BGE
A series of open-source text embedding models released by the Beijing Academy of Artificial Intelligence (BAAI). BGE models demonstrate strong performance on multiple semantic retrieval benchmarks, support both Chinese and English, and are among the most widely used open-source embedding models.
Core Facts
OpenAI的text-embedding-ada-002和开源的BGE系列是常用的Embedding模型
BGE (BAAI General Embedding) is a Chinese embedding model open-sourced by the Beijing Academy of Artificial Intelligence.
OpenAI's text-embedding-ada-002 can map text into a high-dimensional vector space, where semantically similar texts are closer together in the vector space.
Timeline (last 90 days)
Embedding models are neural network models that map discrete inputs into dense vectors, with representative examples including OpenAI's text-embedding-ada-002, Voyage AI's voyage series, and open-source BGE, E5, GTE and others.
Some later BGE versions such as bge-m3 have simplified the query prefix, but when using classic versions such as bge-large-zh, the requirement to add a prefix on the query side must still be followed.
The classic BGE versions (such as bge-large-zh) have an input limit of 512 tokens.
BGE uses an asymmetric retrieval training approach in which queries and documents use different encoding templates; omitting the query prefix causes the encoding distribution to be inconsistent with the training distribution.
When using the BGE embedding model for retrieval, a specific instruction prefix must be added on the query side, otherwise similarity scoring may be distorted, producing an inversion where irrelevant sentences score higher than relevant ones.
Commonly available Chinese-capable Embedding models include OpenAI's text-embedding-3, BGE, M3E, Tongyi Qianwen text-embedding, and others.
RAG成本优化杠杆包括：嵌入模型换用本地BGE可省90%以上，向量数据库自托管可省80%，简单查询路由到小模型可省50-80%，高频查询加Redis缓存可省30-60%
Open-source embedding models such as BGE and Nomic-Embed can run on ordinary GPUs or even CPUs without requiring an internet connection.
Even in modern embedding models optimized through instruction fine-tuning or contrastive learning (such as BGE and E5), the anisotropy problem has not been completely eliminated, only reduced in degree.
Mainstream embedding models such as OpenAI's text-embedding-3-small and lightweight versions of the BGE series typically have parameter counts in the range of 100M–500M.
11 more timeline events
All Facts (20)
OpenAI的text-embedding-ada-002和开源的BGE系列是常用的Embedding模型
80%VerifiedBGE (BAAI General Embedding) is a Chinese embedding model open-sourced by the Beijing Academy of Artificial Intelligence.
80%VerifiedOpenAI's text-embedding-ada-002 can map text into a high-dimensional vector space, where semantically similar texts are closer together in the vector space.
65%UnverifiedBGE和text-embedding-3是当前可选的向量嵌入（Embedding）模型
90%Unverified知识库检索中，文档分段后每个文本块通过嵌入模型（如text-embedding-ada-002或BGE系列）转换为高维向量存储在向量数据库中
90%UnverifiedEmbedding models are neural network models that map discrete inputs into dense vectors, with representative examples including OpenAI's text-embedding-ada-002, Voyage AI's voyage series, and open-source BGE, E5, GTE and others.
50%UnverifiedSome later BGE versions such as bge-m3 have simplified the query prefix, but when using classic versions such as bge-large-zh, the requirement to add a prefix on the query side must still be followed.
50%UnverifiedWhen using the BGE embedding model for retrieval, a specific instruction prefix must be added on the query side, otherwise similarity scoring may be distorted, producing an inversion where irrelevant sentences score higher than relevant ones.
50%UnverifiedBGE uses an asymmetric retrieval training approach in which queries and documents use different encoding templates; omitting the query prefix causes the encoding distribution to be inconsistent with the training distribution.
50%UnverifiedThe classic BGE versions (such as bge-large-zh) have an input limit of 512 tokens.
50%UnverifiedCommonly available Chinese-capable Embedding models include OpenAI's text-embedding-3, BGE, M3E, Tongyi Qianwen text-embedding, and others.
50%UnverifiedRAG成本优化杠杆包括：嵌入模型换用本地BGE可省90%以上，向量数据库自托管可省80%，简单查询路由到小模型可省50-80%，高频查询加Redis缓存可省30-60%
50%UnverifiedOpen-source embedding models such as BGE and Nomic-Embed can run on ordinary GPUs or even CPUs without requiring an internet connection.
50%UnverifiedEven in modern embedding models optimized through instruction fine-tuning or contrastive learning (such as BGE and E5), the anisotropy problem has not been completely eliminated, only reduced in degree.
50%UnverifiedMainstream embedding models such as OpenAI's text-embedding-3-small and lightweight versions of the BGE series typically have parameter counts in the range of 100M–500M.
50%UnverifiedStarting from version 2.8, Blender gradually removed the built-in Game Engine (BGE); the community primarily uses UPBGE or exports scenes to external engines such as Godot/Unity to implement game logic.
50%UnverifiedThe BGE series models on HuggingFace offer better support for Chinese and can be deployed locally to avoid data-export compliance risks.
50%UnverifiedOpenAI text-embedding, BGE, and Jina are different Embedding models to choose from.
50%UnverifiedEmbedding models (such as BGE and text-embedding-3-small) are much smaller than LLMs and can typically run on CPUs or low-spec GPUs, with inference latency in the millisecond range.
50%UnverifiedRepresentative models for dense vector retrieval include DPR, BGE, and text-embedding-ada.
50%