Docs › How answers are verified
How answers are verified
ProLyt's position is simple: computation the platform owns is safe by construction; computation the model emits needs checking. The model writes the query; the database computes the numbers. Fixed rules around the model then decide what reaches you. This page describes those rules as they work today — including their limits.
The pipeline
- A router decides whether the question is about tables, documents, or both.
- Table search picks the few tables most relevant to the question (6 by default) and shows the model their names and columns — never whole tables.
- The model writes SQL and calls a read-only query tool; the database runs it and returns rows (at most 50 rows and about 6,000 characters per result are shown to the model).
- For documents, the model searches passages or reads a document in capped windows.
- The model writes the answer; the guards below run on it; the answer, its SQL, the tables read and the passages are shown together.
The guards
| Guard | What it does | What it does not do |
|---|---|---|
| Read-only SQL | Statements that write — INSERT, UPDATE, DELETE, DROP, CREATE, ALTER and similar — are refused before reaching the database. | It does not block read-only functions a database exposes. Connect with a read-only user; the role is the primary control. |
| Grounding | An answer that states figures while no query read your data and no document was cited is replaced by a refusal, and the credit is refunded. | An answer backed by a query is shown as is — its figures are not re-checked one by one against the rows. The SQL is shown so you can. |
| Unverified figures | In a document-based answer, every number that does not appear in the passages the model read is listed in a chip: *"N figures not found in the documents read"*. | A flagged figure may be a correct calculation, a reformatted value, or an error — the chip states presence, not correctness. |
| Join check | When the executed SQL joins tables in a way that is likely to multiply rows (a one-to-many join aggregated the wrong way), ProLyt rewrites the aggregation deterministically where it can, and otherwise warns on the answer. | Relationships confirmed on the schema map are recorded as the team's declaration; automatic enforcement of confirmed relationships is planned, not built. |
| Completeness | When a list is built from documents and only part of the corpus was read, the answer says so, or is refused instead of being served as complete. | It applies to document enumerations, not to SQL results. |
| Empty answers | An empty final answer is replaced by a platform message and refunded, whatever ran before it. | A rendered chart counts as an answer. |
| Prediction refusal | A question asking for a forward number is declined before any model call, at no charge, with the Forecast page offered instead. | A question about the past that mentions the future in passing may be misread; the decline costs nothing and can be rephrased. |
| Charts | A chart is drawn from the rows of the query it names; a request for columns the query did not return is refused with the available columns. | It cannot tell whether the query was the right one. |
What the model never gets
- Your whole dataset. It sees table names and columns, the capped rows its own queries return, document passages, and — on plans with memory — recent questions in the same chat with their SQL.
- Write access. Every query goes through the read-only guard and your database user's permissions.
- Your prediction. Forward numbers are the Forecast page's (deterministic) or yours (recorded in the ledger with your name).
What none of this proves
That the model chose the right query. Two defensible readings of "net revenue" produce two different correct-looking numbers. That is why every answer shows its SQL and its sources, why a workspace can keep a metric glossary, and why an answer can be sent for review by a colleague. Measurements of how often the model picks the right table and the right definition are published in How we measured.