Skip to content

LLM Wiki — an agent-maintained knowledge repository

A worked pattern that composes PageIndex + GraphIndex + Ontology into a single knowledge repository an agent compiles, queries, and contributes to. Runnable example: examples/knowledge_wiki/.

The idea

Classic RAG re-synthesises an answer from raw text on every query and throws the work away. An LLM wiki (after Andrej Karpathy's framing) flips that: the agent compiles sources into durable, cross-linked pages, and files its own knowledge back into the repository. The LLM does the editorial bookkeeping — cross-referencing, summarising, filing — that a human wiki maintainer would do by hand. Each conversation can leave the knowledge base better than it found it.

AI-Parrot already ships every piece needed to build this; the example wires them together without changing the framework.

How the pieces map

LLM-Wiki concept AI-Parrot subsystem Module
raw/ sources seed documents examples/knowledge_wiki/raw/
wiki/ pages PageIndex parrot.knowledge.pageindex
the knowledge graph GraphIndex parrot_tools.graphindex.GraphIndexToolkit
entity layer Ontology parrot.knowledge.ontology.OntologyRAGMixin
lint health report examples/knowledge_wiki/wiki.py::wiki_lint

PageIndex — the wiki pages

PageIndex stores a document as a lean tree: a JSON table-of-contents (titles, summaries, metadata) plus per-node markdown sidecars. It is writable — an agent authors durable pages with add_node, insert_content, tag_node, and update_node_content, and retrieves them with hybrid BM25 + LLM-walk search that returns citable section titles. See PageIndex.

GraphIndex — the knowledge graph the agent grows

GraphIndexToolkit exposes 19 tools, of which 7 are write tools: create_concept, create_node, link_nodes, unlink_nodes, attach_summary, tag_node, merge_nodes. They mutate the same in-memory graph and FAISS index the read tools use, so a concept the agent files mid-conversation is immediately searchable via find_node / relevance and joins the Louvain communities surfaced by list_communities. This is the mechanism that lets the LLM contribute its own knowledge rather than only consume a static corpus.

Ontology — the structured entity layer (optional)

OntologyRAGMixin adds tenant-scoped, authority-aware retrieval over a graph database. It is built for graceful degradation: with no tenant_manager it returns not_configured; with ArangoDB unreachable it returns vector_only — it never raises. The example therefore runs end-to-end without it, and lights it up when an ArangoDB-backed stack is supplied.

Wiring it together

The example keeps all glue inline in examples/knowledge_wiki/wiki.py:

  • build_pageindex_toolkit(...) / seed_wiki_from_raw(...) — compile sources into pages.
  • graph_seed_from_tree(...) — bridge the two indices by deriving graph DOCUMENT/SECTION nodes from the page tree.
  • build_graphindex_toolkit(...) — the ergonomic helper for a write-enabled GraphIndexToolkit (assembles the graph, embeds seed nodes, and hands the toolkit a consistent graph / faiss_index / node_map / node_id_list / assembler / embedder / nodes set).
  • WikiAgent(OntologyRAGMixin, BasicAgent) via build_wiki_agent(...) — one agent with both tool surfaces plus the entity layer.
  • wiki_lint(...) — surfaces orphan concepts and community structure.

The loop

  1. Ingest raw sources into PageIndex pages.
  2. Bridge the page tree into graph nodes/edges.
  3. Build a WikiAgent with both toolkits (+ Ontology).
  4. Query — grounded, cited answers.
  5. Contribute — the LLM files a new concept, cross-links it, attaches a summary, and adds a wiki page.
  6. Re-query — the new knowledge is searchable.
  7. Lint — report repo health.

Running

# Offline (no API key): real PageIndex + GraphIndex toolkits, deterministic.
python examples/knowledge_wiki/llm_wiki_agent.py --no-llm

# Full agentic demo (needs GOOGLE_API_KEY for ingest + agent).
export GOOGLE_API_KEY=...
python examples/knowledge_wiki/llm_wiki_agent.py

# Tests (offline, no API key / DB).
pytest examples/knowledge_wiki/ -v

Notes

  • The example's default HashingGraphEmbedder is offline and deterministic; swap in parrot.knowledge.graphindex.embed.GraphIndexEmbedder for production semantic similarity.
  • PageIndex auto-persists to examples/knowledge_wiki/store/ (git-ignored); the GraphIndex graph is in-memory in the demo.