When Ask refuses — and what to do

lumnik Pro

lm ask, the TUI's :ask and the PWA's Chat need the AI. On lumnik Open, or with the AI off, lm ask says so: this hub's build has no Ask on Open, this hub has the AI switched off with the AI off.

lm ask turns a question into SQL and runs it, read-only, in one scope. One rule decides every case below: a refusal is better than a confident wrong answer. A refusal costs you a retry. A wrong "there are none" costs a business decision.

The rules in full, for an auditor: Ask — the rules, in detail.

Read an answer

$ lm ask --scope desk "Combien de clients GOLD sont à Lyon ?"
Il y a 2 clients GOLD à Lyon.
  data as of: 2026-10-24T21:35:11.811114Z (last completed run)

  disputed — a source disagrees on a value this answer filtered on:
    P-1008  city = 'Lyon'  counted  t_acme_desk_erp: Lyon · t_acme_desk_crm: Villeurbanne

  COUNT
  2
  1 row returned

  SQL: SELECT COUNT(*)
FROM "connector"."v_acmecustomer"
WHERE "segment" = 'GOLD' AND "city" = 'Lyon'
LIMIT 500

  try: lm conflicts list  — where this scope has source conflicts or lookup ambiguities, row by row

Line	Tells you
the first sentence	the answer, written by the model from the rows below
`data as of`	the last run that fed this scope: the answer is as fresh as that
`disputed`	a record counted on a value its sources disagree on: here, P-1008 is Lyon in the ERP and Villeurbanne in the CRM
the rows	what the query returned. Always shown
`SQL`	the query that ran. Read it: it is what was really asked

You see → it means → you do

You see	It means	Do this
`Error: API error 409: no schema discovered for scope '…' — apply an entity or run discovery first`	Ask knows no table in this scope yet	Run a connector of that scope, or apply an entity. Then `lm source rediscover --scope <tag>`
I have no data declared for this scope yet…	the scope exists, with no table	Same
Je ne peux pas répondre à partir des données disponibles pour ce périmètre — la question ne correspond pas au schéma déclaré.	the question names nothing the scope holds	Rephrase with the words of the columns: `lm source schema --scope <tag>` lists them
… not a declared value of segment (known: GOLD, SILVER, BRONZE)	you asked about a value the entity does not declare	A typo: ask again. A real value: add it to the entity's `values`, re-apply
… not a declared column of this scope	the question needs a column nobody declared	Add it to the entity, or ask with an existing column
… not a table in this scope	the answer would read a table outside the selected scope	Ask in that scope, or add this scope to the connector
I couldn't translate that into a safe query.	the model wrote something that is not one plain read	Ask a simpler question
La requête n'a renvoyé aucune ligne. Il peut n'y en avoir réellement aucune…	zero rows. Maybe none exist; maybe the query missed the question	Read the SQL before saying "there are none"
That filter matches N rows, so the question has no single answer…	a name shared by several records	Ask again with the code the message lists
The answer summarises the first 48 of the 312 rows returned.	the sentence is written from a sample; every row is still printed	For an exact figure, ask for the count
The SQL query stops at 500 rows: there may be more.	the result was cut at the query's limit, 500 at most	Narrow the question
The phrased answer stated a figure these rows do not back, so it is withheld.	the model wrote a number the rows do not support	The rows and the SQL are the answer
I couldn't tie this question to the previous one…	a follow-up lumnik could not resolve	Ask the question in full
I couldn't run that query — the database refused it	the query failed	Read the SQL; if it names a missing table, `lm source rediscover --scope <tag>`

The hub's own sentences follow --lang, else your locale (fr_FR above). The model answers in the question's language.

A zero-row answer is not a "no"

$ lm ask --scope desk "Quels clients sont à Villeurbanne ?"
La requête n'a renvoyé aucune ligne. Il peut n'y en avoir réellement aucune, ou la requête n'a pas saisi votre question — vérifiez le SQL avant d'affirmer qu'il n'y en a pas.
  data as of: 2026-10-24T21:35:11.811114Z (last completed run)

  disputed — a source disagrees on a value this answer filtered on:
    P-1008  city = 'Villeurbanne'  not counted  t_acme_desk_erp: Lyon · t_acme_desk_crm: Villeurbanne

  no rows returned

  SQL: SELECT "name" FROM "connector"."v_acmecustomer" WHERE "city" = 'Villeurbanne'
LIMIT 500

No customer reads Villeurbanne in the fused view. But the CRM places P-1008 there: not counted says so. The disagreement is yours to settle: lm conflicts list lists every one.

Ask questions that work

Use the words of the data. « Quels articles sont en rupture de stock ? » is declined: no column says "rupture". « Quels articles ont une quantité à zéro ? » names the column, and works.
A name or a text. « de la catégorie Pompes » matches the name exactly. « dont la description contient Badu » searches the text. Check which one the SQL used: = or ILIKE.
Dates: write 2026-07-02, or a month with its year: juillet 2026. 01/07/2026 is ambiguous and left to the model. The day is read in UTC, even on an entity that declares a zone:.
An identifier: a terse lookup such as status SO-000004 works when the identifier column declares an example in the entity.

A question that follows the previous one

Pass the previous question with --previous. The PWA's Chat does it for you:

$ lm ask --scope desk --previous "Combien de clients GOLD sont à Lyon ?" "Et à Paris ?"
Il n'y a pas de clients GOLD à Paris.
  understood as: Combien de clients GOLD sont à Paris ?
  data as of: 2026-10-24T21:35:11.811114Z (last completed run)

  COUNT
  0
  1 row returned

  SQL: SELECT COUNT(*) FROM "connector"."v_acmecustomer" WHERE "segment" = 'GOLD' AND "city" = 'Paris'
LIMIT 500

understood as is the question really answered. Read it before trusting the figure.
« Et X ? » replaces a value; « seulement … » narrows; « aussi … » widens.
Only the previous question is remembered, never its rows: « le deuxième » is refused.
How a follow-up is resolved, case by case: the rules.

When an answer looks wrong

The guards lower the odds of a wrong answer. They do not remove it.

Read the SQL that came with the answer.
Fix the grounding: a value or a column Ask should have known goes into the entity manifest; re-apply it. lm source rediscover --scope <tag> refreshes what Ask knows.
Ask again: the same question gives a SQL you can compare.
Report it to contact@lumnik.fr, with the question, the SQL and the scope.