Training your first assistant

An assistant in Docutrain is a document you can talk to. You hand it content, Docutrain reads it, splits it into small overlapping sections and indexes them, and from that point on the assistant answers from that material and nothing else.

The whole thing takes about two minutes of your attention and somewhere between thirty seconds and eight minutes of the computer's. Here is what happens in between.

Five ways in

From your dashboard, Create Assistant opens a window with a tab per content type. Most people use the first one and never look at the rest, which is a shame, because two of the others are the interesting ones.

Audio is the one people are surprised by. Point it at a recorded lecture, a board meeting or a podcast episode and it transcribes the recording, then trains on the transcript like any other text. An hour of talk that nobody was ever going to re-listen to becomes something you can ask questions of.

Form is stranger still. It uploads nothing, because there is nothing to upload yet. It sends you off to build a form, and the responses people submit accumulate into a knowledge base you can then chat with. Forms that become knowledge explains why that is more useful than it sounds.

The one question worth reading carefully

Pick a PDF and Docutrain asks something before it will go any further: What type of PDF is this?

It matters more than a setup question usually does, because it changes what your assistant can show people, not only what it can say.

Choose Document (reports, papers, handbooks) and it pulls individual figures, tables and charts out of the file, ready to drop the right one into an answer when it fits. Choose Slide deck and it captures every page whole, so the assistant can present an entire slide the way you designed it.

Get this wrong on a deck and the answers still work. They are just words where they should have been the slide. You can change the choice any time before the upload starts. What it affects downstream is in images and video.

One related trap. PowerPoint files uploaded as .pptx give up their text but not their pictures, so if you want the slides themselves, export to PDF first and choose Slide deck.

Spreadsheets get read, not skimmed

Assistants think in prose, so a spreadsheet has to become prose before it is any use. Rather than flattening a workbook into a wall of cells and hoping, Docutrain profiles each sheet and writes it out.

Columns are classified first: numeric, date, categorical, rating scales, and free text. Then each type is handled on its own terms. Ratings and numbers become means and distributions. Categories become top-value counts. Dates become trends bucketed by month. Cross-tabulations pair them up, so a survey yields lines like Satisfaction by Region: East mean 4.51 (n=975), and contingency blocks carry row percentages. Reference-shaped sheets are serialized a row at a time as Region: East | Product: Widget, which is what makes a price list or a directory answerable.

Free-text columns matter most and are treated accordingly: the actual comments are preserved as quote lists, tagged with the segment they came from, instead of being averaged into nothing.

The part worth trusting is that every one of those figures is computed in code, never by the AI. The model reads arithmetic that already happened. A mean in an answer is a mean, not a plausible-looking number.

Where a sheet genuinely is mostly numeric with little written text, it still trains, and you get a straight notice saying answers drawn from it may be thin and that a written report would give richer ones. Password-protected files and legacy .xls are rejected outright.

The limits, plainly

Files run up to whatever the upload window states, 30 MB in a typical installation. PDFs get a warning past 300 pages and a hard stop at 500. The fix is to split the file, train on the first part, then add the rest through Retrain with Add to existing data. Up to six documents can be queued at once, and they train one at a time.

The limit you will actually meet is the plan one: how many trained documents your group may keep active, which is one on Free, five on Pro, ten on Business, and uncapped above that. An amber notice appears at 80 percent, and the allowance is rechecked as each queued document comes up, so a large batch stops partway instead of silently overrunning. See plans and limits.

Watching it train

Close the upload window and the Processing panel takes over. It is more informative than a progress bar has any right to be.

Behind it, the pipeline examines the file, extracts the text (transcribing audio, reading scanned pages visually where there is no text layer), chunks that text into overlapping sections, generates the embeddings that let the assistant find relevant passages later, and finishes by writing an AI summary and pulling out keywords. Embedding is almost always the slow part.

Rough timings: a small text document, thirty seconds to a minute. A typical PDF, one to two. Presentations and larger PDFs, two to four. Long documents and audio, four to eight or more.

The card shows you where it has got to, with chips for file size, chunks so far, which embedding batch it is on, elapsed time and time remaining. If your PDF has complex diagrams, an extra reading pass runs and the card says so instead of leaving you to wonder. And if nothing moves for five minutes, the status flips to Stuck and offers Force Retry. Small thing, but it is the difference between a stalled job you can fix yourself and one you have to email someone about.

Documents with slides or figures build their gallery in the background once the text is ready, so you can start asking questions while the images are still arriving.

What you get by default

Left alone, a freshly trained document arrives sensibly configured: your title, a web address derived from it, an AI-generated cover built from its top keywords, an AI summary, and access restricted to your owner group. Private until you decide otherwise, rather than the other way round.

One default is worth revisiting straight away. Source references are hidden in answers to begin with, and since the citations are the reason to trust the thing at all, most owners turn them on. That switch and everything else lives in the document settings.

For more than a couple of files at a time, the same window will queue them with per-file settings. That is the batch upload queue.