PROJECT INSIGHT ·
AI That Genuinely Saves Me Work
More than ten years of scanned mail — and today my archive answers questions. Locally, with verifiable sources and clear boundaries.
This series began with a question about the deductible on my home contents insurance. The answer was in a letter I had scanned years ago and never opened again. Today I simply type questions like that and receive the answer together with its evidence. Between those two moments came text recognition, an index that does not forget prematurely, search by words and by meaning, a local language model and a research assistant with a fixed toolbox.
*AI-generated illustration: A closed local path leads from the scanned letter to the answer and visibly back to its source.*
It Is Finished for My Purpose
Software can always be extended. There are new file formats, new models and new ideas. That does not make Operus Search unfinished to me. It solves the problem for which I built it: I can reach the information in my mail without knowing where the letter is stored.
My letters do not leave my network. And if the language model is unavailable, search remains usable.
Operus Search is finished for me because it helps me today — not because there are no ideas left.
Every Question Takes Its Own Path
Looking back, this may be the most important lesson from the series. Questions to my archive all sound similarly straightforward. Technically, they require completely different paths — and Operus Search does not force them all through the same language model.
| Question | Path |
|---|---|
| Where does it say “property tax”? | full-text search, entirely without AI |
| What is my deductible? | full text and meaning, answer from the letter, source |
| How far in advance must I cancel? | the same — the period is quoted, not calculated |
| Who did I spend the most with in 2024? | deterministic database query, program code calculates |
| How much more expensive was the invoice than the quotation? | research assistant with comparison tool and verification |
| When do my statutory rights expire? | receipt plus verified rule, assumptions remain visible |
The language model has an important but limited role in this table. It is most valuable at the transitions: between my everyday question and the application’s precise functions, and between several passages and an understandable answer. Calculation, counting and determining periods are left to code.
Three Questions I Deliberately Do Not Get Answered
What Operus Search does not do says at least as much as its answers. In the thorough mode from Part 8, it looks like this:
“Delete all invoices from 2022.”
Operus Search refuses. It can read my letters but cannot modify or delete anything — not because it is polite, but because no such tool exists.
“What is the weather in Berlin today?”
There is no forecast, only a notice that no internet access exists. The architectural decision from Part 2 becomes very concrete here.
“How High Is Mount Everest?”
A general language model would answer immediately. Operus Search says that it found no relevant documents. That is true — my mail says nothing about Mount Everest. This is precisely why I can trust the other answers. An application that stays silent when evidence is missing does not invent anything when the subject is my insurance either.
What Could Come Next
Because ingestion, indexing, search and answers are cleanly separated, new sources can be connected without reinventing the system.
The email inbox is an obvious candidate. Many invoices no longer arrive by post but as PDF attachments. A small additional process could detect new attachments and place them in a watched folder. From that point, the familiar path takes over:
new PDF attachment
↓
watched folder
↓
watcher and job
↓
extract or recognise text
↓
immediately searchable and ready for questionsThe index, search and AI do not have to be rebuilt. The new source ends at an existing entry point.
The place where Operus Search runs may change too. I plan to move it into a Docker container on an NVIDIA DGX Spark, where larger language models can also run quickly. The principle remains unchanged: that device will also stand in my home.
Why Operus Search Matters So Much to Me
I have built many technically interesting projects. Yet I measure the value of Operus Search by something else: it helps me with my own data. An archive that grew for more than ten years has become a tool that makes important information accessible in a short time.
An impressive demo can show what a model can do in principle. Operus Search shows me in everyday life what my insurance policy says, how an invoice differs from a quotation and what evidence supports an answer.
The files remain my files. The folders remain my folders. But access to what they contain has changed fundamentally.
Good AI does not replace the dependable parts of an application. It connects them so that a person can reach the goal faster.
My archive can now do more than be searched. I can ask it questions — and trust the answers because I can follow their path all the way back to the scanned letter.
End of the series — and the beginning for every additional source that can use the same dependable path.
Back to the beginningI have been scanning my mail for more than ten years — how it all began with a question about an old insurance letter. Part 1
