Oracle Database Guidelines for Generative AI Agents

Generative AI Agents doesn't manage the database, so you must set up your existing database so Generative AI Agents can connect to it. The following requirements

Database Table

Create an Oracle Database 23ai table with the following fields:

Required Fields

• DOCID: An id assigned to each record or document
• BODY: The actual content that you want the agent to search
• VECTOR: The vector generated from an embedding model based on the body content

Optional Fields

• CHUNKID: An id for the chunked document, in case the document is larger than 512 tokens.
• URL: A URL reference for the content, if available.
• TITLE: The title of the body content

Tip

You can create text embeddings in OCI Generative AI.

Database Function

A function is a subprogram that can take parameters and return a value. You need to set up a database function can return vector search results from each query. Here are the requirements:

Input Parameters

• p_query: The query string
• top_k: Number of top results

Return Type

• SYS_REFCURSOR

Requirements

• The embedding model that you use for the function's query field must be the same as the embedding model that transforms the table's BODY content to vector embeddings.

• The function's return fields must match the table's required (DOCID, BODY and SCORE) and optional fields (CHUNKID, TITLE, and URL).

  The BODY content is processed by an embedding model to produce a VECTOR for each content. The function calculates the distances between the query vector and each VECTOR, returning the top_k rows. The SCORE field represents the distances between the query vector and the VECTOR from the BODY.

• If the function's return field names don't match the table field names, you must use aliases in the function.

Example

Here is an example of a database function:

create or replace FUNCTION RETRIEVAL_FUNC (p_query IN VARCHAR2,top_k IN NUMBER) RETURN SYS_REFCURSOR IS
    v_results SYS_REFCURSOR;
    query_vec VECTOR;
BEGIN
    query_vec := dbms_vector.utl_to_embedding(
        p_query,
        json('{
            "provider": "OCIGenAI",
            "credential_name": "OCI_VECTOR_CREDENTIAL",
            "url": "https://inference.generativeai.us-chicago-1.oci.oraclecloud.com/20231130/actions/embedText",
            "model": "cohere.embed-english-v3.0"
        }')
    );

    OPEN v_results FOR
        select l.id as DOCID, lv.chunk_id as CHUNKID, l.file_name as TITLE, 
        'https://objectstorage.us-chicago-1.oraclecloud.com/n/{namespace}/b/{bucket_name}/o/' || l.file_name as URL,
        lv.chunk_txt as BODY, vector_distance(embed_vector, query_vec) AS SCORE
        from legislation_vector lv, legislation l
        where l.id = lv.id
        order by SCORE 
        fetch first top_k rows only;

    RETURN v_results;
END;

Here are high-level networking guidelines:

Networking Requirements

• Add the database to the private subnet of a virtual cloud network (VCN) in OCI.
• Set up the following properties for the for the subnet's ingress rule:
  • Source Type: CIDR
  • Source CIDR: The VCN's IPv4 CIDR Block (such as 10.0.0.0/16)
  • IP Protocol: TCP
  • Source Port Range: All
  • Destination Port Range: 1521-1522

Security Requirements

• Create a Vault to save the secrets for the database.
• Create the vault in the same compartment as the database.

You can create reusable connections to an Oracle database in OCI, by using the Database Tools service. Follow these guidelines to create a connection for your database.