Assistant api don't access the database

Prompting
, , , ,
1

[not a DEV] SOS struggling with gpt4 turbo prompts (:brazil:)

Hi, guys! I’m using the Assistant API (apparently) and on the front end made by a DEV (which I am not), I see we’re using the GPT4 turbo.

So far, no problem…
I have access to the following basic functions:
Choose the model, temperature, description, instruction, and whether or not it can respond out of context. Also choose which type of AI to use (Assistant or Chat, but Assistant allows me to associate files for responses, and chat does not).

In addition, I can attach files to the database, so that they are consulted by the template before answering user questions.

My problem starts when it seems that AI is stubborn with me, it knows the answers, but it insists on saying that it doesn’t know, and when I say “yes, you know,” it gives me the answer that is in the database document.

It doesn’t seem to obey the prompt! Does it not obey because I give the instructions in Portuguese? It is my native language, and also the language its speaks.

I have even tried to upload a XPTO.txt file to the database with not even 50 characters, and I put in the prompt instructions: when u don’t know the answer to a question send the text of the document. Well, it didn’t work, I was frustrated.

:bulb: Has anyone been through this? I can show my PROMPT, the attached document and its answers, in case the problem is me!

:triangular_flag_on_post::triangular_flag_on_post::triangular_flag_on_post: nothing involving RAG will help me, because I do not have access to the code!

2

It seems it can actually get answer from your database so perhaps the problem is only in the instruction/system prompt. You need to know how it is getting answer from your database. Then edit your instruction/system prompt to nudge it to use it.

1 Like
3

Yes, I imagined it was that too. But my prompt is very clear, following all possible methodologies such as CO-STAR, and I also don’t have access to the codes. :sweat: :sweat:

Is there a way to know how it is getting responses from my database without touching the codes?

My fear is that I mess with the prompt too much and the error lies in the files structure. But I can’t find documentation on good practices and examples anywhere (specific for assistants ofc).

Also…Do you know where I can get information on how I should structure a document to add to its database? It seems that the Assistant API is always lacking in explanation for laypeople. :face_with_spiral_eyes:

4

It definitely uses RAG. But you just do not know how it is implemented.

Well, if you are unsure, you do not have to edit the instructions/system prompt. You can just include it in the user message. It is an acceptable way to augment instructions.

There are many discussions in this forum about such topic. For example, this one.