Searchable project files: a knowledge base with AI

An AI knowledge base first looks in your project documents for the matching passages. Only then, and only if you want it, does a language model write an answer from those passages and name the source reference. The method is called retrieval-augmented generation, RAG for short, and whether it works for you depends heavily on the state of your files.

The principle: look it up first, then answer

A language model does not know your projects. It has never seen your variations, minutes or construction programmes. If you ask it anyway, it answers from general knowledge and, when in doubt, with whatever sounds plausible.

RAG solves this by giving the model a reference work. The basic idea comes from a 2020 publication: a pretrained language model is connected to a searchable vector index (in the paper, an index of Wikipedia). Microsoft calls the pattern an industry-standard approach in its architecture documentation, a way of letting language models work with data they did not know beforehand.

Think of a colleague who knows nothing by heart but keeps a very good index. She looks it up, lays the pages in front of you and summarises what they say.

1. Read in documents PDF, Office, emails, scans with OCR
2. Split into sections Each section covers one topic
3. Build an index Full text and meaning, with document and page AI prepares
4. Ask a question In plain language, as to a colleague
5. Find source passages The best-matching sections with an excerpt AI prepares
6. Answer with source Short text, every statement with a reference AI prepares

Text version of the graphic

• 1. Read in documents: PDF, Office, emails, scans with OCR
• 2. Split into sections: Each section covers one topic
• 3. Build an index: Full text and meaning, with document and page (the AI prepares)
• 4. Ask a question: In plain language, as to a colleague
• 5. Find source passages: The best-matching sections with an excerpt (the AI prepares)
• 6. Answer with source: Short text, every statement with a reference (the AI prepares)

What happens during set-up

The first three steps run once at the start and then for every new or changed file.

First the files are read in: PDF, Office documents, exported emails. Scans without a text layer need OCR first, otherwise there is nothing to search.

Then the programme splits each document into sections, according to Microsoft in such a way that each section belongs together in content. Sections that are too small often hold too little context to answer the question. Ones that are too large are expensive, can exceed the model’s limits and give poorer results. For minutes, splitting by agenda item suggests itself, and for a bill of quantities (BoQ) splitting by groups of items.

Finally each section goes into an index, together with document name, page and date. The search runs two ways at once. Full-text search finds exact terms. Search by meaning finds passages that fit the question even when different words are used. Microsoft names product codes, technical jargon, dates and names as cases for an exact match. For you these would be file references, item numbers or plan indices.

Two stages: source passages and answers

The search can be run in two stages, and you do not have to take both.

	Stage 1	Stage 2
Result	List of hits with excerpt, document and page	Short answer, every statement with a source
Who writes text	nobody, things are only found	a language model, from the hits
Checking	You read the passage yourself	You check the answer against the original
Risk	Hits are missing if the files are incomplete	in addition: the answer may summarise something wrongly

Stage 1 has a value of its own. It quickly shows you whether your files are searchable, and nobody has to trust a summary. Stage 2 builds on it. The model is instructed to answer only from the passages found and otherwise to report “not found”. That does not rule out errors, which is why the source reference matters.

A permissions filter belongs with it: each person sees only hits from documents they could also open in the file store.

A walk-through with an invented example

A question in the office: “What did we agree in the Musterstraße project about extending the construction period when supplies run short?” Full-text search finds all documents containing the phrase “extension of the construction period”. Search by meaning also finds a set of minutes that speaks only of “postponement of completion because of material shortage”.

In stage 1 you see three hits: building contract page 12, minutes of the regular site meeting page 3, variation 2 page 1, each with an excerpt and a date. You read the three passages yourself. In stage 2 a draft answer of two sentences stands above them with the same three source references. Before anything is taken into a letter, someone opens page 12 of the contract. All names and details in this example are invented.

How to start

A pilot on a completed project or on one area of folders makes sense, set at four to six weeks at the start as an assumption. A contact person on your side helps decide which folders go in and assesses the hits against the collected test questions. The pilot tells you how many questions the search finds, where hits are missing and whether extending it to more projects is worthwhile. Whether the knowledge base should run locally or with a provider in the EU is covered in Knowledge assistant on your own documents: local or EU cloud.

Why order comes before technology

Search finds only what is in the files and readable. An online survey by Fraunhofer IRB of March 2026 among architecture and engineering practices gives sober figures, which apply to the respondents and not to every practice. Mainly architecture practices (45 per cent) and structural engineering practices (29 per cent) took part. 65 per cent rate their project data as partly structured but in need of improvement. 16 per cent consider it so unstructured that it is hardly usable for AI. Only 8 per cent call it very well structured, and only 5 per cent of companies capture, archive and maintain their data systematically and to a high standard.

What that means in practice:

• Scans without a text layer appear in no search.
• If the old and new versions of a plan or contract lie side by side, the search can quote the old one. Date and path therefore belong in every hit.
• Questions with no hits show gaps. They should reach a named person who looks after the files, as a list.

Microsoft recommends collecting representative test documents and test questions before building and assessing each step separately. For a practice that means collecting everyday questions, noting where the answer actually is, and then measuring whether the search finds it.

What a knowledge base does not do

It does not find knowledge that is not written down anywhere. It does not replace checking against the original when something goes into a contract, a variation or a letter. And it needs rules on which folders are allowed in. Personnel files and tenant files generally do not belong in a general project search. Whether and how you may process the documents is for your data protection officer to clarify, and in an individual case your lawyer.

What you can do now

1. Collect ten questions that were asked in the office in recent weeks and note where the answer was.
2. In a completed project, open ten random PDFs and try to highlight one sentence in each. Where that is not possible, it is a scan without a text layer.
3. Name one person who is responsible for keeping the files in order.
4. Decide which folders never belong in a shared search.