=8D=E5=9C=A80-1=E4=B9=8B?= =?UTF-8?q?=E9=97=B4=E7=9A=84=E9=97=AE=E9=A2=98?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- configs/kb_config.py.example | 6 ++++-- startup.py | 2 +- 2 files changed, 5 insertions(+), 3 deletions(-) diff --git a/configs/kb_config.py.example b/configs/kb_config.py.example index c9ff431..6be2848 100644 --- a/configs/kb_config.py.example +++ b/configs/kb_config.py.example @@ -21,8 +21,10 @@ OVERLAP_SIZE = 50 # 知识库匹配向量数量 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 -SCORE_THRESHOLD = 1 +# 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， +# 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 DEFAULT_SEARCH_ENGINE = "duckduckgo" diff --git a/startup.py b/startup.py index 3bb508a..11f60b6 100644 --- a/startup.py +++ b/startup.py @@ -128,7 +128,7 @@ def create_model_worker_app(log_level: str = "INFO", **kwargs) -> FastAPI: args.conv_template = None args.limit_worker_concurrency = 5 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.disable_log_requests = False