Add custom data into 3.5-turbo


How to add custom data (ex. CSVLoader) into code with model 3.5-turbo, for example into ChatCompletion call?


1 Like

Great topic, subscribing. I thought about the same problem and the thoughts are:

  1. use the system role message to provide the data in the very first prompt in the text format explaining what is this data and how the model should use it;
  2. process the completion output respectively and define the query parameters needed depending on the user’s input, e.g. if the user mentioned some product name and characteristics we can query the database based on this input and based on the response compose another prompt by providing this data as a part of the input.

Yes, I did exactly as you wrote, it’s more or less ok, but in this case I can’t:

  • check if no product in the database
  • if there are not enough about of it
  • if some product has no parameters
  • can’t provide choice of materials for example.
    So, I can say - as a POC it works, but as production ready solution - not yet.

Then there is option 3: Finetuning.

Kindly share your experience here if You’d try that in your project.

P.S. However, you can only fine-tune InstructGPT (Ada, Davinci etc) models and GPT-3 models, not GPT-3.5 models, as stated in the official OpenAI documentation

Also, as option 4, it is possible to create embeddings vectors from custom data organized as multiple text documents. For searching the information for a user, the GPT API call retrieves the most relevant document using just these embeddings. Then with another prompt, GPT extracts the matching answer from the most relevant document.

Finetuning is not a option for such task. It’s good for text, but not for concrete numbers. Numbers can not be probability, they must be equal. You can have answer “it seems we have iPhone on the stock”. The same issuew with converting concrete numbwes/dates/facts into embeddings and search by relevant.

For GPT-3 we can add customer data source, it works with bugs, but ok.

For 3.5 - I can’t find and asked.

For GPT-3 we can add customer data source, it works with bugs, but ok.

Do you mean the data sources such as CSV, not text embeddings? Could you share more?

Yes, exactly. Langchain allows to add it easy.