Embeddings & Semantic Search¶

Embeddings power semantic tool selection and codebase search. Use local or remote providers.

Quick Setup¶

# ~/.victor/profiles.yaml
profiles:
  local:
    embedding_provider: sentence-transformers
    embedding_model: BAAI/bge-small-en-v1.5

Or with Ollama:

profiles:
  local:
    embedding_provider: ollama
    embedding_model: qwen3-embedding:8b

Tool Selection Strategies¶

Strategy	Description
`keyword`	Fast, no embeddings needed
`semantic`	Embedding-based matching
`hybrid`	Blends keyword + semantic (default)
`auto`	Chooses based on availability

tool_selection_strategy: hybrid

Models¶

Model	Dimensions	Size	Use Case
BAAI/bge-small-en-v1.5	384	~130MB	Default
all-MiniLM-L12-v2	384	~120MB	Low memory
qwen3-embedding:8b	4096	~4.7GB	High quality

Air-Gapped Mode¶

For restricted environments without network access:

profiles:
  airgapped:
    provider: ollama
    model: qwen2.5-coder:7b
    airgapped_mode: true
    embedding_provider: sentence-transformers
    embedding_model: BAAI/bge-small-en-v1.5

Behavior:
- Web tools disabled
- Only local providers allowed
- Falls back to keyword selection if embeddings unavailable

Architecture¶

User Request → Embedding Provider → Vector Comparison → Tool Selection
                     ↓
              Embedding Cache

Providers:
- Local: sentence-transformers, Ollama, vLLM, LM Studio
- Remote: Cloud embedding APIs

Troubleshooting¶

Issue	Solution
Tools not triggering	Use `keyword` strategy
Slow cold starts	Pre-cache embeddings
Memory issues	Use smaller model (bge-small)