Skip to content

Labor · Knowledge

Turning Documents into Usable Knowledge

We process incoming documents and existing document holdings without losing track of their origin. This makes content findable, verifiable, and usable for further work.

Practical build 01Structural diagram · processing chain

Capturing and filing documents in a structured way

Invoices, delivery notes, and contracts arrive in different formats and quality levels. Without a shared processing step, their content stays unstructured and hard to find.

Test stand

Make the capture worse and see which fields the rules catch.

Capture

Delivery note

Delivery note no.LS-204817

Delivery date14.03.2025

Hausmann Baustoffe GmbHAm Sonnenweg 1260311
ItemArticleQty
1Cement, 25 kg bag12
2Paving stone, grey40
3Kerb stone, added by hand6

Total quantity58

Received byK. Brandt

Filing

  1. Delivery note no.LS-204817LS-204817filed
  2. Delivery date14.03.202514.03.2025filed
  3. SupplierHausmann Baustoffe GmbHHausmann Baustoffe GmbHnone, free textno rule
  4. Delivery addressAm Somnenweg 12Am Sonnenweg 12none, free textno rule
  5. Postcode6031160311filed
  6. Qty item 11212filed
  7. Qty item 24040filed
  8. Qty item 366filed
  9. Total quantity5858filed
  10. Received byK. 8randtK. Brandtnone, free textno rule

Try this: with the phone photo every rule passes, yet two fields are wrong, because no rule checks free text. With the fax, the sum rule catches the total quantity. Switch that rule off and the error goes into filing unnoticed.

Filed directly

7

rule passed, no human look needed

For review

0

a rule flagged the field

Wrong and unnoticed

2

only possible where no rule checks

Sample document. The reading errors are a visible model of typical character mix-ups, not a measurement of any text recognition. The build makes no claim about recognition quality or throughput.

How the setup works Choose how well the sample delivery note was captured. Fixed rules check every field, and only what passes goes into filing without a human look. Switch rules off and compare with the original to see what slips through unnoticed.
Scope and limitations

Not part of this demo: The reading errors are a model of typical character mix-ups. The build makes no claim about format coverage, recognition quality, or throughput.

The build shows a documented processing chain, not a specific installation.

Practical build 02Product concept · Sovereign Knowledge Engine

Making confidential holdings searchable and verifiable

Contracts, reports, and meeting minutes accumulate scattered across years. Anyone looking for a specific statement has to search through folders or rely on what one particular person happens to remember.

Test stand

Ask the collection a question.

Try

Collection

Answer

The answer is always a passage from an approved document, quoted word for word.

Document

Maintenance framework agreement

§ 3 The contractor takes on maintenance and minor repairs to the client's buildings.

§ 7 The warranty for completed work is 24 months from acceptance. Defects must be reported in writing within 14 days of discovery.

§ 11 The agreement runs for an indefinite period. It can be terminated with three months' notice to the end of a quarter.

Approved documents

4/5

only these may answer

Matching passages

0

in approved documents

Matches, but locked

0

stays out

Sample collection of made-up documents. The search compares words in the browser, no language model is involved. The build makes no claim about match quality, collection size, or response time.

How the setup works Ask a question of a small sample collection. The answer is the passage that was found, quoted word for word, with document and section. Approve the draft or lock a contract and see that only approved documents answer.
Scope and limitations

Not part of this demo: The build uses a simple word search, no language model is involved. It shows the principle, not match quality, corpus size, response time, or a specific installation.

The stations follow the knowledge lifecycle of the BKS Sovereign Knowledge Engine.

Practical build 03In-house development · closed trial

Searching existing folder structures without restructuring them

Existing file shares hold folder structures that have grown over years. Their paths and structure cannot be changed without risk, yet their content still needs to become findable.

Test stand

The folders are already indexed. Change a file, and the index still searches the old version.

Folders, unchanged

  • Projects
  • 2019_Hall-North
  • Quote_v3.docxchanged, not yet indexed
  • Acceptance
  • Acceptance_report_final_FINAL.pdfchanged, not yet indexed
  • 2021_Daycare-Sonnenweg
  • Framework_agreement.pdfchanged, not yet indexed
  • Admin
  • Contracts
  • Lift_maintenance_2020.pdfchanged, not yet indexed
  • Scans
  • Scan_0047.pdfchanged, not yet indexed
  • Old
  • Copy of Quote_v3 (2).docxchanged, not yet indexed

Index

First run: 6 read.

Search

FilterNo accepted suggestions yet.

    Suggestions

    Machine suggestions are not matches. Only accepted suggestions become filters.

    • Type QuoteQuote_v3.docx
    • Client Logistik NordQuote_v3.docx
    • Type ReportAcceptance_report_final_FINAL.pdf
    • Client Logistik NordAcceptance_report_final_FINAL.pdf
    • Type ContractFramework_agreement.pdf
    • Client Sonnenweg daycareFramework_agreement.pdf
    • Type ContractLift_maintenance_2020.pdf
    • Type InvoiceScan_0047.pdf
    • Client Logistik NordScan_0047.pdf
    • Type QuoteCopy of Quote_v3 (2).docx

    Files moved

    0

    paths stay as they are

    Read in the last run

    6

    of 6, the rest was unchanged

    Open suggestions

    10

    waiting for a human

    Made-up sample folder structure. Search and index run in the browser. The suggestions are stored in the content and not generated by a model, one of them is wrong on purpose. The build makes no claim about the scope, format coverage, or match quality of k2a.

    How the setup works Search a grown sample folder structure. Change a file: the index only knows the new content after the next run, and that run reads only what changed. Suggestions become filters only once you accept them.
    Scope and limitations

    Not part of this demo: k2a is currently being tested against our own document holdings. Scope, format coverage, and match quality are still under active development.

    The build shows a feature of k2a, our in-house, not publicly available tool for knowledge access.

    Three levels of maturity, clearly separated
    The first build shows a documented processing chain, the second our product concept for confidential knowledge holdings. The third shows k2a, an in-house development that we are currently testing exclusively against our own document holdings.

    Back to the Labor overview →

    Which document set would make a sensible first pilot?

    A clearly bounded document set is enough to test search paths, source references, and required approvals in practice.

    Schedule a conversation