Skip to content

Semantic Tool Selection

Victor can select tools by comparing embeddings of the user request to tool descriptions. This reduces tool noise and keeps prompts smaller when you have many tools.

When To Use

  • Large tool catalog
  • Ambiguous or multi-step requests
  • You want fewer irrelevant tools per turn

Requirements

  • An embedding provider (local or remote)
  • Optional: cached tool embeddings for faster selection

Strategy Options

  • keyword: available without embeddings, fast, deterministic
  • semantic: embedding-based matching
  • hybrid: blends keyword and semantic
  • auto: chooses based on availability and settings

Configuration

Profiles or environment variables:

# Modern strategy setting (preferred)
tool_selection_strategy: auto  # keyword, semantic, hybrid, auto

# Legacy toggle (if your config still uses it)
use_semantic_tool_selection: true

# Embedding settings
embedding_provider: sentence-transformers
embedding_model: BAAI/bge-small-en-v1.5

Notes

  • Offline use is possible with a local embedding model installed.
  • Cold starts are slower until embeddings are cached.
  • Quality depends on model choice and tool descriptions.