An AI assistant that answers from your own documents: what we learned building one

Every firm has documents that already answer most of the questions people ask: BIM execution plans, modeling standards, product manuals, course materials. And people still ask a colleague, because finding the paragraph takes longer than asking.

An AI assistant that answers from your own documents closes that gap. The technique is called retrieval-augmented generation, or RAG. This article shares what we learned building one, and where the same approach fits a BIM team.

The project

DIM is the AI assistant of Parametric Center, a training center in Ukraine that teaches Autodesk software. A disclosure: Parametric Center shares its founder with Consedge Tech, and our team built DIM for it.

The brief was narrow on purpose. Students ask questions about their course at all hours. The answers are in the lessons. DIM answers each student from the lessons of the course that student is enrolled in, on the website, in Telegram and by email.

How it works

1. Preparation. Course materials, including video lessons as transcripts, are split into passages and stored with labels: course, module, lesson.
2. Identification. The assistant works out who is asking and which course they are enrolled in.
3. Retrieval. The question is matched against passages from that course only.
4. Answer. A language model writes the answer from the retrieved passages and points to the lesson it came from.
5. Fallback. If the passages do not cover the question, the assistant says so and hands over to a person.

The course filter sits before retrieval, so a student never gets an answer built from a course they are not taking.

None of these steps is exotic. The lessons are in how each one fails.

Six lessons

1. Document preparation is most of the work

The quality of the answers is set before the model is involved. A transcript split in the middle of an explanation gives half an answer. Two versions of the same lesson give two contradictory ones.

Most of the effort went into the material: cleaning transcripts, splitting by topic instead of by length, removing superseded versions, and labelling every passage.

2. Access rules belong in retrieval, not in the prompt

A student should get answers from their own course and no other. The reliable way to enforce that is to filter before the search. The model never receives a passage the user may not see.

Asking the model in its instructions to "only use course A" is a request. A filter is a guarantee.

3. "I don't know" is a feature

A wrong answer about a course costs more than a missing one. The assistant answers only from retrieved passages, names its source, and says plainly when the material does not cover the question.

That honesty is also what makes people trust the answers it does give.

4. One brain, several channels

Students use the website, Telegram and email. The answering logic is one service, and each channel is a thin adapter. The hard part is identity: knowing that a Telegram account and a website login are the same student.

5. The questions it cannot answer are the most useful output

Every unanswered question is logged. Read together, they show where the course materials have a gap. The assistant becomes a feedback loop for the content itself.

6. Test with a fixed question set

Before each change to prompts, models or materials, the same set of real questions is run again and the answers compared. Without it, an improvement for one question quietly breaks another.

Where a rule engine beats a language model

The same lessons apply when the subject is a building model instead of a course, with one addition.

Many BIM questions have exact answers. Is this parameter filled in? Does this view name follow the convention? Is this door family on the approved list? A language model can answer those, but it is the wrong tool. It is slow and costly at scale, and it can be wrong.

A rule engine answers them the same way every time. Rules written as plain data, in a format such as YAML, can be reviewed by a BIM manager and versioned like code.

The language model earns its place around the rules:

• finding the paragraph of the standard a rule comes from,
• explaining a failed check in plain words,
• drafting a rule from a sentence in the BIM execution plan, for a person to approve.

Rules decide. The model explains. That division keeps the result auditable.

Is it worth it for a BIM team?

The approach fits when three things are true:

• The answers exist in writing, even if scattered across PDFs, wikis and recordings.
• The same questions come back: onboarding, standards, "how do we do X here".
• Someone owns the content and will fix the gaps the assistant exposes.

Good candidates are BIM standards and execution plans for new hires, support for a software product, technical catalogs for a manufacturer, and training material.

It is a poor fit when the knowledge lives only in people's heads. An assistant cannot retrieve what was never written down.

What to ask before you build

1. Who will ask, and what do they ask most often? Collect fifty real questions.
2. Which documents hold the answers, in what formats, and who keeps them current?
3. Who may see what?
4. Which channels matter: a website widget, Teams, Telegram, email?
5. Who takes over when the assistant cannot answer?
6. What are the data rules: where may documents be stored and which model providers are acceptable?

If you already have a prototype, request a Production Readiness Audit. If you are starting from the questions above, a Feasibility Review gives you a written verdict on scope and risks. Our AI engineering page describes the work, and this article covers the same demo-to-production gap for Revit plugins.

FAQ

Which AI models does an assistant like this use?

The design does not depend on one model. We choose per project for language quality, cost and data rules, and keep the model replaceable.

Is our data used to train the models?

Not with the API access we use, where customer data is not used for training. Documents stay in storage you control.

What happens when the answer is not in the documents?

The assistant says so and hands the question to a person. The question is logged, so the gap can be closed.

Which document formats work?

PDFs, wiki pages, spreadsheets and video through transcripts. Scanned documents need text recognition first.