Official letterDistrict office · 2 min ago
ReviewCase 03 · Document automation
One scan —
and the document does the rest.
Paper mail is read, classified, archived and turned into clear tasks. Later, a normal question is enough to find the right case.
Self-hosted OCRRule-based validationCross-language search
1×scan once, then find it anywhere in the workflow
2search methods: full-text and semantic
Reviewwhen required fields are missing
The system is in production; the documents, names and figures shown here are anonymised sample data.
PyPython
AIVision OCR
PGPostgreSQL
Vpgvector
GGoogle Drive
TGoogle Tasks
TGTelegram
DDocker
Thursday · 23 JulyIncoming documents
⌕ Search everything
Processed today14+3 since 08:00
Automatic1286% without intervention
To review2with rule explanation
Insurance addendumInsurer · 18 min ago
ArchivedSupplier invoiceSupplier · 09:42
DoneSDocument assistant
ASK INSTEAD OF SEARCHING
When is the reply to the district office due?
From 2 sources24 July 2026
The deadline is on page 2 of letter #0412.
Open original →The starting point
The paper was filed.
Its consequences were not.
A tidy folder still does not answer who must act, which deadline is running or where the eventual reply belongs.
Read and name
Document type, sender and case had to be identified and entered manually for every scan.
Transfer deadlines
A deadline in a letter became a task only after someone consciously copied it into another system.
Reconstruct cases
Incoming letter, reply and follow-up correspondence were stored together — but not automatically as one coherent case.
From intake to action
Every step creates
an auditable state.
The language model reads and suggests. Rules decide whether processing may continue automatically.
Text, type and sender are extracted.
Existing text layers are used; otherwise the system renders pages and runs vision OCR.
Rules decide what runs automatically and what needs review.
Specific checks, not a vague confidence score, determine the next step.
One question is enough as a search field.
Full-text and semantic search are combined — even across languages.
Ingest · OCR and classification
Page 2 / 3300 dpi
Reference 41-ABReply due by 24 Jul 2026
ERGEBNISStructured fields
OCR complete- Document type
- Official letter
- Sender
- District office
- Reference number
- 41-AB
- Reply deadline
- 24.07.2026
Review queue · Uncertain cases
✓ ✓ ! !
SYSTEM SUGGESTIONOfficial letter · District office
Review requiredDocument type is clearType matches sender and register
BESTANDENDeadline found in the originalPage 2 · Text position saved
BESTANDENReference number already existsSimilar case #0398 recognised
REVIEWReply reference is unclear2 open tasks may match
REVIEW
Telegram · Cross-language search
SDocument assistant
Когда истекает срок ответа?Question in Russian
2 sources · 0.8 sReply deadline: 24 July 2026Source: official letter #0412 · Page 2
Open originalCase #0412
System views
More than filing.
One complete case from start to finish.
The views show intake, review, tasks and search without real names or document contents.
A complete end-to-end process
From letter to task.
The reply returns to the same case.
An outgoing reply scan can close exactly the matching open task — not merely a similar one.
ScanPDF intake
OCRText & fields
ValidatorAutomatic or review
ArchiveReadable structure
TaskDeadline & status
SearchFull text + vector
Automation with clear limits
The model reads.
Rules decide.
That keeps the result repeatable, reviewable and correctable.
No gut-feeling approvals.
Specific rules validate each condition. If something does not match, the system explains why and routes it to a person.
Reply and intake stay connected.
The system closes only an unambiguous matching task. Ambiguous matches are never decided silently.