oomol-lab/wiki-graph

distill any book down to its spine

What it solves

LLMs often struggle with large source materials, requiring them to re-read raw data from scratch for every query. Wiki Graph solves this by compiling long text into a maintainable, structured knowledge base (the "LLM Wiki" concept) that preserves entities, relations, and direct evidence back to the original source.

How it works

The tool uses a CLI to ingest text (from PDFs, EPUBs, web pages, etc.) and stores it in a custom .wikg archive format. It employs an LLM to extract entities—aligned with public datasets like Wikipedia/Wikidata via WikiSpine—and relations (triples) from the text. It also builds a "Reading Graph" by chunking text and connecting conceptually relevant segments, which allows for generating summaries that remain traceable to the original source sentences.

Who it’s for

  • AI Developers building agentic workflows that require stable, verifiable knowledge retrieval.
  • Knowledge Workers looking to transform long documents into searchable, structured, and portable knowledge bases.
  • Researchers needing to trace specific claims or entities back to their exact source chapters and sentences.

Highlights

  • Traceable Knowledge Graphs: Generates structured networks of entities and relations that include direct evidence links to the source text.
  • Portable .wikg Format: Uses a specialized archive format to store source text, chapter trees, and knowledge graphs in a single, shareable file.
  • Public Entity Alignment: Uses Wikipedia/Wikidata QIDs to provide stable semantic boundaries for extracted entities.
  • Context Packing: Includes a pack command to turn specific chunks or entities into portable text for use in LLM prompts.
  • URI-based Navigation: Employs a unique URI system to navigate and query specific scopes, chapters, or entities within an archive.