Finding documents by their content

The task

The agency keeps its projects in folders on its own server. In our conversation it named three problems: the search mostly found only file names, its index kept failing, and new staff didn’t know the numbering scheme of the project folders.

What was needed was a search that understands what you mean, without anyone moving or renaming files or transferring them into a new program.

An example

The question asks about ‘fast internet’; the email says ‘fibre optic’. So the keyword search finds nothing, while the search by meaning finds the email.

How it works

1. The search reads the existing folders directly. Nothing is moved or renamed.
2. It extracts the text from PDF, Word and text files. Text recognition reads scanned pages.
3. Every query searches in two ways: for the exact words and for the meaning. Exact numbers, such as an invoice number, come first.
4. Each document appears as one result; versions such as ‘_final’ or ‘(1)’ are grouped together.

Status and limits

Prototype

Developed in two days as advance work for a proposal and shown in a client meeting on 6 October 2026. No pilot has been commissioned so far.

Tested

On a deliberately messy test server with 487 invented files, the right document came first in 22 of 25 test searches; keyword search alone managed 17. I created the files and the test searches myself.

Data

The search index is a file next to the search. For text recognition and meaning, text extracts and scanned pages go to AI models at Mittwald, a hosting provider in Germany. Whether the search would run on an office computer or at Mittwald in day-to-day use is still open.

Outages

If the AI service can’t be reached, the keyword search keeps working.