Quick answer
How to start with Docsi AI: drop one readable file onto the widget at the top of your screen, describe the topic, scope and format (“Summarise chapter 2 as a one-page study sheet in PDF”) and let the task run. Docsi reads the file locally, sends only the text it needs to Docsi AI and returns a finished file with source markers. Check two or three important statements on the page given before you share the result.
Key takeaways
- Start with one readable file and one clear goal.
- A good request names source, result, length and format.
- Docsi reads files locally and sends only the text needed to Docsi AI.
- Source markers show file and page; check the most important statements there.
- Phrase corrections in the same conversation instead of starting over.
Prepare your file and goal
An active subscription unlocks Docsi. The small widget lets you work at the edge of your screen; the main window gives you more room for several sources, history and longer documents.
Choose a file you can read yourself, such as a lecture PDF. A specific task names the topic, length and output format. “Summarise chapter 2 as a one-page study sheet with five review questions” is more precise than “Do something with this”.
From input to result
- Drop the file onto Docsi or add it in the main window. Check that its content was recognised as a source.
- Choose a useful suggestion or write what you need. Include the language and file format you want.
- Let the task run. Docsi AI tasks need a connection; conversions and PDF merging are processed locally.
- Read the result, compare important statements with the source and save a separate output file.
Ask a useful follow-up
A conversation keeps its context. After a result, try “Explain point 3 more simply” or “Add a comparison table”. If you want a finished file, say so explicitly; Docsi can answer a simple question in the conversation.
A request made of four building blocks
A good first request answers four questions: what is the source, what should be produced, how long may it be and which format do you need? “Turn the project brief into a discussion prep with five open questions, one page at most, as a Word file” contains all four. If a building block is missing, Docsi has to guess or asks once – that is how he is built: for an unclear request he asks one precise question instead of producing an arbitrary document.
It also helps to add one sentence about the audience (“for a second-semester exam”, “for my team lead”) and a rule for gaps (“Do not invent dates; flag missing information”). Such boundaries work better than long style wishes. General guides on writing AI requests recommend the same: keep context, task, format and examples clearly apart.
What happens behind the scenes
Docsi first reads your file locally: PDFs page by page, Word files with headings, lists and tables, Excel workbooks with every sheet. Only when an AI task starts do the excerpts it needs go to Docsi AI, with numbered source markers. The assistant works in up to ten steps, reads further pages when needed, calculates table values exactly and then decides whether a short reply or a document fits.
Before you see the result, Docsi matches every source marker against the original text. Markers without a matching passage are removed. That does not replace your own reading, but it makes visible where a statement comes from. Your original file stays unchanged; Docsi always saves the result as a new file.
Making good use of the first week
The free trial lasts three days and includes 150 credits. Use it with real but manageable tasks: a study sheet from one chapter, an analysis of a small table, a merged PDF. Local tools such as converting, merging and text recognition cost no credits, so you can try them without limits.
Note which wording gives good results and turn it into your own templates. If a result is too long, say so in the same conversation (“Shorten it to one page”); Docsi keeps the context of the previous steps. For a new, unrelated task, start a new conversation so that old sources do not carry over.
Widget or main window?
The widget is the slim bar at the top of your screen. It suits quick requests with one file or one window: drop it in, write a sentence, done. When you hover over the bar, a small input field opens with your credit balance. Docsi answers short questions right there without creating a file.
You need the main window as soon as things get bigger: several sources at once, a longer conversation, the preview of a PDF with text recognition, or the history of earlier results. It also holds the settings for language, AI, updates and your subscription. Both views share the same sources and results, so you can start in the widget and continue in the main window.
Checking the first result properly
Do not just read the first result for style but for content. Three questions help: are all important points from the source included? Do figures, names and dates match at the locations given? Has anything been added that the source does not say? The last question matters most, because language models sometimes add plausible-sounding general knowledge.
If something stands out, correct it in the same conversation: “Point 4 is not in the document, please remove it” or “Back section 2 with verbatim quotes”. Docsi then revises the existing result instead of starting over. Only once the content is right is it worth refining layout and format.
A worked scenario
Suppose you have a twelve-page journal article as a PDF for a seminar and need to present its key points in five minutes tomorrow. You drop the PDF onto the widget and write: “Summarise the article for a five-minute talk: research question, method, three findings, one limitation. One page at most, Word file, sources with page numbers.”
Docsi reads the article, searches for the relevant passages and delivers a one-page Word file. You check the three findings on the pages given, notice that the limitation is phrased too generally, and write in the same conversation “Make the limitation precise using the wording on page 9”. You then add your own transitions in Word and have a solid basis for your talk within minutes.
Key terms explained
- Widget
- The slim Docsi bar at the top of the screen. It opens when you hover over it and accepts files, windows and quick questions.
- Source
- A file, window or pasted text Docsi reads for a task. Sources are numbered (Q1, Q2 …).
- Source marker
- A reference in the result to a passage of a source, with the page number for PDFs. Docsi checks every marker against the original.
- Credit
- The unit of usage for the built-in Docsi AI. Local tools use no credits.
- Conversation
- A connected sequence of requests. Docsi considers up to 16 earlier steps in it, but no other conversations.
A worked example
You have a six-page project brief and need to prepare for a discussion. Start with that file alone and ask about the goal, unresolved points and next steps. A precise first task shows whether the text was imported completely. Request a short overview first; an extensive document becomes useful after checking the sources. If the brief mentions attachments you have not added, the result needs to identify that gap explicitly.
Adapt source and scope, then use the text in Docsi. This website starts no AI task.
Check the result
If something is wrong
| Observation | Next step |
|---|---|
| Task does not start | Check licence status and the selected AI connection. Local functions also require an active licence. |
| Answer does not match the file | Inspect the source list and imported text. Restart with a clearly bounded section. |
| Document is too long | Specify the audience, section count and desired length in a follow-up task. |
Common questions
Is my original file overwritten?
The import acts as a source. Save the generated result under a separate name and review it before sharing.
Why do I get a reply instead of a file?
Docsi distinguishes questions from document requests. Name a format such as PDF or Word and describe the document you want.
Do I need experience writing AI requests?
No. One sentence with source, goal, length and format is enough. If something is unclear, Docsi asks one specific question.
Can I use several files at once?
Yes, in the main window. Name the role of each file in your request, for example “slides are the main source, notes add detail”.
Where do I find the finished result?
In Docsi’s history. From there you save it as PDF, Word, Excel, CSV or Markdown wherever you like.
Can I ask Docsi something without a file?
Yes. Docsi answers short questions right in the widget. Without a source, however, there are no source markers, and the answer rests on the model’s general knowledge.
Sources and further reading
Statements about Docsi are based on the product documentation by Impulse Apps. These primary sources explain the underlying formats, standards and methods.
- Anthropic: Prompt engineering overviewPrinciples for clear requests to language models: context, examples, format.
- OpenAI: Prompt engineeringGuidance on how instructions, roles and output format steer a result.
- Mozilla PDF.jsThe library Docsi uses to read text from PDF files page by page.