ReferenceDraft · not yet verified
Embeddings and the model gateway
The embedding model the engine searches by meaning with, and the model gateway that serves it.
| Setting | Engine default | Shipped | What it does |
|---|---|---|---|
EMBEDDING_BASE_URL | — (required) | http://gateway:4000/v1 | The OpenAI-compatible endpoint the engine embeds through |
EMBEDDING_MODEL | — (required) | nomic-embed-text | The route the gateway serves. The model name is the embedding space: a new name opens a new space, and the embedding sweep re-embeds every record into it |
EMBEDDING_API_KEY | empty | empty | A bearer key for the endpoint, if it needs one |
OPENROUTER_API_KEY | — | — | In gateway.env, read by the gateway only. Needed only to switch to the shipped OpenRouter route (EMBEDDING_MODEL=text-embedding-3-small) |
EMBED_MAX_CHUNKS | 64 | — | The most chunks a record is embedded in; text past them has no vector |
WINDOW_EMBED_SIZE | 5 | — | How many records a window vector covers |
The gateway's routes are in conf/gateway.yaml: each names a model, where it is served, and its vector width (output_vector_size), which the engine requires.