BetriebsGehirn - RAG knowledge infrastructure case note

Short version

BetriebsGehirn treats AI answers not as a chatbot demo, but as knowledge infrastructure. The prototype asks how company knowledge must be prepared so that AI answers stay source-grounded, bounded, reviewable, and honest when evidence is missing.

Problem

A chatbot hides the hard questions

Many company AI ideas start with the same assumption: import internal documents, connect an LLM, and add a chat window. That can work as a demo, but it often hides search scope, source quality, permissions, evidence, and accountability.

Positioning

RAG is more than vector search

A useful answer depends on source scope, document structure, retrieval quality, evidence status, and failure behavior. If the evidence does not carry the answer, the system should return a non-answer or a review hint instead of plausible prose.

Architecture idea

First clarify search scope, source status, and evidence. Only then generate an answer.

Source scope before open-ended search across every available document.
Source status and document structure as part of answer quality.
Chunking along usable document units instead of only token boundaries.
Retrieval as candidate selection, not as a proof of truth.
Structured output and evidence status before publishing an answer.
Review gates for gaps, conflicts, and insufficient evidence.

Knowledge flow

The diagram shows architecture flow, not a product interface, document views, provider logos, or internal runtime details.

Sources synthetic or approved material
Structure parsing, metadata, chunks
Search scope scope, status, permissions
Retrieval candidate passages, not truth
Orchestration LLM bounded by the scope
Structured output claim, sources, status
Evidence supported, insufficient, not found
Answer or non-answer no invented replacement claim
Human review review gate and pilot decision

Grounded answer

Short, traceable, and limited to the source basis.

Non-answer

A deliberate result when the evidence is weak or missing.

Review gate

Human clarification for conflicts and pilot decisions.

Boundaries: local, synthetic, no customer-data operation, no guaranteed correctness.

Evidence status and non-answers

SUPPORTED

The answer can be grounded sufficiently in the defined source scope.

INSUFFICIENT_EVIDENCE

There are hints, but not enough evidence for a reliable answer.

NOT_FOUND

The defined search scope does not contain a usable basis for the question.

Conflict

Sources point to different states and need human clarification.

Boundaries / not claimed

BetriebsGehirn is described as a local architecture and development state. The limits are part of the case, not a footnote.

not a finished SME product
not a production-ready knowledge platform
not running customer data
not a legally approved or GDPR-safe solution
not enterprise-ready operation
no guarantee of correctness or hallucination-free answers
no finished cloud connectors, agents, or knowledge graph features

Next pilot slice

one concrete knowledge domain
a defined search scope
synthetic or explicitly approved test data
clear user questions
success criteria for grounded answers
stop criteria for weak evidence
visible review gates for conflicts and non-answers

Consulting bridge

AI Architecture Review before building

If an AI pilot should become more than a demo, the first useful step is an architecture check: search scope, data sources, output shape, evidence, review gates, risks, and realistic next steps.

AI Architecture Review 30/60/90 pilot planning