Wiki-LLM
The recommended mode for small to medium curated corpora. It compiles sources into persistent markdown pages, maintains AGENTS.md, index.md, and log.md, and answers from that memory before rereading raw sources.
Platform
Three clear ways to work knowledge: Wiki-LLM, Agentic RAG, and Graph-LLM.
EIGENVERTEX should no longer be read as a vague hybrid stack. The product is structured around WikiLLM for durable memory, Retrieval for evidence recall, and Graph-LLM for relational reasoning.
Three strategies
Each strategy has its role, its scale, and its promise
Agentic RAG
The recommended mode for large or exploratory corpora. It rereads the evidence layer, combines lexical and vector retrieval, and produces grounded synthesis without assuming a fully compiled wiki already exists.
Graph-LLM
The relational strategy. It connects concepts, methods, documents, and tensions to surface explainable paths, useful neighborhoods, and contradictions worth investigating.
Explicit workspace modes
A workspace is created either as WikiLLM or Retrieval. Ingestion, query, and maintenance behavior follow from that choice without ambiguity.
Immutable raw sources
PDFs, web pages, audio, video, notes, and imports remain the source-of-truth layer. The system reads them, indexes them, or compiles them, but does not rewrite them.
LLM-maintained wiki
WikiLLM materializes durable memory as source, topic, entity, concept, question, and analysis pages, with explicit lint and maintain operations.
Document retrieval engine
Retrieval mode prepares chunks, indexes, and evidence paths for questions that require broad recall, exploration, and precise citations.
Relational reasoning
Graph-LLM is not a decorative graph. It is the path that helps connect the strong points of a domain, expose tensions, and guide transverse views.
Chat, API, and outputs
Human teams and product applications can consume memory through chat, the native API, the OpenAI-compatible facade, and structured outputs.
Product outcome
What the system enables in practice
- Choose the right strategy for the size and nature of the corpus.
- Build durable memory when the topic requires capital accumulation.
- Reread raw evidence only when document recall is the right move.
- Plug product applications into a private, deployable domain intelligence layer.
Principles
An architecture built to stay legible
- 1 No implicit mixing between the WikiLLM stack and the Retrieval stack
- 2 The wiki is persistent memory, not a cosmetic cache
- 3 Retrieval rereads evidence instead of pretending to be compiled memory
- 4 The graph adds relational reasoning without becoming the main storage layer