Skip to content

Embedding Architecture

Victor supports pluggable embedding providers for semantic tool selection and codebase search. The design keeps providers swappable while sharing a common configuration model.

Components

  • Embedding configuration (provider, model, dimensions)
  • Embedding service (generates vectors)
  • Cache (stores tool embeddings)
  • Vector store (optional, for codebase search)

Provider Types

  • Local: sentence-transformers, Ollama, vLLM, LM Studio
  • Remote: cloud embedding APIs

Data Flow

  1. Text input is embedded by the provider.
  2. Embeddings are cached for reuse.
  3. Similarity search selects tools or documents.

Notes

  • Local providers enable offline use when models are installed.
  • Latency and quality depend on model choice and hardware.
  • You can override providers per profile.