Vector databases
Flexible Vector Database is a managed retrieval store backed by OpenSearch. You choose the embedding model (text-only or multimodal), push documents in, and search by text, image, or a combination. Recommendations are a first-class feature: record a view, then ask for similar items weighted by what the user interacted with.
You need an embeddings instance
Vector databases are decoupled from inference. Create an embeddings instance first and reference its serving name when you create the database. The instance bills per active hour regardless of database activity.
Database fields
| Field | Notes |
|---|---|
| id | Internal database id. Used in API paths. |
| slug | URL-safe handle. Stable across renames. Use it in client code. |
| name | Display name. Free text, can be changed at any time. |
| description | Free-text description shown in the list and detail pages. |
| embedding_serving_name | Serving name of the embeddings instance that encodes the documents. |
| embedding_model_ref | Hugging Face repo id resolved from the serving name. Read-only. |
| opensearch_index_name | Backing index. Read-only, managed by the platform. |
| status | ready | provisioning | failed. Failed databases show an error message. |
Where to go next
Create a database →
Pick a model, set the slug, and prepare the first document set.
Indexing →
Push documents in via index-direct (sync) or index-jobs (async).
Searching →
Text, image, and multimodal queries with optional recommendation bias.
Recommendations →
Record views, get a recommend token, and bias search results.