Asked again when something moves
Every golden question is asked again after an edit to the glossary, a schema approval, a change to the manuals, a change to your own AI keys, an access profile saved, or a change to what a key reads, and once a week.
Accuracy
How it stays right
If an answer is wrong, you should be the one who finds out first. This is what checks its answers, what it counts, and what it does when it cannot answer, for every application on the platform.
Where answers come from
Grounded in your material, or it says so
Answers come from your records and your written material, never from what a model remembers about the world. When your material cannot support an answer, it says so. If it declines a question your records might answer without having read them, Sonela makes it look in your database before the decline stands.
The figure check
Every figure, held to its source
Every figure in an answer is checked against what the assistant was shown: your records, your manuals, your business context and the question itself. The check is rules, not another AI. An answer with a figure it could not find is counted on Gaps, and is never remembered to answer the next person.
Golden questions
The questions you trust, asked again after every change
Write the questions you know the answers to, and what each answer must do: state a value, never state another, cite a manual, read a table, never read one, return rows, or decline.
Before and after, side by side
Each run keeps every question's verdict beside its last one, so you see whether a change made the assistant better or worse. A golden that fails is asked once more before it counts as failed.
Asked as the people who ask
A golden can be asked through one of your keys, the public one included, or as an access profile, so it tests what that audience is given. A replay spends none of your question credits; on your own AI key, your provider bills its model calls as usual.
By email
A weekly accuracy report, for workspaces that switch it on
Once a week Sonela asks every golden question again and emails the result: passes as counts, failures in the words you wrote them. Everything in it is counted, and nothing in it is written by an AI. It is on every plan, the trial included.
Usage → Gaps
Where it fell short, counted
Gaps lists, for any period, the questions it declined, manuals with nothing on a question, answers given while your database was away, figures it could not verify, tables its queries kept failing on, manuals never used, and answers people reported. It holds counts and the names of your tables and manuals, and no question’s words. The Monday usage email carries the same lines.
From the evidence
A golden question from what went wrong
One form makes a golden question from an answer on the Ask screen, from an answer someone reported, or from a table Gaps says kept failing, so the questions you check are the ones that went wrong.
Verified queries
Approve the query, then hold every replay to it
From What ran, save the database query behind an answer as verified. Each replay of that golden question then runs it, and every figure the new answer states must be one of the values it returns.
A query a new schema no longer accepts is marked stale until a schema accepts it again. The approved statement is not yet shown to the assistant when it answers.
The daily schema check
A changed column, found before a question fails on it
Once a day Sonela compares each database with the schema you approved, reading its structure and none of its rows. Sonela records a dropped or changed column for your review and emails you once; answers keep reading the approved schema until you approve the new one.
When your database is away
It answers from what is written, and says so
If your database or gateway cannot be reached, it answers from your manuals, your business context and your API lookups, and adds one sentence saying the records could not be read. When a gateway goes quiet, we email you once for that silence.
Remembered answers
Fast for the same question, and only where that is safe
The same question asked again can be answered from memory in milliseconds — but never one whose answer read your records or a lookup, and any change to your manuals, context or rules retires what was remembered. You can switch it off, or empty it, in your dashboard.
The record
What ran, and what people reported
Every answer that read your database opens to show the query the AI wrote beside the one your database ran: see what ran. Under every answer, a person can also report it as incomplete or inaccurate; you see the report in your dashboard and can make a golden question from it. The security brief sets out what a report keeps.
Start with the questions you know the answers to
Open a workspace, give it your material, and save the first answers you trust as golden questions.
14 days, 320 question credits, no card. No AI key to supply — trials run on Sonela's key.