Embedding Model
A model that converts data such as text and images into high-dimensional vector representations. Embedding models serve as a core component of RAG systems and semantic search pipelines, enabling semantic retrieval through vector similarity computation.
Core Facts
Timeline (last 90 days)
Embedding models and generative models are two different types of models: the former outputs fixed-dimension vectors, while the latter outputs text.
Embedding models only need to map inputs to fixed-dimensional vectors without autoregressive decoding; they have short sequence lengths and low memory requirements at inference time, making them one of the most seamlessly deployable model categories for in-browser AI.
In RAG, unstructured data is converted into high-dimensional numerical vectors by an Embedding Model, and the distance between vectors in mathematical space represents the semantic similarity of the content.
For short fragmented notes, insufficient vector information can cause false recall rates to rise significantly, and when the volume of records is insufficient, the association network becomes too sparse.
In RAG implementations, text is converted into high-dimensional vectors by embedding models and stored in vector databases such as Pinecone, Weaviate, and Milvus; at query time, the closest segments are found via cosine similarity.
嵌入向量通常为768维或1536维
更换或升级嵌入模型时，旧向量与新向量处于不同语义空间无法直接比较，往往需要全量重建索引
All Facts (11)
RAG（检索增强生成）技术架构将文档切片、向量化、存入向量数据库，查询时先检索相关片段再交给模型生成回答
90%Verified嵌入模型将文本转换为高维向量时，通常生成768维或1536维的浮点数数组
80%Verified向量数据库将文本内容通过嵌入模型（Embedding Model）转化为高维数值向量，检索时通过余弦相似度等算法找出语义最接近的内容
80%VerifiedIn RAG implementations, text is converted into high-dimensional vectors by embedding models and stored in vector databases such as Pinecone, Weaviate, and Milvus; at query time, the closest segments are found via cosine similarity.
75%VerifiedRAG通过将文档切分为语义块后经嵌入模型转化为高维向量存入向量数据库，查询时通过近似最近邻（ANN）搜索找到语义最相近的文档片段
65%UnverifiedEmbedding models and generative models are two different types of models: the former outputs fixed-dimension vectors, while the latter outputs text.
50%UnverifiedEmbedding models only need to map inputs to fixed-dimensional vectors without autoregressive decoding; they have short sequence lengths and low memory requirements at inference time, making them one of the most seamlessly deployable model categories for in-browser AI.
50%UnverifiedIn RAG, unstructured data is converted into high-dimensional numerical vectors by an Embedding Model, and the distance between vectors in mathematical space represents the semantic similarity of the content.
50%UnverifiedFor short fragmented notes, insufficient vector information can cause false recall rates to rise significantly, and when the volume of records is insufficient, the association network becomes too sparse.
50%Unverified更换或升级嵌入模型时，旧向量与新向量处于不同语义空间无法直接比较，往往需要全量重建索引
50%Unverified嵌入向量通常为768维或1536维
50%