Embeddings + answers Endpoint

Prompting
1

Hi,

I am experimenting on a possible use case of querying across large documents. I am currently using embeddings followed by answers endpoint setup. For 1.2 MB of data “text-search-curie-doc-001” model is taking 6 hrs. Is there any way to reduce the time or any alternate approach?

Thank you.

2

Hi @mounika.alavala! I think I can help. A few questions:

  • How many queries does that 1.2MB come out to?
  • Are you batching the queries or are you sending them one at a time?
  • I’m a little confused about how you’re using embeddings and answers together. Can you walk me through your steps?
3

I noticed it took about 20 seconds for 70 queries even on ADA but I was doing them individually. I suspect that batching them will be much faster.

4

Hi @hallacy
Following are the steps:

  1. I clubbed 1.2MB of text data and each line is sent to get embeddings using “text-search-curie-doc-001” model. This step is taking 6hrs time.
  2. After that this csv of embeddings is used by “text-search-curie-query-001”. From the output of “text-search-curie-query-001”, I take 100 search results and feed them to “answers endpoint”.

→ Only 1 query is asked at once.
→ To get embeddings it taking the maximum time. To get the query results it takes only 25 sec (Using “davinci” for answers endpoint).

Reference used : openai-python/Semantic_text_search_using_embeddings.ipynb at main · openai/openai-python · GitHub

5

How many transactions/queries is this in total?