Corpus2Skill
A research implementation that compiles a bounded document corpus into a navigable hierarchy of Anthropic Skills, so an LLM agent answers questions by browsing documents instead of querying a vector search service.
Corpus2Skill is the official implementation accompanying a paper accepted to Findings of EMNLP 2026 on distilling enterprise knowledge into navigable agent skills for question answering and retrieval-augmented generation. It converts a collection of documents into a structured tree of skill files; at query time an LLM agent navigates that hierarchy, reading top-level and sub-topic index files and drilling down, then fetches full documents on demand by ID, without embeddings, a vector store or BM25 at serving time.
Embeddings are used only at compile time, to cluster and label documents into the hierarchy before it is handed to an agent. This initial release includes the compile, serve and evaluate pipeline along with a quick start built on the WixQA dataset, and is licensed MIT.