Stiekema Studio

Step 1 / 6

Understand first, then answer

A question like "travel expenses" can mean anything: one specific question about a single scheme, a comparison between two documents, or an overview across the whole collection. DocumentChat recognises what you probably mean and picks the approach to match.

If the question is too broad or context is missing, the assistant does not guess but asks one specific follow-up question first. And operations that take real work only start once you confirm that choice.

Guided intake

“difference in travel expenses old and new scheme”

Answerone specific question

Comparetwo schemes side by side

Summarisea whole document

The assistant asks: which two schemes do you want to compare?

[1] the intake recognises the intent and asks a follow-up when in doubt

Step 2 / 6

Searching the way an expert does

Behind the scenes your question is turned into several searches that run over the collection at the same time. It searches by meaning and by literal text, so that both "commuting allowance" and "travel expenses" hit the mark - even when the document uses slightly different words than you do.

If a search round turns up too little of use, the system widens its view automatically. Results are merged and de-duplicated, so the language model gets the best possible selection of source material.

The search layer

Searches by meaning and by text:travel expensescommuting allowance

[2] meaning and keywords, in parallel across the whole collection

Step 3 / 6

Answers that account for themselves

The language model gets only the document fragments that were found as its context, and is strictly instructed to answer from those alone. Every factual claim gets a footnote pointing to the exact document and the page.

If the answer is not in the documents, or only partly, the assistant says so plainly. Answers appear on screen as they stream, with the sources in a panel alongside that folds open.

The conversation

How much annual leave do I get after 10 years of service?

At ten years of service the accrual is 25 statutory and 3 non-statutory days1. From the age of 45 a seniority day is added2.

Sources [1] cao chapter 6, p. 41 · [2] annex 3, p. 87

[3] a footnote per claim, a page per footnote

Step 4 / 6

Comparing, from two to all of them

For questions like "what is the difference between scheme A and B" the system groups the context it found per document, so that the answer sets the two out side by side instead of mixing them up.

If the question covers many or all documents at once, DocumentChat switches to a separate mode: every document is assessed on its own and the outcomes land in a readable comparison table. Work that would take days by hand is there in minutes.

Comparison table

Subject

Scheme A

Scheme B

Travel allowance per km

€ 0.211

€ 0.232

Home-working allowance

€ 2.151

none2

Start date

1 Jan1

1 Jul2

Sources [1] scheme A, art. 9 · [2] scheme B, art. 11

[4] assessed per document, brought together in a table

Step 5 / 6

Document processing that respects structure

Before there is anything to ask, documents have to be read properly. The pipeline handles PDFs and Word documents, scans included, and picks per document the reading strategy that gives the best result.

When splitting them into searchable fragments, the system follows the document's structure of meaning - chapters, articles, tables - instead of cutting bluntly on length. Tables stay intact, with their heading and surrounding context attached. It is precisely in tables, such as salary scales, that the real answer often sits.

The document pipeline

Ingestpdf & docx

Readscans too

Splitby meaning

Enrichthemes

Indexsearchable

Tables stay intact, with their heading and context attached.

[5] from pdf to searchable fragment, without loss of structure

Step 6 / 6

Quality you can see

Administrators see per document how well the processing went and where attention is needed. For every question asked, it is possible to look back diagnostically at what the system did: what it searched for, what it found and what the answer rests on.

Under the bonnet, a suite of over a thousand automated tests watches the behaviour of the system on every change. That keeps the quality standing as development continues.

The memory

Document

Version

Quality

Status

2026-03

current

2025-11

current

2026-01

current

[6] every processing run recorded, every question traceable

From question to grounded answer, in six steps.

Understand first, then answer

Searching the way an expert does

Answers that account for themselves

Comparing, from two to all of them

Document processing that respects structure

Quality you can see

Curious what this does with your documents?

Pick a time now

Book a demo

Send a message

Call me back