Evidence Engine
RAIVA is the evidence-based AI platform for enterprise knowledge. It connects retrieval, analytical methods and language models so that claims can be checked against their sources.
Documented architecture. Not proof of production approval.
Architecture
Document preparation provides the foundation. Semantic, lexical and graph retrieval complement each other in parallel. Role profiles coordinate research and can initiate further searches. Generation and evidence verification remain distinct tasks.
Documented architecture. Not proof of production approval.
Model limitations
Language generation can produce plausible but unsupported claims. Missing context, unsuitable matches, conflicting document versions and faulty conclusions are distinct sources of error.
ChatGPT and Claude can themselves use research tools and sources. The specific configuration and workflow matter. Citations alone establish neither a claim nor the completeness of research. Sampling affects the output, not its truth.
Verification chain
Choose a case and a component. Switching cases preserves your component selection. This shows functional responsibilities, not a fixed technical execution order.
Illustrative example · fictional documents
Fictional Lindenhof documents. Illustrative verification, no live query.
Supported within the scope shown; other documents have not been exhaustively reviewed.
Search again when evidence is missing
v1 · Page 8 · § 12
The lease is concluded for a fixed term ending on 31 December 2030.
The addendum is signed by the tenant only. Whether it is effective is unresolved; the applicable lease end remains open.
Search again when evidence is missing
v1 · Page 8 · § 12
The lease is concluded for a fixed term ending on 31 December 2030.
v2 · Page 2 · § 1
The term is extended until 31 December 2033. The tenant may terminate once, effective 31 December 2031.
Uses the end dates in the rent roll. If the 2nd addendum is effective, unit A runs until 2033 with a break option in 2031 and the WALT must be recalculated. Correct arithmetic confirms neither the completeness nor the accuracy of the inputs.
Search again when evidence is missing
v1 · As of 30 Jun 2026 · Rows 1–2
Unit A: €229,200 p.a., ends 31 Dec 2030. Unit B: €110,400 p.a., ends 30 Jun 2029.
Analytical AI
Deterministic means that a defined calculation or rule can be reproduced with the same inputs. It does not apply to all model decisions or complete answers.
(€229,200 × 4.5 yrs + €110,400 × 3.0 yrs) / €339,600 = 4.01 yrs
Rent roll · v1 · As of 30 Jun 2026 · Rows 1–2Even a correctly calculated WALT is wrong for the decision if the rent roll does not reflect the 2nd addendum.
Applications
Set diverging end dates, rents and break options side by side and leave unresolved effectiveness open.
Match every tenancy in the rent roll to a lease and flag gaps with their source.
Show rents, end dates, reporting date and formula alongside the result.
The source shown supports the claim within the stated scope.
Sources disagree; clarification is required.
No adequate source was found.
Do not claim a successfully completed check.
Individual citations do not replace exhaustive research.
Common questions
The workflow makes source binding, checks and open limits explicit. AI chat can also use tools; configuration matters.
It helps find documented relationships and connected sources. A relationship alone does not establish substantive validity.
It prioritizes matches by relevance. It does not determine whether a claim is true.
It influences the selection of text continuations during generation. It is not fact-checking.
No. Defined calculations may be reproducible; this does not guarantee correct model decisions or complete answers.
The gap remains visible. Further research or human clarification is needed; unsupported claims are not presented as confirmed.
Next step
Discuss your documents, question and required review scope.
Product principles, not a production or legal assurance.
Documented architecture. Not proof of production approval.
