Managed embeddings

Keep searchable text indexed as records change.

Define an embedding index for the label and string property you want to retrieve by meaning. Use managed embeddings for the shortest path or external vectors for model control.

Start free View docs

Embedding pipelines become another system to operate.

Generating vectors, backfilling old records, tracking index state, and keeping updates synchronized adds infrastructure before semantic retrieval provides any product value.

Before

Run a separate embedding worker
Store vectors in another database
Schedule backfills after index changes
Keep source records and vectors synchronized

With RushDB

Create one label and property index policy
Use server-side embeddings for managed indexes
Backfill existing records through the index lifecycle
Bring your own vectors when your application requires it

What it enables

Semantic memory without a separate vector pipeline.

Embedding indexes are policies scoped to a label and string property. That keeps vectorization explicit while letting your app keep writing ordinary records.

Index only useful text

Choose the specific property that should be recalled by meaning, such as an agent output, document body, or product description.

Backfill stays attached

Managed indexes can generate vectors for existing records and keep updated values searchable as records change.

Model control remains possible

External indexes let teams supply their own vectors when compliance, local inference, or domain models matter.

How it works

Start with the smallest useful path.

Choose the searchable property

Create an index policy scoped to a record label and one string property, such as MEMORY.output or PRODUCT.description.

Choose managed or external vectors

Managed indexes generate embeddings server-side. External indexes accept vectors supplied by your application.

Wait for ready state

Check index lifecycle state before search. Once ready, matching records can be retrieved by semantic similarity.

Flow

How text becomes searchable memory

The index policy tells RushDB which records and property should carry vectors. Once the index is ready, search can rank matching records by meaning.

Pick label + property

Create index policy

Backfill vectors

Write new records

Search by meaning

Implementation sketch

Create an index, then search with plain text.

Managed search accepts a natural-language query. External indexes use a pre-computed queryVector instead.

Know the operational boundary.

One-time setup is explicit

Searchable properties are not embedded by accident. Create an index policy first, then monitor its status before querying.

Read the embedding index guide

External vectors remain available

Use BYOV when a domain model, local provider, or compliance requirement means vector generation belongs in your application.

Read the BYOV guide