APEX 24.2 RAG Search with apex_ai - Search your Database Records

In this blog,

1. We are going Vectorize our EMP table by creating embeddings.

2. We’re going to perform a Vector Search for ‘Who is the President?’

3. When the record comes back we’ll use Generative AI with the same question to give a natural language response.

I want to present this blog in as few steps are possible. If you want to search PDFs, click here

I like to think of embeddings like the Oracle Date Datatype. I imagine the Date Datatype as some mysterious value that I cannot ever see. I can only ever see a TO_CHAR’ed version of it.. ultimately Oracle probably store it as a non-text value. Same with embeddings; its just some internal format which when I TO_CHAR it, it looks like an array of numbers.

Prerequisites

• A 23ai Oracle Database - get one here

• An Generative AI Service - get one here

• A Vector Provider i.e ONNX (I dont have a guide yet) or OCI (Click Here)

• Install The EMP/DEPT APEX Sample data set

Database Configuration

1. Add a VECTOR column

    alter table EMP add
    ("EMBEDDINGS" VECTOR);
   

2. Update all Rows with the Embedding of a JSON Object representing the row.

    BEGIN
    FOR x in ( SELECT empno, JSON_OBJECT(*) AS emp_dept_json FROM EMP_DEPT_V )
    LOOP
      UPDATE EMP 
         SET EMBEDDINGS =  apex_ai.get_vector_embeddings(
            p_value             => x.emp_dept_json,
            p_service_static_id => 'OCI_VECTOR' )
        WHERE EMPNO = x.EMPNO;
    END LOOP;
    END;
   

3. Test it out by performing a Vector Search for the President

    SELECT JSON_OBJECT( v.* ) x
      FROM EMP e, EMP_DEPT_V v 
      WHERE e.empno = v.empno
      ORDER BY vector_distance(embeddings, apex_ai.get_vector_embeddings(
            p_value             => 'Who is the President?',
            p_service_static_id => 'OCI_VECTOR' ))
      FETCH FIRST 1 ROW ONLY;
   

4. Ask AI to make sense of all this: Paste this as a computation or something

5.  DECLARE
      -- Constants
      c_prompt CONSTANT VARCHAR2(4000) := 'Who is the President?';
      c_system_prompt CONSTANT VARCHAR2(4000) := 
        'I will pass you the question followed by a JSON representation of the record. 
         Your answer must be less than 255 Characters. 
         Your response must only relate to the JSON I send you.';
      c_service_static_id_vector CONSTANT VARCHAR2(100) := 'OCI_VECTOR';
      c_service_static_id_ai CONSTANT VARCHAR2(100) := 'oci';
   
      -- Variable
      l_response CLOB;
    BEGIN
      -- Query to fetch JSON object of the closest match
      SELECT JSON_OBJECT(v.*) x
        INTO l_response
        FROM EMP e, EMP_DEPT_V v 
        WHERE e.empno = v.empno
        ORDER BY vector_distance(
                  embeddings, 
                  apex_ai.get_vector_embeddings(
                    p_value             => c_prompt,
                    p_service_static_id => c_service_static_id_vector))
        FETCH FIRST 1 ROW ONLY;
   
      -- Generate AI response using APEX_AI package
      l_response := apex_ai.generate(
        p_prompt            => 'Who is the President and when was he hired ' || l_response,
        p_system_prompt     => c_system_prompt,
        p_service_static_id => c_service_static_id_ai);
   
      RETURN l_response;
    END;
   

Troubleshooting

If you see this… you cannot call apex_ai.generate outside an APEX Session.

ENJOY!

What's the picture? Its a view of The Inspire at Hornbeam Park. Take a look inside. Visit Yorkshire!