Embedding Models

Embedding models convert text into vector representations, enabling semantic search and knowledge base operations. Similar concepts produce similar vectors, which is what powers retrieval-augmented generation (RAG).

How Embeddings Work

  1. Documents are chunked and converted to embeddings.
  2. Embeddings are stored in a vector database.
  3. When a task queries the knowledge base, Xagent searches for similar embeddings.
  4. Retrieved content is provided as context to the LLM.

When to Configure

An embedding model is required for knowledge base functionality — uploading and searching documents, building RAG systems, and semantic retrieval.

Supported Providers

ProviderModelsBest for
OpenAI & compatibletext-embedding-3-small / -large, text-embedding-ada-002General-purpose embeddings.
DashScope (Alibaba Cloud)text-embedding-v4 / v3 / v2Chinese / Asian-language optimization.
Xinferencebge-large-en-v1.5, bge-base-en, all-MiniLM-L6-v2, other HuggingFace modelsPrivacy and self-hosting.

HuggingFace models

To use HuggingFace embedding models, deploy them via Xinference and point Xagent at the Xinference Base URL.

Parameters

ParameterDescription
DimensionsVector size. Lower (384–768) is faster and cheaper; higher (3072+) improves accuracy at more storage cost.
Chunk sizeTokens per chunk when indexing (typically 512–1000). Smaller chunks are more precise; larger chunks carry more context.

Best Practices

  • General use: OpenAI text-embedding-3-small — good cost/performance.
  • Highest quality: text-embedding-3-large or bge-large-en-v1.5.
  • Match chunk size to typical query length and use overlap to preserve context.

Next Steps