xarical/wof-wiki

Project folders

• README.md - You are here!
• source/ - Original Mediawiki xml file, as well as a cleaned version. The cleaned version is the one used to generate documents/ and chromadb/.
• documents/ - Converted into markdown files.
• chromadb/ - Converted into a chromadb to eliminate the markdown conversion and embedding steps to use the dataset for RAG. It is highly recommended that the embedding model used for retrieval is the same as the one used to generate the embeddings. The embeddings in the current version of the chromadb were generated with Google's text-embedding-004.

Notes:

• Mediawiki-to-GFM is used to convert the Mediawiki xml file to GitHub Flavored Markdown (GFM)
  • The cleaned version of the MediaWiki xml file removes [[]], variations of '''', and a wikitable (Fatespeaker's) that was causing issues with Mediawiki-to-GFM.
  • The "Category:" files were removed, as they do not provide any meaningful information.
  • There are approximately 580 markdown files, corresponding to 580 wiki pages.
• unstructured's chunk-by-title method was used to chunk the documents for embedding.
  • The markdown files were run through a Python script to add a new line at the top of each document with the text "## Overview", which is markdown for a section header. It is included because the chunk-by-title method was ignoring the first overview section of the documents.
• Questions that works very well for testing:
  • (Easy) "Who is <character name>?"
  • (Easy) "Who is <side character name> (<piece of relevant info to narrow down the search>)?"
  • (Medium) "What is the personality of <main character>?"
  • (Hard) "What are the motivations of <side character>?"
    • Prone to hallucination in simple RAG systems due to irrelevant documents
  • (Hard) "Who is <obscure character>?"
    • Prone to hallucination in simple RAG systems due to irrelevant documents
  • (Hard) "Who is <character who shares a name with another character> (from <specific book name>)
    • Prone to hallucination in simple RAG systems due to shared name; somewhat mitigated by larger chunk sizes to improve context understanding

Credits:

Original Mediawiki xml file is sourced from the The Wings of Fire Wiki, which is licensed under CC-BY-SA 3.0.
This dataset is a fan-made project, and is not associated with Wings of Fire or the Wings of Fire Wiki.