Knowledge Graph & LLM Wiki¶

Classic RAG re-synthesises an answer from raw text on every query and throws the work away. This chapter covers the three subsystems AI-Parrot ships to do the opposite — compile knowledge once, keep it structured, and let agents contribute back to it.

The three layers¶

Layer	Unit of work	The question it answers
PageIndex	One document	Where inside this document is the answer?
GraphIndex	A whole corpus	How do these things relate to each other?
LLM Wiki	A durable page set	What have we already established?

They compose. The LLM Wiki is built on GraphIndex pages and edges; GraphIndex can hand hierarchical documents to PageIndex so a SECTION node resolves to real prose instead of a chunk.

PageIndex — tree-based, vectorless retrieval¶

Builds a hierarchical semantic tree from a PDF or Markdown document using LLM reasoning, then navigates that structure to find relevant sections. No vector database required, and it works with any provider AI-Parrot supports.

→ PageIndex reference

GraphIndex — typed graph over code, docs and skills¶

A 6-stage pipeline (extract → embed → assemble → resolve → persist → analyze) that emits UniversalNode / UniversalEdge across domains, infers cross-domain edges by similarity, and produces an interactive graph.html map plus a GRAPH_REPORT.md naming god nodes, communities and knowledge gaps. The code pass is deterministic and LLM-free.

→ GraphIndex reference

LLM Wiki — an agent-maintained knowledge repository¶

A machine-first knowledge graph of pages and typed edges the agent compiles, queries, and files its own knowledge back into — "the agent forgets, the graph does not". wikitoolkit build creates it, wikitoolkit query reads it, wikitoolkit remember writes to it, and namespaces let several corpora (a codebase, an Obsidian vault, a Jira backlog) federate behind one query.

→ LLM Wiki overview · Complete guide

Decision matrix¶

Use case	Reach for
Retrieve from one long PDF without a vector DB	PageIndex
Map how a codebase hangs together	GraphIndex CLI (`parrot-graphindex`)
Give an agent memory that survives restarts	`GraphMemoryMixin` — Persistent graph memory
Let a coding assistant consult the repo before grepping	WikiToolkit as Claude Code infrastructure
Query Jira tickets as a knowledge base	Jira issues namespace