Skip to content

menkhus/wiki_knowledge_graph

Repository files navigation

wiki_knowledge_graph

Thesis

Wikipedia is already the target pattern, not just a source to mine: it is simultaneously a knowledge graph (entities, cross-links, categories) and a human-readable, book-structured document (headers, sections, a references list, citations). The dual nature is the point — proof the pattern holds at a scale nothing else reaches, maintained by humans, for decades.

The piece that's actually broken is the citation itself, not the article. A reference today points to something that's frequently static, stale, unstructured, or link-rotted. Post-AI document publishing means the cited document — a paper, a manual, a spec, another wiki page — should carry the same apparatus a book already gives a human reader: frontmatter (title, author, date), a table of contents, a back-matter index, and citations. That structure is reference-supporting for a human and reference-grounding for an AI, at the same time, through the same mechanism.

An AI built this way avoids relying on trained-in recall of a document's content — recall that is real, measurable, and legally contested (see notes/reference.md §0) — and instead fetches and cites the current, public, specific source at query time, the same way a well-prepared human researcher would.

The one-sentence version: the AI should talk the way a well-prepared human does — citing a specific, current source it just looked up, not asserting from memory.

What this project is, concretely

A gap in an otherwise well-developed idea space (see notes/reference.md §2 for the full directory index): nothing in this workspace yet connects to Wikidata/Wikipedia's actual APIs, or builds the document-publishing standard described above. auto_knowledge_graph already has the reusable schema, provenance model, and embedding-based semantic index this project would build on rather than duplicate.

Notes

  • notes/reference.md — the processed reference document: directory index, dated placement in the 2026 idea-space (Karpathy's LLM Wiki vs. RAG vs. knowledge graphs), research on structure-guided retrieval vs. embeddings, the Intel SDM worked example, and the copyright/recall analysis behind the thesis above.
  • notes/SESSION_2026-07-14_consolidation.md — the raw session log this was distilled from.

About

Post-AI document publishing: grounding AI answers in front-matter/back-matter/citation structure instead of RAG chunks or trained recall

Topics

Resources

License

Code of conduct

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages