Discourse AI + Data Explorer?

[EricGT]

@simon

SELECT * FROM poll_votes WHERE poll_id = 74;
 poll_id | poll_option_id | user_id |         created_at         |         updated_at
---------+----------------+---------+----------------------------+----------------------------
      71 |            270 |       2 | 2023-08-14 06:59:00.382873 | 2023-08-14 06:59:00.382873
      71 |            271 |     121 | 2023-08-14 07:02:57.364892 | 2023-08-14 07:02:57.364892
      71 |            271 |       1 | 2023-08-14 07:03:36.464304 | 2023-08-14 07:03:36.464304

Is that a typo?

Should WHERE poll_id = 74 be WHERE poll_id = 71.

---

I did not check the entire prompt, was just looking to see if you included an example of what the results of the query should look like. In other words you gave examples of the table values but I did not see an example of the expected result for a prompt, perhaps I missed it. The expected result could be used to validate if the SQL is correct.

---

Suggestion:

In the example Discourse database prompt the use of 1 for id is used in multiple tables. While as humans we get that the 1 only has meaning with the context of the field and the table, the AI will not know this and that could give the AI a chance to make a choice. Therefore as a suggestion modify the Discourse database prompt to use different numbers for each id.

To go one step further, check that all of the values in each example table is a single token by using the OpenAI Tokenizer page. I know you may want to keep some of them as words or even worse strings but does the AI really care? Will values of multiple tokens lead to more variation and possible hallucinations?