Embeddings

General definition

Embeddings are produced by passing content through an embedding model, which outputs a fixed-length list of floating-point numbers (the vector). The key property is that items with similar meaning produce vectors that are close together by cosine similarity or Euclidean distance, while unrelated items produce vectors that are far apart. A sentence about “cardiac care” and one about “heart treatment” will land near each other even though they share no words.

• Text embeddings - produced by models such as OpenAI text-embedding-3, Cohere Embed or open-source alternatives (BGE, E5, Nomic)
• Multimodal embeddings - represent text and images in a shared space (e.g. CLIP), enabling cross-modal search
• Sparse vs dense - dense embeddings (a full float vector) capture semantics; sparse embeddings (a mostly-zero vector) capture keywords; hybrid search combines both
• Dimensionality - typical embedding sizes range from 384 to 3,072 dimensions depending on the model

Embeddings are the input to vector databases and the foundation of semantic search. In a RAG pipeline, documents are chunked and embedded at index time; at query time the user’s question is embedded and the nearest document chunks are retrieved and passed to the LLM.

In the Ethora ecosystem

The Ethora AI SDK’s RAG bot framework embeds your knowledge base content at ingestion time and queries those embeddings at runtime to retrieve the most relevant context before generating a response. You can bring your own embedding model, including open-source ones that run locally, so no content needs to leave your infrastructure during either indexing or retrieval.

In self-hosted deployments this matters especially for sensitive domains: clinical notes, legal documents or financial records are embedded and searched entirely within your own environment. The self-hosted LLM agent configuration supports OpenAI-compatible embedding endpoints, meaning any compliant local model (Ollama, vLLM) can serve as the embedder.