Skip to content
Totsum

Known limits

We would rather you know these before downloading than discover them after. This list is kept current. The free trial exists so that you can verify everything that matters to you before paying.

Hardware and platforms

Section titled “Hardware and platforms”
  • The hardware requirement is real. Totsum needs a GPU with roughly 12 GB of VRAM, or a Mac with an M3 Pro chip or better and 18 GB+ of unified memory, and many current computers don’t qualify. Check the compatibility guide first.
  • Answers take minutes, not seconds. The model runs on your own machine, and short of a top-end GPU you should expect a few minutes per answer. The time depends on your files as much as on your question, so a long wait does not mean that the question was wrong. Ask, do something else, and come back. If the same machine becomes much slower than it used to be, see the FAQ.
  • No Linux build currently. Totsum runs on Windows and on macOS (Apple Silicon) only.
  • Setup needs one download. The installer fetches the AI model (~7 GB) once, during setup. Plan for it on slow connections, or pre-provision offline.

What it answers

Section titled “What it answers”
  • Totsum answers questions your data can answer. It queries the files you import. The model carries general knowledge from its training, but it knows nothing about your company, your clients, or the assumptions made when the dataset was built, such as what a column really means, how the data was collected and what was left out. That context is yours, and a data dictionary can supply part of it.
  • Interpretation will sometimes miss. Like any AI, Totsum will get some answers wrong. That is a certainty over enough questions, and it happens more often on some data and question shapes than on others. Precise questions help, but they don’t make it zero. The verdict on a result belongs to your reviewer, via the follow-up report. Two things soften the miss. With use, you build a feel for which of your questions and files need the closest look. A wrong answer understood at review usually teaches you something about the data itself, such as what a column actually holds or what the dataset silently assumes.
  • Complex queries get no Query Flow. The flow is only drawn when each step can be translated faithfully back to the SQL. When it can’t, Totsum withholds the flow, says so, and points you at the SQL instead. That is deliberate, because a diagram that nobody can verify would be worse than none.
  • Some questions are harder than others, and it is hard to tell which from the outside. What is easy or difficult for the model depends on your files as much as on your wording. When the preview or the query flow does not match what you asked, rephrase the question, split it into steps, or name the file and the column you mean.
  • It is not a data-science workbench. There are no Python notebooks, no model training and no data-cleaning pipelines. Totsum takes you from a file to a verifiable preview and chart, and deliberately nothing more.
  • A model for questions about your data, not a general assistant. The model on your machine is chosen and set up to turn questions about your files into queries: counts, sums, comparisons, trends, top and bottom lists. It is not there to write text, to explain the world or to reason beyond your data. It will not always refuse such a request, so the answer to one is not worth much.

Product

Section titled “Product”
  • English-first. The interface and documentation are in English at launch.
  • Trial limits are firm. The trial gives unlimited questions on the 3 sample datasets, then 10 on your own files. Charts are watermarked, and exploration on screen is complete but there is no export at all. The trial is for verifying that Totsum works on your files and hardware.
  • Data export is Commercial-only. Personal exports the follow-up report, watermarked and without the data section, and lets you save a chart or a query-flow diagram from it as an image. The data section and the frozen sample rows as CSV require Commercial.
  • Self-service support. Support consists of the documentation, the FAQ and a bug channel. We read the reports and don’t promise a reply, and a bug worth fixing gets its fix in a later version. Security and conformity updates are a separate, firm commitment, described in the support model. Your statutory consumer rights are unaffected.

If you spot something missing from this list, tell us. An accurate limits page matters more to us than a flattering one.