Wiki-RAG
An experimental retrieval-augmented generation system that ingests any MediaWiki site through its API and answers questions about that content through an OpenAI-compatible API, built as a Moodle Research project.
Wiki-RAG pulls structured knowledge from a MediaWiki site's own API, indexes it for retrieval, and combines that retrieved content with a language model to produce more contextually relevant and accurate answers than the model alone, using any MediaWiki installation as its knowledge base. Vector similarity search is provided by Milvus, run standalone or distributed, with the project's own Docker Compose setup recommended for testing and development.
Configuration is split between a config file for non-secret settings such as which MediaWiki sites and namespaces to ingest, which models to use and search parameters, and a separate env file for secrets such as API keys and tokens; a migration script converts an older, env-only setup into this newer split configuration automatically. The project is labelled experimental and described as part of Moodle's research initiative into educational-technology applications of retrieval-augmented generation.