TablesSQLDatabaseTool does not use similarity criterion on selecting the tables from few_shot_examples #474
rafaelpadilla
started this conversation in
General
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
First of all, congratulations for such project and keeping it opensource 🙌
I see that the tool TablesSQLDatabaseTool uses examples provided by the user to retrieve the most similar tables.
The tool uses the few-shot examples to complete the final answer, which includes:
few_shot_examples
However, IMHO, this approach is strange, since it does not use any criterion on selecting the tables from the
few_shot_examples
.In a case where the few_shot_examples contains a query:
"SELECT BRANDS FROM TABLE_WATCHES"
The table
TABLE_WATCHES
would be included (with the highest similarity score) for any input question. Even if the input question is totally unrelated to the TB_WATCHES. For instance: "How many liquid products are the in the warehouse?"I'm pretty sure you guys have baseline datasets to evaluate the precision of the output SQL queries. And I believe that including a similarity criterion in the few_shot_examples, the results will be improved.
Beta Was this translation helpful? Give feedback.
All reactions