Local Embeddings
Local Embeddings
Section titled “Local Embeddings”Turn text into vectors on this device with an in-browser embeddings model. Connect it to Local Knowledge to index and search documents without an API key or any data leaving your browser.
What it does
Section titled “What it does”Exposes an on-device embeddings model as an Embeddings dependency. Inference runs in the shared background worker, so the editor stays responsive. The default is all-MiniLM-L6-v2; you can also pick the legacy Universal Sentence Encoder or any installed embeddings model.
The embeddings carry their model’s identity, so a knowledge store knows which model it was built with and never mixes vectors from different models.
When to use it
Section titled “When to use it”- Build a private, offline knowledge base for the Q&A Agent or RAG Agent.
- Avoid per-token embedding costs when indexing many documents.
Inputs and settings
Section titled “Inputs and settings”| Setting | Notes |
|---|---|
| Model | An on-device embeddings model (default all-MiniLM-L6-v2, recommended). Universal Sentence Encoder keeps older knowledge stores readable. |
Outputs
Section titled “Outputs”Returns an embeddings dependency for Local Knowledge and other nodes that accept embeddings.
Choosing a model for a knowledge store
Section titled “Choosing a model for a knowledge store”A knowledge store is built with one embeddings model and remembers it. Pick the model before you first fill a new store:
- A new store uses whatever embeddings are connected the first time it runs (or all-MiniLM-L6-v2 if none).
- Connecting a different model to an existing store later doesn’t re-embed it — the store keeps using its own model, with a warning in the run log.
See Local Knowledge for details.
Troubleshooting
Section titled “Troubleshooting”- “Embeddings model … isn’t available. Install it from Local AI.” — install it from the Local AI page, or download it from the picker.
- “Model mismatch” warning — the store was built with a different model; its own model was used. To switch models, create a new store.
- First run is slow — the model loads (and downloads if needed) on first use, then stays warm for the session.
Related nodes
Section titled “Related nodes”- Local Knowledge
- Q&A Agent, RAG Agent
- Local AI — install and manage on-device models.