Arbol
A folder of notes should be a useful source of context, not just a pile of text.
The problem
Retrieval quality starts before a model sees a prompt. Unstructured notes make it hard to distinguish knowledge from navigation, recover relevant context, and trace an answer back to a source.
The approach
Arbol puts a schema, metadata, stable identifiers, validation, and structured exports ahead of retrieval. Its documented pipeline combines BM25 lexical search with a link graph, reranking, and diversity selection. Context packs carry a citation manifest.
The engineering choice
The project uses Python’s standard library and keeps the model call outside the retrieval layer. This makes the retrieval component independently testable and lets different agent clients consume the same sources through MCP.
What this demonstrates
Data modeling, pipeline design, retrieval architecture, developer tooling, and the discipline to evaluate whether a more complicated approach actually helps.
Current scope
The public repository documents a single-user, local-first system. Lexical retrieval has vocabulary limits, and results depend on the corpus and evaluation set. This portfolio does not claim a universal retrieval benchmark.