Frequently asked questions

Upload your data. Get a model you can read.

Diafunc is the Symbolic AI Platform. You upload a table and choose the column you want to predict. An automated analysis then discovers a readable model and writes the full report around it. You do not need to write code. Every new account starts with 50,000 free credits, which is about five full analyses, and no credit card is required.

Link to this answer

1. One readable model for each target you chose, with its quality measured on rows the search never saw.
2. A full report in plain language. It opens with a headline and the concerns the analysis flagged, shows the discovered model, and then covers your data, the model and what it means for you.
3. Answers to your own questions, if you typed any in before launch.
4. The model as standalone code, a prediction endpoint for it, and the report as a PDF.

Link to this answer

A short piece of explicit code that says how your target depends on the other columns. You can read every line, check it against what you know about your subject, and show it to someone else. The report prints the model and restates it in plain English right next to it.

Link to this answer

The report tells you, and it does not flatter. The headline states how much of the variation the model explains and how far it beats the strongest simple baseline. A Concerns chapter lists everything the analysis flagged, the critical items first: thin data, a column that gives the answer away, a model that barely beat its baseline. The Model section shows predictions against real values, feature importance and sensitivity. Every claim shows its work: the data, the math, the validation behind it.

Link to this answer

Quality is how well the model predicts, marked down for the model's size, and measured on rows the search never saw. 0 % means no better than predicting the average. 100 % means exact. Because it is measured on unseen rows, a model cannot score well by memorising your table.

Link to this answer

Yes, in three ways.

1. As code. The model exports as standalone code in Python, JavaScript, SQL, R and more. It runs on its own, with no library or runtime to install.
2. As a prediction endpoint. Each target can be deployed as an authenticated REST endpoint that returns a prediction with a 95 % interval.
3. As a document. The report downloads as PDF or EPUB, exports to Excel, and can be shared through an unlisted link.

Link to this answer

Symbolic AI produces results a person can read: explicit structure such as code and formulas. Most current AI produces millions of trained numbers that can predict well but cannot be read. With Symbolic AI the model itself is the explanation, because you can read it from the first line to the last.

Link to this answer

Symbolic Regression is the method underneath. Ordinary regression asks you to pick the form of the model first (a straight line, a polynomial) and then works out the constants. Symbolic Regression searches for the form and the constants together, directly from the data. Ufinq, Diafunc's Symbolic AI technology, is built on it.

Link to this answer

It is the way most people use Diafunc, and it has four steps: create a project, upload your data, launch the analysis, read the results. The analysis profiles your table, searches for a model for each target, tests it on rows it held back, and writes the report. All of it runs without code.

Link to this answer

Very little.

1. The target: the column you want to predict. Diafunc suggests one and says why. You can choose several.
2. Questions (optional): anything you want the report to answer, in your own words. The report answers them in a Questions & Answers chapter.
3. A segment (optional): a column to break the results down by, such as a region or a product line.
4. The depth: how long the search may run. Standard, four hours, is preselected.

Everything else has a sensible default. Before you launch, a check warns you about thin data, a target that never varies or many missing values, so you find out before any credit is spent.

Link to this answer

No. You upload, choose, launch and read. If you do write code, the whole platform is also open to you through notebooks, the API and the command line.

Link to this answer

Any table: rows of observations, columns of measurements or attributes. Diafunc imports dozens of formats, among them CSV, Excel, Parquet and JSON, as well as SQLite and the files of statistics packages such as SPSS and Stata. You see a preview before anything is imported, and delimiters and encodings are detected for you. You can also connect a source that refreshes: Google Sheets, an HTTPS URL, S3-compatible storage or a SQL database.

Link to this answer

There is no fixed minimum. Large tables are fine: size is bounded only by your credits. As a guide, the check before launch warns you when a table has fewer than about a hundred rows per target, because a quality figure measured on so few rows is uncertain. More rows help, and so do columns that plausibly have something to do with the target. If your data is thin, the report says so in its Concerns chapter and tells you how to read the result.

Link to this answer

Five kinds:

1. A quantity, such as a price, a temperature or a duration.
2. A count, such as visits per day.
3. A rating, such as a score from 1 to 5.
4. Yes or no, such as whether a customer renewed.
5. A category, such as a product type.

Diafunc suggests the right reading for each target and you can change it. The choice shapes the report: a yes-or-no target, for example, gets the chapters that make sense for classification.

Link to this answer

You set the time, from one hour to a full day. The first chapters of your report appear within the first few minutes, so you can start reading about your data early. The search for the model then runs for the depth you chose:

The time is a ceiling for the whole run: the search stops when the time or the credit budget is reached, whichever comes first. You can close the browser and have Diafunc email you when the analysis has finished. A run can be canceled at any point.

Link to this answer

A list would hand the hardest part of the job back to you. While it runs, the search compares a very large number of candidates on accuracy and simplicity. It returns the one that holds up best on rows it never saw, together with the evidence. If you chose several targets you get one model for each.

Link to this answer

It can. A connected data source refreshes when you ask it to or on a schedule you set, and an analysis can be re-run once or on a schedule. Together that gives you a report that is re-run on a schedule as your data grows.

Link to this answer

AutoML tools try many kinds of model, tune them, and return the most accurate one. That is usually a large ensemble or a neural network: you can call it, and you cannot read it. Diafunc searches for a model small enough to read and tells you what it gives up for that, because the report sets the model against simple baselines on unseen rows. You also get the written report, which AutoML leaves to you. If raw accuracy is all that counts and nobody will ever ask why, a black-box model can be the right tool.

Link to this answer

A chatbot writes a plausible answer in seconds. Ask again and the answer may differ, and the figures in it are hard to check. A Diafunc analysis works through your whole table, holds rows back, searches for a model for an hour or more, and measures that model on the rows it held back. The result is a piece of code that is the same tomorrow and gives the same prediction for the same input every time. You can use both: a chatbot for a quick look, Diafunc when someone will act on the answer.

Link to this answer

A trendline fits a shape you pick (a line, a polynomial, an exponential) to one input column. Diafunc looks for the shape itself, across all your columns at once, and then tests the result on rows it did not use. If a straight line through one column answers your question, the spreadsheet is enough, and the report will tell you so: it always states how far the discovered model beats the strongest simple baseline.

Link to this answer

A dashboard shows what happened: totals, trends, breakdowns. Diafunc discovers the relation underneath, meaning how your target depends on the other columns, as a model that predicts and can be read. The two answer different questions and sit well side by side. Diafunc does not replace your dashboard.

Link to this answer

You, and the people you invite. A project is private by default and visible to its members only. You grant access per person, with a role for each, and a project becomes public only if its owner makes it public. A report can also be shared through an unlisted link, which you can let expire or revoke at any time. The recipient sees the report and never the uploaded data.

Link to this answer

In the EU. Diafunc runs on dedicated servers in Germany and Finland, and data storage and email delivery stay within the EU as well. Connections are encrypted, and stored secrets such as data-source credentials are encrypted at rest. Diafunc is a hosted service; there is no on-premise version.

Link to this answer

It never sees the rows of your table. A third-party language-model provider writes the narrative parts of a report and powers the Assistant and the suggested setup. For that it receives column names, summary statistics, the discovered models with their scores, the questions you asked, and group labels such as region = NW. The rows of your table are never sent.

Link to this answer

No. Diafunc trains nothing on customer data. Your table is the input to your analysis and to nothing else.

Link to this answer

Yes, and deletion is immediate once you confirm it. Deleting a project removes everything in it, and the data is physically purged within 24 hours. Deleting your account removes its projects, uploads, tokens, chats and analyses, with no waiting period. Invoices are kept for the period the law requires, stripped of identifying data.

Link to this answer

Yes, at any time. Tables export in dozens of formats, including CSV, Excel and Parquet. Reports export as PDF, EPUB and Excel, and models as code.

Link to this answer

50,000 free credits to start. That is about five full analyses, and the credits don't expire. On top of that the Free plan adds 1,000 credits every month. You get the full platform, and you do not need a credit card.

Link to this answer

Start free. Scale with credits. Every plan includes the full platform. What you choose is the credit allowance.

Annual billing saves up to 25 %. Prices are the same figures in euros and US dollars, and the listed price is the total you are charged. Free and Premium come with the documentation, the community and the Assistant. Ultimate adds professional support, with requests answered within three business days.

Link to this answer

Credits are the one meter for everything you use: compute, storage, transfer and AI usage. For an analysis, the depth you choose sets the budget: Express 500 credits, Quick 2,000, Standard 10,000, Thorough 40,000, Deep 150,000. A run is charged at most its budget, and it starts only if your balance covers the whole budget, so an analysis never stops halfway for lack of credits.

Link to this answer

Nothing is lost. Your projects stay visible, your data is safe, and an analysis that has already started runs to the end. New work that consumes credits waits until your monthly credits refill, or until you upgrade or add an Extension Pack. Unused monthly credits do not carry over to the next month. The 50,000 credits you started with do.

Link to this answer

Yes. There are no setup fees, and you need no credit card to start.

Link to this answer

Yes. A team has one subscription at €99 per month, which brings 100,000 pooled credits, an admin console, an activity log, single sign-on (SAML) and SCIM provisioning. Each member is a seat at €19 per month. A team of three therefore pays €156 per month. On annual billing the subscription is €79 and a seat €15 per month. Extension Packs add credits to the pool.

Link to this answer

The whole platform is an API. Everything you can do in the browser is available over REST at /api/v1/..., with API tokens scoped to a project or a team, structured errors, and an OpenAPI spec for each service.

Link to this answer

Yes. Diafunc runs an MCP server at diafunc.com/mcp. An agent that speaks the Model Context Protocol connects with an API token and can work with projects, tables and analyses the way you do.

Link to this answer

Yes. Install it with npm install -g @diafunc/cli, or with the installers for Linux, macOS and Windows. It signs in with a device code, prints JSON (TSV when piped), and includes an interactive shell.

diafunc auth login
diafunc analyze data.csv --targets price --hours 4 --budget 10000
diafunc project list -o json

Link to this answer

Yes, with the FDK, the Function Development Kit. You write functions in Java, Kotlin, Scala, JavaScript or Python, they run on the platform, and they can reach all the data in your project.

Link to this answer

Ufinq is Diafunc's Symbolic AI technology, built on Symbolic Regression and years of foundational research. It discovers small, readable programs that describe how data behaves. Candidate programs evolve in parallel and compete on accuracy and simplicity. Each one is verified, simplified and has its constants tuned along the way, and one readable model comes out at the end.

Link to this answer

Benchmark-tested. Evaluated on SRBench, the standard Symbolic Regression benchmark. A benchmark sweep runs every week on public datasets, and the results are published at ufinq.com/benchmarks, with no cherry-picked workloads. We make no claim to beat anyone.

Link to this answer

No. Ufinq is open in its ideas and closed in its source. The papers, designs and benchmark results are public at ufinq.com. The technology itself is available through the Diafunc Platform: in the browser, over the API, from the command line, through the MCP server and the FDK. That keeps one maintained, benchmark-tested version in service for everyone, and it is how an independent company pays for long-term research. You can start free.

Link to this answer

Diafunc is an independent Austrian company, founded and operated by Markus Pollak in Purbach, Austria. You can reach it at contact@diafunc.com or through the contact form at diafunc.com/contact.

Link to this answer

The product is in English: the interface, the reports and the documentation. You are welcome to write to us in German.

Link to this answer