Ask a question

You ask the AI Assistant, in the right pane, a question about your data in plain language, and the answer arrives as a result: a written answer in the chat, and a preview in the center with the chart, the rows behind it and the query that produced them. The AI runs on your machine.

Before the first question

The first time you ask, Totsum shows the notice An AI answers this question: “Your question will be answered by an AI that runs locally on this machine. The generated answer — and the visualizations, data and query steps that come with it — is a first look: before you rely on it, it must be verified by an expert.” Continue sends the question; Continue and don’t ask again sends it and records that you have read the notice.

The pane opens with a short welcome: ask about your data in plain English, then refine with follow-ups such as “Change it to a pie chart”, “Filter to USA only” or “Why is Q4 the highest?”; load a file, then use Suggest some questions; and you can ask how to use Totsum itself.

What you can ask

A chart or a figure: “Show sales by country”, “Top 10 customers by revenue”, “What’s the total revenue?”. The result opens in the center and lands in Explore.
A table: “Show me the raw sales data” produces a grid you can sort.
A subset or a combination of your data: “Filter to USA only”, “Join customers with orders”. The answer is a new result that holds just those rows, and the query that selects them is on its SQL tab, where you can read it or change it (see the SQL page).
A follow-up: “Now filter to 2024”, “Make it a pie chart”, “Change the title”. The assistant remembers the conversation, so a follow-up applies to the last result.
A question that spans several files: Totsum looks for the tables the question needs and joins them on the relationships it found when the files were loaded.
An exploration: “What’s interesting in this data?”, “What should I explore next?”.
Data quality: “Which records look unusual?”, “Are there outliers in the price column?”.
An explanation: “Why is this category highest?”, “Explain the trend in this chart”.
A question about the schema: “What tables do I have?”, “Describe the customers table”, “How many rows?”.
A calculation: “What’s the average of 10, 20, 30?”.
An action: “Delete the sales chart”, “Rename that table to monthly_sales”. The assistant asks you to confirm before it acts.
A question about the app: “How do I build a report?”. The assistant answers from this manual.

Be specific about what you want to see, and name the table or the column when a word could mean two things. A question that a person would understand is a question the assistant can work with.

Create a chart

You create a chart by asking for it in the AI Assistant, in plain language: “Show sales by country as a bar chart”, “Plot price against size”, “Show monthly revenue as a line chart”. The chart opens in a preview in the center, with the rows behind it, its statistics and the query that produced it on the preview’s other tabs. The assistant draws bar, horizontal bar, stacked bar, line and pie charts, histograms, plots, single figures and tables; when you name none, it picks one that suits the result. To change a chart, ask again in the same conversation: “Make it a pie chart”, “Only 2024”, “Color by region”.

When the assistant asks back

If a question can be read in more than one way, the assistant asks which reading you meant, often as a numbered list. Reply with the number, “2” for instance, and it continues. If a request would change something, such as deleting a chart or renaming a table, the assistant asks for a yes before it does it. These replies are part of the same conversation and, on the free trial, they do not count as questions.

While it works

The chat shows what the assistant is doing as it goes: reading the data, writing the query, building the chart. The Send button becomes Stop while a question is being answered, so you can cancel a request that is taking longer than you want.

While the assistant works, a small chip next to the progress line says which kind of answer is coming; it goes away when the answer arrives. Fast means the question matched a ready-made query and is answered at once. Deep means the AI is reasoning through the question and writing a custom query, which can take a few minutes; you can switch to something else, and the answer will be waiting when you return. On most machines, a deep answer is a matter of minutes, not seconds.

Reading the answer

The written answer in the chat is a first look at your data, not an analysis. Every result comes with the means to check it: the rows behind it, the statistics it was built from, the SQL itself, and, when Totsum can check its retelling, the query written out step by step in plain language, on the tabs of the preview in the center. A share (“what share of respondents…”) is taken over the rows that have a value in the column it tests, not over every row of the file; when rows were left out, the answer says how many. The Results page describes them. What you decide to rely on goes to your reviewer through a follow-up report.

Helping the assistant

When fewer than half of your tables, or fewer than half of your columns, have a description, a banner at the top of the chat says “Add descriptions to improve AI answers”, with a Describe My Data… link and a ✕ that hides it for this workspace. Descriptions are what let the assistant know that cust_id and customer_ref are the same thing, or that amount is in cents. The Ingest page explains the three ways to add them.

When Totsum finds a README or a data dictionary beside a file you loaded, the assistant offers to read it: “Found DataDocumentation.txt — use it to describe your data?”. Accept, and a notice reports how many descriptions it added.

The conversation

Enter or Ctrl+Enter sends the question; Shift+Enter or Alt+Enter starts a new line in the input box.
Up and Down, at the first or last line of the box, recall earlier questions.
The ··· menu at the top right of the pane has Clear conversation, which starts a fresh conversation and clears what the assistant remembers of the previous one. The results in Explore stay.

The status bar shows how much of the AI’s working memory the conversation occupies. In a long conversation the assistant has more to read before each answer; clearing it is the way to start light again.

On the free trial

The free trial includes a fixed number of questions about your own files; the licensing page gives the number. Questions asked while only an example dataset is loaded do not count, and the chat says so in a banner: “Example data — questions here are free and don’t use your trial.” Your replies to a clarification or a confirmation do not count either. When few questions remain, the chat says so after each one: “You have 3 free trial questions left. Register a Personal or Commercial license via Help → License… to remove the limit.”