Embeddings Configuration

Embeddings

Synapses uses vector embeddings for semantic search (search(mode="semantic")), memory recall, and hybrid context carving.

{
  "embeddings": "builtin",
  "embedding_model": "nomic-embed-text",
  "embed_pool_size": 3
}

Field	Type	Default	Description
`embeddings`	`string`	`"builtin"`	Mode: `"builtin"`, `"ollama"`, or `"off"`
`embedding_model`	`string`	`"nomic-embed-text"`	Ollama model name (only used with `"ollama"` mode)
`embedding_endpoint`	`string`	`""`	OpenAI-compatible HTTP endpoint (alternative to Ollama)
`embed_pool_size`	`int`	`3`	ONNX inference workers (1-8, only for `"builtin"` mode)

Uses the bundled nomic-embed-text ONNX model (~137MB). No external dependencies. Runs inference in-process with configurable worker pool.

{ "embeddings": "builtin", "embed_pool_size": 3 }

Uses Ollama for embedding generation. Requires Ollama running locally with the embedding model pulled.

{
  "embeddings": "ollama",
  "embedding_model": "nomic-embed-text"
}

Disables vector embeddings entirely. Semantic search falls back to FTS5 keyword search.

{ "embeddings": "off" }

Use this for air-gapped environments or when you want minimal resource usage.

search(mode="semantic") — HyDE-enhanced vector search
memory(action="search", query=...) — Semantic memory retrieval
Context carving with HybridLambda > 0 — Blends structural and semantic similarity
Node embedding index (HNSW) for fast approximate nearest neighbor lookup