A plain vanilla OpenAI natural language AI application (ChatPGT) can answer general questions about topics such as the ones you can find on Wikipedia, because the underlying LLM model was trained using many sources, including Wikipedia. But in order to answer questions about information in proprietary documents, such as a company handbook, you need to feed the proprietary documents to the system. One simple but crude way to do this is to directly include the text from a document into the query/context/question. But if you are working with large documents, or if you have many documents, you probably need to use RAG (retrieval-augmented generation).

Until recently, creating a RAG system was extremely difficult and time-consuming. For a set of PDF documents, you had to use OCR to extract the text, chunk the text into blocks, convert the blocks into vector embeddings, store the embeddings form of the documents into an external vector database (such as Chroma), write tricky code to query the vector database, and write code to integrate the vector query results into the question/context. This process is a huge challenge.

But a few months ago, the OpenAI Responses API simplified creating RAG systems dramatically. The Responses API does most of the work for you. But to use the Responses API for RAG, you must create a special vector database store on the OpenAI servers, and upload your specialized documents into it.

I decided to investigate. My first step, in this blog post, was to create a vector store. And I wanted to make sure I could also delete the vector store so I wouldn’t be charged storage fees for all eternity.

The ideas are best explained by the examining the output of my demo and the program code. I create a vector store and then immediately delete it:

Begin OpenAI vector store demo

Creating vector store
Done
{'id': 'vs_68b9a8e212b48191b8375930e6c3ed41',
 'name': 'dummy_vector_store', 'created_at': 1756997858,
 'file_count': 0}

Deleting the vector store
Done
VectorStoreDeleted(id='vs_68b9a8e212b48191b8375930e6c3ed41',
 deleted=True,
 object='vector_store.deleted')

End demo

The demo program code is essentially some OpenAI documentation from platform.openai.com/docs/api-reference/vector-stores that I refactored a bit.

# create_openai_vector_store.py

from openai import OpenAI

def create_vector_store(store_name):
  # assumes a global-scope client object
  try:
    v_store = client.vector_stores.create(name=store_name)
    store_info = {
      "id": v_store.id,
      "name": v_store.name,
      "created_at": v_store.created_at,
      "file_count": v_store.file_counts.completed
    }
    return store_info
  except Exception as e:
    print("FATAL in create_vector_store() " + str(e))

def delete_vector_store(store_id):
  try:
    info = client.vector_stores.delete(vector_store_id=store_id)
    return info
  except Exception as e:
    print("FATAL in delete_vector_store() " + str(e))  
   
# -----------------------------------------------------------

try:
  print("\nBegin OpenAI vector store demo ")

  key = "sk-proj-_AX7bGTXUwg-qojh2T5Z2CVXrox" + \
    "xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx" + \
    "xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx" + \
    "xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"

  client = OpenAI(api_key=key)

  print("\nCreating vector store ")
  store_info = create_vector_store("dummy_vector_store")
  print("Done ")
  print(store_info)

  store_id = store_info["id"]    # need to delete store

  print("\nDeleting the vector store ")
  delete_info = delete_vector_store(store_id)
  print("Done ")
  print(delete_info)

  print("\nEnd demo ")

except Exception as e:
  print("Fatal error in program" + str(e))

OK, that was an interesting investigation. Next, I need to explore how to use a vector store together with the Responses API to implement a RAG system.

In most AI systems, you send a request over the Internet and a response is sent back to you. In the late 1930s, there were fascinating telephone based systems to request and get music. These were the days before juke boxes.

Right: This is a “Choice by Voice” machine, used by a different network system.