BlazeSQL in AI Database Agents

BlazeSQL failed to surface unanswered questions as failures. In the recording, Query review showed one ordinary row marked 'Not verified' and no failures or unanswered view appeared anywhere in the sections shown. The social-media complaints question appeared in Chat with a normal answer, and the other unanswerable question is not shown in the admin view.

BlazeSQL answered the plain fact in prose, so the evidence does not show the result being confined to a chart or a one-cell table. In the captured run, the reply said there were 100 orders ever, and the same frame also showed a one-row result table and no chart.

BlazeSQL passed this scenario: it showed the same revenue answer in a different form. In the recorded conversation, the first result appeared as a table with seven categories, then the follow-up request switched that same result to a bar chart. The bar lengths matched the table values within the chart's scale.

BlazeSQL replaced the earlier July answer with the corrected June answer. In the visible exchange, the first turn returned 21 orders for July 2026, and the follow-up turn rebuilt the query for June 2026 and returned 20 orders, matching the corrected period only.

BlazeSQL offered what to ask next after answering the revenue question. It ended the August 2026 revenue reply with “Would you like a breakdown by category, or any other analysis?”, so the capture shows one specific follow-up rather than only a generic closing.

BlazeSQL understood the follow-up as referring to the prior answer. In the first turn, it returned the top five customers by spend. In the follow-up, it counted orders for those same five customers, kept them in the same order, and did not ask who they were again. The shown SQL carries that customer list forward in a named CTE.

BlazeSQL failed this scenario. In the shown run, it gave a confident answer instead of saying it could not answer, even though the needed table was excluded and the SQL used only permitted tables.

BlazeSQL said it could not answer from the available data when asked how many 5-star reviews were received this month. It listed ten available tables, gave no number, and pointed to adding review data or using a related source instead. The key finding is that it declined rather than inventing a count.

BlazeSQL did not answer revenue as a bare number; it stated the definition it applied and showed the breakdown it used. In the captured Sep 2, 2026 run, it reported order revenue, subscription revenue, refunds, and an adjusted net revenue total, with the order line explicitly marked as excluding cancelled orders.

BlazeSQL saves an answer and its query, then reopens the saved item with the same three rows under the saved name. The saved item appears in a separate Saved area, and the reopened table matches the earlier chat answer.

BlazeSQL used the user’s documentation to answer from the cryptic table and column. It reported 6 returns currently awaiting manual fraud review, split across 5 approved and 1 received. The stored note was applied to returns.flg_x2, and the follow-up answer came from that documented column.