oomol-lab/wiki-graph

distill any book down to its spine

What it solves

Wiki Graph addresses the challenge of managing large volumes of long-form text (such as books, meeting records, and internal documents) by converting them into a maintainable, structured knowledge base. Instead of forcing an LLM to re-read raw material from scratch for every query, it compiles the text into a durable format where entities and relations are extracted and traceable back to the original source.

How it works

The tool uses a CLI to process text into .wikg archives. It leverages WikiSpine to identify public entities aligned with Wikipedia/Wikidata QIDs, which provides a stable semantic foundation. An LLM is then used for disambiguation and filtering.

Key components include:

  • Knowledge Graph: Extracts entities and triples (subject --predicate--> object) to create a structured network of information.
  • Reading Graph: Breaks text into cognitive chunks and organizes them into reading chains to generate compressed, traceable summaries.
  • Evidence Tracking: Every extracted entity or relation is linked to specific chapters and sentences in the source text for verification.
  • URI System: Uses a stable URI-based addressing system (e.g., wikg://book.wikg/entity/Q8018) to access specific scopes or objects within the archive.

Who it’s for

It is designed for developers and AI agent creators who need to build knowledge-intensive workflows where information must be searchable, verifiable, and portable. It is particularly useful for those implementing the "LLM Wiki" concept for long-text knowledge production.

Highlights

  • Source-Traceable: Every piece of knowledge is backed by evidence pointing back to the original source text.
  • Public Entity Grounding: Uses Wikidata/Wikipedia alignment to ensure stability and avoid the hallucinations common in purely generative extraction.
  • Portable Knowledge: The .wikg format allows knowledge bases to be shared, backed up, and delivered like a book or course.
  • CLI-First Design: Provides a comprehensive command-line interface with a discoverable help network for easy integration into agent workflows.

Related

  • Project
  • Project
  • Project
  • Project
  • Project