WikiLLM Technology
From raw documents to compiled corporate memory — not RAG, but knowledge compilation.
The core idea: Compiled knowledge instead of RAG
Traditional RAG
Nothing accumulates. No learning effect.
WikiLLM (ROZUM)
Every source enriches the total knowledge.
Most AI systems use RAG: with every query, document fragments are searched and assembled. Knowledge is rebuilt from scratch every time — nothing accumulates.
WikiLLM works differently. Instead of re-deriving knowledge with every query, the system builds a persistent, interlinked wiki — a structured network of summaries, entities, concepts, and cross-references. New knowledge is compiled once and then kept current.
The idea is based on Andrej Karpathy's LLM Wiki concept: the LLM becomes a disciplined wiki maintainer. It reads sources, extracts key information, updates existing pages, and maintains cross-references — automatically and traceably.
Four-Layer Architecture
ROZUM extends the WikiLLM idea with enterprise governance.
Source & Provenance
Immutable raw data with SHA-256 hash, versions, ACL snapshots, and provenance tracking. SharePoint, file shares, ERP, PDF archives.
Compiled Knowledge
Summaries, entities, concepts, comparisons, contradictions — structured, linked, and with source citations. The persistent corporate wiki.
Retrieval & Reasoning
Hybrid search: tree reasoning, lexical and vector search, graph traversal, reranking, source verification, calibrated refusal on insufficient evidence.
Governance & Control
Identity management, policies, ACL synchronization, audit trail, retention, deletion — applied at every transition, not as an outer shell.
Permission-Aware Knowledge
The critical part that WikiLLM alone doesn't solve: when a knowledge page is created from three documents with different permissions, a user only sees it if they have access to all relevant sources. Permissions are inherited, not ignored.
Cost comparison: Local vs. Cloud
Open-source LLMs on your server eliminate ongoing token costs. The difference grows with every additional use.
| ROZUM Local | Cloud API | |
|---|---|---|
| Token-Kosten/Jahr | ~€200-500 (Strom) | €10.000-20.000 |
| Hardware (einmalig) | €4.000-8.000 | — |
| Scaling | Free | Linear increase |
| Vendor lock-in | None | High |
Supported Models
All models: MIT/Apache license, commercial use allowed, no vendor lock-in.