Why two searches instead of one?

Because each is good at something else. Keyword search hits “§ 7” or “batch 2024-08” exactly but misses paraphrases. Semantic search understands paraphrases but tends to confuse similar numbers. Together, with reranking, they find both.

What does the knowledge graph do that search does not?

Search finds passages. The knowledge graph records individual facts, such as a delivery period with its validity, each with a source. It does not reorder search hits. When two facts disagree, a responsible person decides.

How quickly do changes reach the answers?

With the next sync of the source. Only changed files are recomputed. When only rights change, nothing is recomputed. A deletion removes passages, vectors and derived facts.

Does souveraen.ai train a model on our data?

Not in levels 0 to 2. Learning happens in the index, knowledge graph and glossary, not in the language model. A custom adapter (level 3) is off by default and needs its own decision.

How do we check quality on our side?

With your own test questions and the sources you expect. Every change to search is measured against them. You see the result in the learning report before a change goes live.

Learning knowledge

From thousands of passages, the eight that matter.

Before anyone asks, souveraen.ai has read your repositories: text, scanned pages and tables, split along headings, with version and access rights on every passage. Each question then runs through two searches and a second scoring pass. Facts from the knowledge graph come along with their own source.

Start for free

How a document is read

Text, scans and tables
Keyword and semantic search
Rights on every passage

Search trace

What delivery period applies to partial deliveries from Müller GmbH?

Purchasing workspace · 9,522 passages

1
Rights and validity
Only passages you may open, in their current version
6,140
2
Keyword search
BM25 over “Müller”, “delivery period”, “partial delivery”
50
3
Semantic search
Vectors (BGE-M3) also find “delivery in instalments”
50
4
Fused
Both lists merged by rank (RRF)
40
5
Reranked
A second model reads question and passage together
8

Sources for the answer

1Framework agreement Müller GmbHv3 · p. 4 · § 5.2
2Purchasing T&C 2026§ 7 (3)
3Supplier handbookch. 4.1 · table 2

Fact from the knowledge graph

Müller GmbH · delivery period · 14 calendar days · valid since 1 Mar 2026

Illustrative run with the default search settings.

Ingestion

What happens to a document before anyone asks.

A framework agreement as a scanned PDF, with a price table on page 6. Six steps later it is searchable in passages that know who may read them.

Framework_agreement_Mueller.pdfWaiting

SharePoint

Contracts folder

Detect changes

Changes only

Read, scans included

Up to 1,000 pages

Split along headings

≈ 600 tokens

Label

Rights per passage

Index twice

2 indexes

Record facts

With source

Keyword index

BM25

Vector index

BGE-M3

Knowledge graph

Facts

Click a step to see what happens there.

Searchable, with version and access rights on every passage

Example run with a single document.

Five building blocks behind every answer.

Each one solves a problem where simple AI search falls short.

Split the way the document is

Fixed-size blocks cut through sentences and tables. souveraen.ai splits at headings and paragraphs, keeps tables whole and overlaps neighbouring passages so nothing is lost at the cut.

Literal and by meaning at once

Keyword search finds item numbers, clauses and names exactly. Semantic search finds the contract even when it says “delivery in instalments”. Both result lists are merged by rank.

A second, closer look

A dedicated scoring model reads the top 40 hits again, this time question and passage together. What only sounds similar drops down. The answer gets the eight passages that actually answer the question.

Facts with origin and validity

Beside search, a knowledge graph records what is known about suppliers, machines and contracts, each fact with its source and validity period. It does not reorder hits; it adds sourced facts. It never settles contradictions by itself.

Filtered before the search runs

Rights, validity and classification are checked on every question, in both indexes. What someone may not open does not appear for them: no title, no preview, no hit count.

Framework_Mueller_v3.pdf

§ 5 DeliveryPassage 1
5.1 Place of delivery and transfer of riskPassage 1
80 tokens overlap
5.2 Delivery period: 14 calendar days from confirmationPassage 2
5.3 Partial deliveries per line itemPassage 2
Table 2: surcharges by quantityPassage 3

Table kept whole

„partial delivery Müller“

Keyword

T&C § 7: “partial deliveries …”

Müller GmbH, order 4471

Supplier list: Müller

Meaning

Contract 5.3: “delivery in instalments”

Handbook 4.1: staggered delivery

T&C § 7: “partial deliveries …”

Fused

1. T&C § 7
2. Contract 5.3
3. Handbook 4.1
4. Order 4471

Rank beforeScore

3
Delivery period 14 calendar days from confirmation
Contract · p. 4
96
1
Partial deliveries count per line item
T&C · § 7
91
5
Staggered delivery on call-off
Handbook · 4.1
74
2
Supplier list with contacts
List · row 12
18

Müller GmbH

Delivery period: 14 days

Contract v3 · p. 4

supplies: steel sheet S235

Order 4471

Contact: J. Krause

Supplier list

Delivery period: 21 days

Price list 2023

Contradiction sent for review

What terms does Müller GmbH have?

L. Berger

Purchasing

8 sources

including contract and price list

T. Hahn

Sales

3 sources

approved supplier list only

Same question, same search. Different rights, different sources.

Framework_Mueller_v3.pdf

§ 5 DeliveryPassage 1
5.1 Place of delivery and transfer of riskPassage 1
80 tokens overlap
5.2 Delivery period: 14 calendar days from confirmationPassage 2
5.3 Partial deliveries per line itemPassage 2
Table 2: surcharges by quantityPassage 3

Table kept whole

Split the way the document is

Fixed-size blocks cut through sentences and tables. souveraen.ai splits at headings and paragraphs, keeps tables whole and overlaps neighbouring passages so nothing is lost at the cut.

„partial delivery Müller“

Keyword

T&C § 7: “partial deliveries …”

Müller GmbH, order 4471

Supplier list: Müller

Meaning

Contract 5.3: “delivery in instalments”

Handbook 4.1: staggered delivery

T&C § 7: “partial deliveries …”

Fused

1. T&C § 7
2. Contract 5.3
3. Handbook 4.1
4. Order 4471

Literal and by meaning at once

Keyword search finds item numbers, clauses and names exactly. Semantic search finds the contract even when it says “delivery in instalments”. Both result lists are merged by rank.

Rank beforeScore

3
Delivery period 14 calendar days from confirmation
Contract · p. 4
96
1
Partial deliveries count per line item
T&C · § 7
91
5
Staggered delivery on call-off
Handbook · 4.1
74
2
Supplier list with contacts
List · row 12
18

A second, closer look

Müller GmbH

Delivery period: 14 days

Contract v3 · p. 4

supplies: steel sheet S235

Order 4471

Contact: J. Krause

Supplier list

Delivery period: 21 days

Price list 2023

Contradiction sent for review

Facts with origin and validity

What terms does Müller GmbH have?

L. Berger

Purchasing

8 sources

including contract and price list

T. Hahn

Sales

3 sources

approved supplier list only

Same question, same search. Different rights, different sources.

Filtered before the search runs

Rights, validity and classification are checked on every question, in both indexes. What someone may not open does not appear for them: no title, no preview, no hit count.

Simple AI search and this pipeline.

Six differences you notice on a single contract.

Not a promise, a way of working: splitting, two searches, reranking, graph and rights.

What matters

Splitting

Rights

Knowledge graph

Quality

Model

souveraen.ai

Headings, overlap, tables kept whole

Keyword and meaning, then a second score

On every question, in both indexes

Facts with a source, without reordering hits

104 test questions, 97.34% source fidelity, measured internally

Unchanged. Learning sits in index and graph

Simple AI search

Fixed character blocks, often mid-sentence

Usually one vector search

Often checked only at ingest

No separate fact store

Unknown until someone objects

Often retrained on the documents

Internal measurements on NVIDIA DGX Spark, as of September 2026. Not contractual figures.

Questions

What IT and the business want to know about search.

Including the limits. Search is only as good as the documents it finds.

Rather see it on your documents?

Bring the files other tools fail on: scans, large tables, old versions. Phone +49 3744 365 2202.

Book a call

PDF, Word, Excel, PowerPoint, email, HTML, text and images such as PNG, JPG or TIFF. Scanned pages go through OCR. A file may be up to 50 MB and 1,000 pages.

Try it yourself

Give it the documents nobody can find.

Connect a repository and ask the questions you call colleagues about today. Free for up to 3 users, no credit card.

Start for free Book a call

Create account1
Connect a repository2
Ask the first question3

From thousands of passages, the eight that matter.

What happens to a document before anyone asks.

Five building blocks behind every answer.

Simple AI search and this pipeline.

Six differences you notice on a single contract.

What matters

souveraen.ai

Simple AI search

What IT and the business want to know about search.

Rather see it on your documents?

Which files can souveraen.ai read?

Why two searches instead of one?

What does the knowledge graph do that search does not?

How quickly do changes reach the answers?

Does souveraen.ai train a model on our data?

How do we check quality on our side?

Give it the documents nobody can find.