Trusted numbers

Numbers come from ForgeQL, not from the model

Whatever a tool returns is saved locally as a dataframe that Cortex can explore with ForgeQL, Smart Forge's SQL query layer. Sum, average, count and group-by become tools it runs, so every figure and every chart is computed from the data, not guessed.

The problem

Ask an AI agent for a number, and it may make one up

A language model predicts text. It is not a calculator, and it cannot hold your whole dataset. Asking it for a statistic directly goes wrong in three ways.

What was our freight spend last quarter?
A model answering from memoryAbout $4.6 million.no query behind it
What the data says$4,812,960.17
  • It hallucinates

    Asked for a total or an average, a model can produce a figure that looks right and is wrong, with nothing to show where it came from.

  • The data does not fit

    A real result can be millions of rows. Not all of it can be pushed into a model, so the model works from a slice and the answer covers only that slice.

  • It burns tokens for a wrong answer

    Every row sent to a model is paid for in tokens. Sending more data costs more, and the arithmetic is still unreliable.

Why it stands out

  • Every result is explorable with ForgeQL

    Whatever a tool returns becomes a table Cortex can query. Sums, averages, counts and grouping are operations it runs, never arithmetic it does in its head.

  • Large results stay out of the model

    A million-row result is saved locally as a dataframe and queried there, so nothing is summarised from a truncated sample.

  • One query across different systems

    Data fetched live from different systems is saved side by side, so a single ForgeQL query can run across all of it, at the moment you ask.

Every result becomes a table Cortex can query

Once a result is saved, ForgeQL works on it. It is standard SQL, so sum, average, count and group all apply. Results that came from different enterprise systems can be joined, too.

  • Each operation is a tool Cortex runs.
  • The answer is whatever the query returns.
  • A join can combine data from your WMS with data from your ERP.

How it works

The usual way, and the Smart Forge way

The usual wayWith Smart Forge
The model adds up the figures itselfEvery figure is computed by a ForgeQL query
Large results are cut down to fit the modelThe full result is saved as a dataframe and queried
Charts are drawn from whatever the model recallsEach chart is drawn from the result of a query