How to train a bot where it can answer to product related queries

[ocean dew]

eg:

Brand_Name	Category	Price
ABC	Bat	$12
XYZ	Bat	$14
MNZ	Bat	$13

Now, when I ask

• a) Q: suggest me the cheapest bat brand and its price?
  A: ABC is of price $12
• b) Q: suggest me the cheapest bat brand and its price?
  A: XYZ is of price $14
• c) Q: suggest me top two cheapest bat brand and its price ?
  A: ABC and MNZ is of price $12 and $13
• d) Q: Suggest me bat brands which has price between 12 and 13 dollars?
  A: ABC and MNZ

I tried using embedding approach which is suggested here ( https://stackoverflow.com/questions/74000154/customize-openai-model-how-to-make-sure-answers-are-from-customized-data )
No luck yet

[lilac bay]

What issues are you having with the embeddings approach? Like the stackoverflow answer, I agree that embeddings are the best way to go here

[ocean dew]

@lilac bay with embeddings, we have to prepare training data with all scenarios

prompt	completion
Price of ABC $12
Price of XYZ	$14
Price of MNZ	$13

Now when i ask what is the cheapest price of bat then it is not able to answer it

[ocean dew]

eg: If I ask which is the cheapest product then with embedding it not able to answer it

[worn rapids]

Sorry but I think you might mistake fine-tune for embedding

https://discord.com/channels/974519864045756446/1078442334670307430

I once sent there a tutorial of searching with text embedding

[worn rapids]

fine tune is actually let the model output your desired answer

If you train it: prompt:A Output: B

then if you send A to it it will more likely to answer B

It’s different from storing a certain data into a model

[ocean dew]

It gives random output for some cases
eg:

Now, when I ask
Enter product name Toe Synthetic Flip Flop Sandal
{
"choices": [
{
"finish_reason": "length",
"index": 0,
"logprobs": null,
"text": "\n\nWomen's Shoes ->\n\nWomen's Shoes ->\n\nWomen's Shoes ->\n\n"
}
],

So, I feel both fine tuning and embedding will not solve my problem

[ocean dew]

Example 2:
Enter product namePrice of Josmo 8190 Plain Infant Walking Shoes, Navy
{
"choices": [
{
"finish_reason": "length",
"index": 0,
"logprobs": null,
"text": " 45.99\n\nPrice of Josmo 8190 Plain Infant Walking Shoes, Navy -> 45"
}
],

Code which I used:
import os
import openai

openai.api_key = "sk-xxxx"

query=input("Enter product name ")
response = openai.Completion.create(
model="davinci:ft-personal:s",
prompt=f"{query} ->",
temperature=0.2,
max_tokens=20,
)
print(response)

[worn rapids]

How many times have you finetuned, It seems that only finetuned after many times can the model work as you desired

I personally prefer textembedding

let me have a try

please wait for hours

[ocean dew]

Sure, Thanks

[olive cypress]

Embeddings= teach new information.

Fine-tuning= teaches new patterns.

[worn rapids]

hello

Finished, I just build it in a discord bot

Maybe a little stupid, just a basic one

I love text embedding!!!💕

I'm improving my code for a better chat experience now

DM me if you need the code frame, I will also post here in the support later as a tutorial