[alberthflores95]

Problem

The analytical sub-queries behind /v1/artifact are written by an LLM that sees only the main SQL and a one-line dialect note: no semantic layer, no server version. In the first GONAC artifact on ClickHouse 22.1, two of three failed and came back empty:

• SUM("facturado") AS facturado, SUM("facturado") / nullIf(SUM("objetivo"), 0): the alias shadows the column, so the second SUM(facturado) is a sum of a sum (ILLEGAL_AGGREGATION).
• SUM(SUM(x)) OVER (...) / nullIf(SUM(SUM(y)) OVER (...), 0): a window result cannot be used inside another expression of the same SELECT, not even through its alias (UNKNOWN_IDENTIFIER).

Nothing retried them. Separately, when no dataset has rows, the document prompt asked for "realistic benchmarks" and "indicative estimates", so the document printed invented figures that read like real ones.

Change

• One repair per failed sub-query. The LLM gets the failed SQL and the database error and returns a corrected statement, which runs once more. A statement that fails twice keeps its error. Repairs run on the request thread because the token tracker is a ContextVar the pool threads do not inherit. The user's sql_override is never rewritten.
• _ERROR_HINTS explains the errors whose text does not name the fix.
• The ClickHouse dialect note carries both rules, worded so the SQL they ask for is valid on any version.
• A failed main SQL is an error, and gets the same repair. run_query_pipeline reports a failure in answer with rows=None instead of raising, so the artifact used to show an empty main dataset with no error and derive the sub-queries from the failed statement. Seen live on the PoC: the chat's own SQL carried SUM(facturado) AS facturado next to SUM(facturado) / SUM(objetivo). If the repair fails too, the main dataset carries the error and no sub-query is built on it. A query that matched nothing is not a failure.
• No data means no figures. The document states the data could not be retrieved instead of estimating.

Verification

• tests/unit/test_artifact_sub_sql_repair.py: 15 tests. The first 10 fail against main, and 3 of the 4 main-SQL tests fail against the first commit (the fourth guards the no-rows case). Together with test_artifact_env.py: 22 passed. Ruff clean.
• Measured with the real model (Anthropic Sonnet 5 on the GONAC PoC server, throwaway container), each repaired statement executed on the production database:
  • Without the hint: the window case failed again, because the model moved the aggregation into a subquery and kept the ratio in the same SELECT.
  • With the hint: 6 of 6 ran (3 per case), with figures matching a hand-written query: cumulative invoiced 3,739.7 M, cumulative progress 96.6 %.

Out of scope

The main chat SQL generator (ask-sql-generation/.../prompts/clickhouse.py) lacks the same two rules and does not retry on execution errors either (BACKLOG group P). Changing it affects every question and needs a benchmark re-run.

🤖 Generated with Claude Code