Embedding Workflows

Choose integrated embeddings or bring your own vectors to create, load, and search vector data with the Oracle VecDB Python SDK.

Both workflows use vector tables, records, metadata, and similarity search. The difference is where embeddings are generated and how the search query is supplied.

In simple terms, integrated embeddings mean that you send text and Oracle AI Database creates the vectors. BYO (bring your own) means that your application or another service creates the vectors and sends them to the database.

Integrated embedding and BYO vector workflows embed source content, store embeddings and metadata in one vector table, and query for similar vectors. Integrated embeddings use an in-database model for content and query embeddings. BYO uses an external provider for both.

Integrated Embeddings

Use integrated embeddings when Oracle AI Database should generate vectors from source content:

Create a vector table that uses an ONNX embedding model loaded in Oracle AI Database.
Upsert records that contain content in the metadata field configured for embedding. Oracle AI Database generates and stores the vector.
Search the vector table with text queries. Oracle AI Database generates the query vector using the table’s embedding model, and applies metadata filters in the same query.

Bring Your Own Vectors (BYO)

Use BYO when your application or an upstream pipeline generates embeddings outside Oracle AI Database:

Create a vector table with a dense vector type.
Generate vectors with an external embedding model.
Upsert records that include vector values and metadata.
Search the vector table with a supplied query vector, and apply metadata filters in the same query.

Reranking is an optional second stage after vector search. See Embedding and Reranking Models for model roles and rerank for the Python SDK workflow.

Related samples

Compare integrated embeddings and BYO in the Embedding & Ingest Patterns notebook.
Build a complete RAG workflow with the Gemini Vector RAG notebook.
See document ingestion and grounded answers in the Doc Chatbot sample application.

Next steps

Run the Quickstart: Build Vector Search with Python to follow the integrated-embedding workflow end to end.
create_vector_table and upsert_vectors.
query.
For the BYO workflow, follow upsert_vectors.

Parent topic: Concepts