Processing configuration
Choose the enrichment and embedding models MarkDB uses for your tenant.
Settings -> Processing in the dashboard controls which models MarkDB's enrichment and embedding pipelines use for your tenant. Leave everything on the defaults to run MarkDB's managed tier.
Enrichment
The model that writes your summaries.
| Setting | Default | Notes |
|---|---|---|
| Provider | openai, anthropic, google, or openai-compatible. | |
| Model | gemini-3.1-flash-lite | Any chat model from the catalog. |
| Max output tokens | 1200 | 0–16000. |
| Temperature | 0.10 | 0–2. |
Embeddings
The model that vectorizes summaries for hybrid search.
| Setting | Default | Notes |
|---|---|---|
| Provider | openai, google, or openai-compatible (not Anthropic). | |
| Model | gemini-embedding-001 | Any embedding model from the catalog. |
| Dimensions | 3072 | Only offered for models with more than one Matryoshka size. |
How settings are applied
- Provider credential required. Whatever provider you pick, add its key under Settings -> LLM keys or the pipeline can't call it. See Authentication.
- Model names aren't validated on save. An unknown model is accepted at save time and only fails later, at dispatch, with a "model not found" error. Pick from the catalog.
- Changes take effect within about a minute. The resolved config is cached per tenant (~60s TTL), so a save isn't instantaneous.
- Changing the embedder rebuilds the index. Switching the embedding model or dimensions changes the vector signature, so MarkDB re-indexes your existing pages against the new embedder. The dashboard warns you before you save.
- Reset removes your overrides and returns the tenant to the managed defaults.
Embedder consistency
Query-time and index-time embeddings must share a model and dimension to be comparable. MarkDB keeps them aligned by re-indexing when you change the embedder -- see Hybrid search.