Use Retrieval-Augmented Generation and Vectors
Provides examples of how to set up Select AI with RAG and use in-database embedding models to generate vector embeddings for document chunks and user prompts.
Topics:
- Example: Set Up and Use Select AI with RAG
This example guides you through setting up credentials, configuring network access, and creating a vector index for integrating OCI Generative AI vector store cloud services with OpenAI using Oracle Autonomous AI Database. - Example: Select AI with In-database Transformer Models
This example demonstrates how you can import a pretrained transformer model that is stored in Oracle object storage into your Oracle AI Database 26ai instance and then use the imported in-database model in Select AI profile to generate vector embeddings for document chunks and user prompts.
Parent topic: Examples Using Select AI
Example: Set Up and Use Select AI with RAG
This example guides you through setting up credentials, configuring network access, and creating a vector index for integrating OCI Generative AI vector store cloud services with OpenAI using Oracle Autonomous AI Database.
Review:
Note:
-
You can specify the content location for creating a vector index by using either an object storage URI or a database directory file. This approach provides flexibility in sourcing files from object storage or from files managed through Oracle AI database directories.
-
Multibyte characters are not supported in file names for input documents. Files with multibyte characters are in the file name are skipped during vectorization.
The setup concludes with creating an AI profile that uses the vector
index to enhance LLM responses. Finally, this example uses the Select AI
narrate action, which returns a response that has been enhanced
using information from the specified vector database.
The following example demonstrates building and querying vector index in Oracle AI Database 26ai.
--Grants EXECUTE privilege to ADB_USER
GRANT EXECUTE on DBMS_CLOUD_AI to ADB_USER;
--Grants EXECUTE privilege DBMS_CLOUD_PIPELINE to ADB_USER
GRANT EXECUTE on DBMS_CLOUD_PIPELINE to ADB_USER;
-- Create the OpenAI credential
BEGIN
DBMS_CLOUD.CREATE_CREDENTIAL(
credential_name => 'OPENAI_CRED',
username => 'OPENAI_CRED',
password => '<your_api_key>'
);
END;
/
PL/SQL procedure successfully completed.
-- Append the OpenAI endpoint
BEGIN
DBMS_NETWORK_ACL_ADMIN.APPEND_HOST_ACE(
host => 'api.openai.com',
ace => xs$ace_type(privilege_list => xs$name_list('http'),
principal_name => 'ADB_USER',
principal_type => xs_acl.ptype_db)
);
END;
/
PL/SQL procedure successfully completed.
-- Create the object store credential
BEGIN
DBMS_CLOUD.CREATE_CREDENTIAL(
credential_name => 'OCI_CRED',
username => '<your_username>',
password => '<OCI_profile_password>'
);
END;
/
PL/SQL procedure successfully completed.
-- Create the profile with the vector index.
BEGIN
DBMS_CLOUD_AI.CREATE_PROFILE(
profile_name =>'OPENAI_ORACLE',
attributes =>'{"provider": "openai",
"credential_name": "OPENAI_CRED",
"vector_index_name": "MY_INDEX",
"temperature": 0.2,
"max_tokens": 4096
}');
END;
/
PL/SQL procedure successfully completed.
-- Set profile
EXEC DBMS_CLOUD_AI.SET_PROFILE('OPENAI_ORACLE');
PL/SQL procedure successfully completed.
-- create a vector index with the vector store name, object store location and
-- object store credential
BEGIN
DBMS_CLOUD_AI.CREATE_VECTOR_INDEX(
index_name => 'MY_INDEX',
wait_for_completion => false,
attributes => '{"vector_db_provider": "oracle",
"location": "https://swiftobjectstorage.us-phoenix-1.oraclecloud.com/v1/my_namespace/my_bucket/my_data_folder",
"object_storage_credential_name": "OCI_CRED",
"profile_name": "OPENAI_ORACLE",
"vector_dimension": 1536,
"vector_distance_metric": "cosine",
"chunk_overlap":128,
"chunk_size":1024
}');
END;
/
PL/SQL procedure successfully completed.
-- After the vector index is populated, we can now query the index.
-- Set profile
EXEC DBMS_CLOUD_AI.SET_PROFILE('OPENAI_ORACLE');
PL/SQL procedure successfully completed.
-- Select AI answers the question with the knowledge available in the vector database.
set pages 1000
set linesize 150
SELECT AI narrate how can I deploy an oracle machine learning model;
RESPONSE
To deploy an Oracle Machine Learning model, you would first build your model within the Oracle database. Once your in-database models are built, they become immediately available for use, for instance, through a SQL query using the prediction operators built into the SQL language.
The model scoring, like model building, occurs directly in the database, eliminating the need for a separate engine or environment within which the model and corresponding algorithm code operate. You can also use models from a different schema (user account) if the appropriate permissions are in place.
Sources:
- Manage-your-models-with-Oracle-Machine-Learning-on-Autonomous-Database.txt (https://objectstorage.../v1/my_namespace/my_bucket/my_data_folder/Manage-your-models-with-Oracle-Machine-Learning-on-Autonomous-Database.txt)
- Develop-and-deploy-machine-learning-models-using-Oracle-Autonomous-Database-Machine-Learning-and-APEX.txt (https://objectstorage.../v1/my_namespace/my_bucket/my_data_folder/Develop-and-deploy-machine-learning-models-using-Oracle-Autonomous-Database-Machine-Learning-and-APEX.txt)
The WAIT_FOR_COMPLETION parameter is included in the
DBMS_CLOUD_AI.CREATE_VECTOR_INDEX call. The default value is
FALSE, meaning the procedure returns immediately after
scheduling vectorization.
If you run NARRATE prompts before the vector index is
fully created, Select AI may produce incomplete or suboptimal responses because not
all documents are available for retrieval. To avoid this, you can set
WAIT_FOR_COMPLETION to TRUE. When set to
TRUE, the procedure does not return until vectorization is
complete. This ensures all documents are indexed before you begin issuing RAG-based
prompts. For large document collections, this may take a significant amount of time.
See DBMS_CLOUD_AI .CREATE_VECTOR_INDEX
Procedure for more details.
You can specify content location for vector index creation using either Object Storage URIs or database directory files.
Object Storage URISpecify a location_uri that points to
the object storage location containing the documents you want to index. For
example,
https://objectstorage.us-ashburn-1.oraclecloud.com/n/namespace/b/bucket/o/documents/Directory Files
You can also use directory
objects to reference local files. Use this format MY_DIR:file.txt.
The directory name (MY_DIR) is case-insensitive
by default. The file name is case-sensitive.
Wildcards are supported only for filenames:
-
Use
*for multiple characters:MY_DIR:* -
Use
?for a single character:MY_DIR:file?.txt -
To specify a case-sensitive directory name, wrap it in double quotes:
"MyDir":* -
To include a quote character in the filename, use two quotes:
MY_DIR:''file.txt
The following example creates a vector index named
RAG_INDEX_DIR using all .txt files located in
the directory object
MY_DIR.
-- Create the profile with the vector index.
BEGIN
DBMS_CLOUD_AI.CREATE_PROFILE(
profile_name =>'OPENAI_ORACLE',
attributes =>'{"provider": "openai",
"credential_name": "OPENAI_CRED",
"vector_index_name": "RAG_INDEX_DIR",
"temperature": 0.2,
"max_tokens": 4096
}');
END;
/
PL/SQL procedure successfully completed.
-- Set profile
EXEC DBMS_CLOUD_AI.SET_PROFILE('OPENAI_ORACLE');
PL/SQL procedure successfully completed.
-- create a vector index with the vector store name, object store location and
-- object store credential
BEGIN
DBMS_CLOUD_AI.CREATE_VECTOR_INDEX(
index_name => 'RAG_INDEX_DIR',
attributes => '{"vector_db_provider": "oracle",
"location": "MY_DIR:*.txt",
"object_storage_credential_name": "OCI_CRED",
"profile_name": "OPENAI_ORACLE",
"vector_dimension": 1536,
"vector_distance_metric": "cosine",
"chunk_overlap":128,
"chunk_size":1024
}');
END;
/
PL/SQL procedure successfully completed. Select AI RAG includes an enable_sources attribute that
controls whether source documents retrieved during vector search appear in the final
response. This attribute is enabled by default. When set to true,
Select AI validates and includes the retrieved source documents in the output. When
set to false, Select AI still validates the sources but does not
include the Sources section in the response. See DBMS_CLOUD_AI Vector Index Attributes for
more information.
The following example shows
enable_sources set to true. Select AI
validates source documents and includes them in the response
output.
-- enable_sources = true (default)
set pages 1000
set linesize 150
SELECT AI narrate how can I deploy an oracle machine learning model;
RESPONSE
To deploy an Oracle Machine Learning model, you would first build your model within the Oracle database. Once your in-database models are built, they become immediately available for use, for instance, through a SQL query using the prediction operators built into the SQL language.
The model scoring, like model building, occurs directly in the database, eliminating the need for a separate engine or environment within which the model and corresponding algorithm code operate. You can also use models from a different schema (user account) if the appropriate permissions are in place.
Sources:
- Manage-your-models-with-Oracle-Machine-Learning-on-Autonomous-Database.txt (https://objectstorage.../Manage-your-models-with-Oracle-Machine-Learning-on-Autonomous-Database.txt)
- Develop-and-deploy-machine-learning-models-using-Oracle-Autonomous-Database-Machine-Learning-and-APEX.txt (https://objectstorage.../Develop-and-deploy-machine-learning-models-using-Oracle-Autonomous-Database-Machine-Learning-and-APEX.txt)
The following example shows enable_sources set to
false. Select AI still validates the source documents but does
not print the Sources section in the
output.
-- enable_sources = false (sources validated but not displayed)
set pages 1000
set linesize 150
SELECT AI narrate how can I deploy an oracle machine learning model;
RESPONSE
To deploy an Oracle Machine Learning model, you would first build your model within the Oracle database. Once your in-database models are built, they become immediately available for use, for instance, through a SQL query using the prediction operators built into the SQL language.
The model scoring, like model building, occurs directly in the database, eliminating the need for a separate engine or environment within which the model and corresponding algorithm code operate. You can also use models from a different schema (user account) if the appropriate permissions are in place.
Parent topic: Use Retrieval-Augmented Generation and Vectors
Example: Select AI with In-database Transformer Models
This example demonstrates how you can import a pretrained transformer model that is stored in Oracle object storage into your Oracle AI Database 26ai instance and then use the imported in-database model in Select AI profile to generate vector embeddings for document chunks and user prompts.
-
your pretrained model imported in your Oracle AI Database 26ai instance.
-
optionally, access to Oracle object storage.
Review the steps in Import Pretrained Models in ONNX Format for Vector Generation Within the Database and the blog Pre-built Embedding Generation model for Oracle AI Database 26ai to import a pretrained transformer model into your database.
The following example shows how to import a pretained transformer model from Oracle object storage into your database and then view the imported model.
- Create a Directory object, or use an existing directory object
CREATE OR REPLACE DIRECTORY ONNX_DIR AS 'onnx_model';
-- Object storage bucket
VAR location_uri VARCHAR2(4000);
EXEC :location_uri := 'https://adwc4pm.objectstorage.us-ashburn-1.oci.customer-oci.com/p/eLddQappgBJ7jNi6Guz9m9LOtYe2u8LWY19GfgU8flFK4N9YgP4kTlrE9Px3pE12/n/adwc4pm/b/OML-Resources/o/';
-- Model file name
VAR file_name VARCHAR2(512);
EXEC :file_name := 'all_MiniLM_L12_v2.onnx';
-- Download ONNX model from object storage into the directory object
BEGIN
DBMS_CLOUD.GET_OBJECT(
credential_name => NULL,
directory_name => 'ONNX_DIR',
object_uri => :location_uri || :file_name);
END;
/
-- Load the ONNX model into the database
BEGIN
DBMS_VECTOR.LOAD_ONNX_MODEL(
directory => 'ONNX_DIR',
file_name => :file_name,
model_name => 'MY_ONNX_MODEL');
END;
/
-- Verify
SELECT model_name, algorithm, mining_function
FROM user_mining_models
WHERE model_name='MY_ONNX_MODEL';These examples illustrate how to use in-database transformer models within a Select AI profile. One profile is configured only for generating vector embeddings, while the other supports both Select AI actions and vector index creation.
Review Perform Prerequisites for Select AI to complete the prerequisites.
The following is an example for generating vector embeddings only:
BEGIN
DBMS_CLOUD_AI.CREATE_PROFILE(
profile_name => 'EMBEDDING_PROFILE',
attributes => '{"provider" : "database",
"embedding_model": "MY_ONNX_MODEL"}'
);
END;
/The following is an example for general Select AI actions and vector index generation where you can specify a supported AI provider. This example uses OCI Gen AI profile and credentials. See Select your AI Provider and LLMs for list of supported providers. However, if you want to use in-database transformer model for generating vector embeddings, then use "database: <MY_ONNX_MODEL>" in embedding_model attribute:
BEGIN
DBMS_CLOUD.CREATE_CREDENTIAL(
credential_name => 'GENAI_CRED',
user_ocid => 'ocid1.user.oc1..aaaa...',
tenancy_ocid => 'ocid1.tenancy.oc1..aaaa...',
private_key => '<your_api_key>',
fingerprint => '<your_fingerprint>'
);
END;
/
BEGIN
DBMS_CLOUD_AI.CREATE_PROFILE(
profile_name => 'OCI_GENAI',
attributes => '{"provider": "oci",
"model": "meta.llama-3.3-70b-instruct",
"credential_name": "GENAI_CRED",
"vector_index_name": "MY_INDEX",
"embedding_model": "database: MY_ONNX_MODEL"}'
);
END;
/This example demonstrates
how to use Select AI with an in-database transformer model if another schema owner
owns the model. Specify schema_name.object_name as the fully
qualified name of the model in embedding_model attribute. If the
current user is the schema owner or owns the model, you can omit the schema
name.
CREATE ANY MINING MODELsystem privilegeSELECT ANY MINING MODELsystem privilegeSELECT MINING MODELobject privilege on the specific model
To grant a system privilege, you must either have been
granted the system privilege with the ADMIN OPTION or have been
granted the GRANT ANY PRIVILEGE system privilege.
See System Privileges for Oracle Machine Learning for SQL to review the privileges.
The
following statements allow ADB_USER1 to score data and view model
details in any schema as long as SELECT access has been granted to
the data. However, ADB_USER1 can only create models in the
ADB_USER1 schema.
GRANT CREATE MINING MODEL TO ADB_USER1;
GRANT SELECT ANY MINING MODEL TO ADB_USER1;
BEGIN
DBMS_CLOUD_AI.CREATE_PROFILE(
profile_name => 'OCI_GENAI',
attributes => '{"provider": "oci",
"credential_name": "GENAI_CRED",
"vector_index_name": "MY_INDEX",
"embedding_model": "database: ADB_USER1.MY_ONNX_MODEL"}'
);
END;
/The following example shows how you can specify case sensitive model object name:
BEGIN
DBMS_CLOUD_AI.CREATE_PROFILE(
profile_name => 'OCI_GENAI',
attributes => '{"provider": "oci",
"credential_name": "GENAI_CRED",
"model": "meta.llama-3.3-70b-instruct",
"vector_index_name": "MY_INDEX",
"embedding_model": "database: \"adb_user1\".\"my_model\""}'
);
END;
/These examples demonstrate end-to-end steps for using in-database transformer model with Select AI RAG. One profile uses database as the provider exclusively created for generating embedding vectors while the other profile uses oci as the provider created for Select AI actions as well as vector index.
Review Perform Prerequisites for Select AI to provide the necessary privileges.
--Grant create any directory privilege to the user
GRANT CREATE ANY DIRECTORY to ADB_USER;
- Create a Directory object, or use an existing directory object
CREATE OR REPLACE DIRECTORY ONNX_DIR AS 'onnx_model';
-- Object storage bucket
VAR location_uri VARCHAR2(4000);
EXEC :location_uri := 'https://adwc4pm.objectstorage.us-ashburn-1.oci.customer-oci.com/p/eLddQappgBJ7jNi6Guz9m9LOtYe2u8LWY19GfgU8flFK4N9YgP4kTlrE9Px3pE12/n/adwc4pm/b/OML-Resources/o/';
-- Model file name
VAR file_name VARCHAR2(512);
EXEC :file_name := 'all_MiniLM_L12_v2.onnx';
-- Download ONNX model from object storage into the directory object
BEGIN
DBMS_CLOUD.GET_OBJECT(
credential_name => NULL,
directory_name => 'ONNX_DIR',
object_uri => :location_uri || :file_name);
END;
/
-- Load the ONNX model into the database
BEGIN
DBMS_VECTOR.LOAD_ONNX_MODEL(
directory => 'ONNX_DIR',
file_name => :file_name,
model_name => 'MY_ONNX_MODEL');
END;
/
-- Verify
SELECT model_name, algorithm, mining_function
FROM user_mining_models
WHERE model_name='MY_ONNX_MODEL';
--Administrator grants EXECUTE privilege to ADB_USER
GRANT EXECUTE on DBMS_CLOUD_AI to ADB_USER;
--Administrator grants EXECUTE privilege DBMS_CLOUD_PIPELINE to ADB_USER
GRANT EXECUTE on DBMS_CLOUD_PIPELINE to ADB_USER;
-- Create the object store credential
BEGIN
DBMS_CLOUD.CREATE_CREDENTIAL(
credential_name => 'OCI_CRED',
username => '<your_username>',
password => '<OCI_profile_password>'
);
END;
/
PL/SQL procedure successfully completed.
-- Create the profile with Oracle Database.
BEGIN
DBMS_CLOUD_AI.CREATE_PROFILE(
profile_name =>'EMBEDDING_PROFILE',
attributes =>'{"provider": "database",
"embedding_model": "MY_ONNX_MODEL"
}');
END;
/
PL/SQL procedure successfully completed.
-- Set profile
EXEC DBMS_CLOUD_AI.SET_PROFILE('EMBEDDING_PROFILE');
PL/SQL procedure successfully completed.
This example uses oci as the
provider.
--Grant create any directory privilege to the user
GRANT CREATE ANY DIRECTORY to ADB_USER;
- Create a Directory object, or use an existing directory object
CREATE OR REPLACE DIRECTORY ONNX_DIR AS 'onnx_model';
-- Object storage bucket
VAR location_uri VARCHAR2(4000);
EXEC :location_uri := 'https://adwc4pm.objectstorage.us-ashburn-1.oci.customer-oci.com/p/eLddQappgBJ7jNi6Guz9m9LOtYe2u8LWY19GfgU8flFK4N9YgP4kTlrE9Px3pE12/n/adwc4pm/b/OML-Resources/o/';
-- Model file name
VAR file_name VARCHAR2(512);
EXEC :file_name := 'all_MiniLM_L12_v2.onnx';
-- Download ONNX model from object storage into the directory object
BEGIN
DBMS_CLOUD.GET_OBJECT(
credential_name => NULL,
directory_name => 'ONNX_DIR',
object_uri => :location_uri || :file_name);
END;
/
-- Load the ONNX model into the database
BEGIN
DBMS_VECTOR.LOAD_ONNX_MODEL(
directory => 'ONNX_DIR',
file_name => :file_name,
model_name => 'MY_ONNX_MODEL');
END;
/
-- Verify
SELECT model_name, algorithm, mining_function
FROM user_mining_models
WHERE model_name='MY_ONNX_MODEL';
–-Administrator Grants EXECUTE privilege to ADB_USER
GRANT EXECUTE on DBMS_CLOUD_AI to ADB_USER;
--Administrator Grants EXECUTE privilege DBMS_CLOUD_PIPELINE to ADB_USER
GRANT EXECUTE on DBMS_CLOUD_PIPELINE to ADB_USER;
-- Create the object store credential
BEGIN
DBMS_CLOUD.CREATE_CREDENTIAL(
credential_name => 'OCI_CRED',
username => '<your_username>',
password => '<OCI_profile_password>'
);
END;
/
--Create GenAI credentials
BEGIN
DBMS_CLOUD.CREATE_CREDENTIAL(
credential_name => 'GENAI_CRED',
user_ocid => 'ocid1.user.oc1..aaaa...',
tenancy_ocid => 'ocid1.tenancy.oc1..aaaa...',
private_key => '<your_api_key>',
fingerprint => '<your_fingerprint>'
);
END;
/
--Create OCI AI profile
BEGIN
DBMS_CLOUD_AI.CREATE_PROFILE(
profile_name => 'OCI_GENAI',
attributes => '{"provider": "oci",
"model": "meta.llama-3.3-70b-instruct",
"credential_name": "GENAI_CRED",
"vector_index_name": "MY_INDEX",
"embedding_model": "database: MY_ONNX_MODEL"}'
);
END;
/
-- Set profile
EXEC DBMS_CLOUD_AI.SET_PROFILE('OCI_GENAI');
PL/SQL procedure successfully completed.
-- create a vector index with the vector store name, object store location and
-- object store credential
BEGIN
DBMS_CLOUD_AI.CREATE_VECTOR_INDEX(
index_name => 'MY_INDEX',
attributes => '{"vector_db_provider": "oracle",
"location": "https://swiftobjectstorage.us-phoenix-1.oraclecloud.com/v1/my_namespace/my_bucket/my_data_folder",
"object_storage_credential_name": "OCI_CRED",
"profile_name": "OCI_GENAI",
"vector_dimension": 384,
"vector_distance_metric": "cosine",
"chunk_overlap":128,
"chunk_size":1024
}');
END;
/
PL/SQL procedure successfully completed.
-- Set profile
EXEC DBMS_CLOUD_AI.SET_PROFILE('OCI_GENAI');
PL/SQL procedure successfully completed.
-- Select AI answers the question with the knowledge available in the vector database.
set pages 1000
set linesize 150
SELECT AI narrate how can I deploy an oracle machine learning model;
RESPONSE
To deploy an Oracle Machine Learning model, you would first build your model within the Oracle database. Once your in-database models are
built, they become immediately available for use, for instance, through a SQL query using the prediction operators built into the SQL
language.
The model scoring, like model building, occurs directly in the database, eliminating the need for a separate engine or environment within
which the model and corresponding algorithm code operate. You can also use models from a different schema (user account) if the appropriate
permissions are in place.
Sources:
- Manage-your-models-with-Oracle-Machine-Learning-on-Autonomous-Database.txt (https://objectstorage.../v1/my_namespace/my_bucket/
my_data_folder/Manage-your-models-with-Oracle-Machine-Learning-on-Autonomous-Database.txt)
- Develop-and-deploy-machine-learning-models-using-Oracle-Autonomous-Database-Machine-Learning-and-APEX.txt
(https://objectstorage.../v1/my_namespace/my_bucket/my_data_folder/Develop-and-deploy-machine-learning-models-using-Oracle-Autonomous-
Database-Machine-Learning-and-APEX.txt)Parent topic: Use Retrieval-Augmented Generation and Vectors