Corpus.
One library for everything I know, read by every AI tool I work in. An AI librarian does the filing, and sensitive pages are handled by a model on my own machine.
IN PLAIN ENGLISH
What this is.
Corpus is the knowledge library underneath my AI tools. Everything stays readable as Markdown, new material should take one action, and sensitive pages are classified and embedded locally before anything cloud-hosted sees them.
01 / THE PROBLEM
Context Drop could deliver better notes than the library could keep.
The old library had long pages, several ideas mixed into one file, and enough capture friction to fall behind real work. Corpus is the rebuild: atomic pages, a clear write contract, and a derived search index that can be discarded and recreated.
02 / WHAT IS BUILT
The source of truth is intentionally boring.
Everything lands as plain Markdown in one vault. The contract every page must satisfy, the write worker that owns commits, and the chunker that splits pages into single-idea pieces are built. Search indexes are derived artifacts, not the thing to preserve.
Reversible indexes
Changing a chunker or embedding model means rebuilding an index, not migrating the notes. A clean-machine restore drill is the proof that this is more than a slogan.
Privacy order matters
A local classifier must decide whether captured content is sensitive before a cloud embedder, reranker, or librarian can read it.
03 / THE BET
Atomic pages should beat hand-curated search.
Context Drop reaches 0.85 precision-at-5 on its curated gold set, while raw hybrid search over the same notes scored 0.56. Corpus is designed to close that gap with structural rules and no hand-curated pack prior. The shared question set is the gate; if the bet fails, the old system keeps serving.
04 / WHERE IT STANDS
Building, with the risky parts named plainly.
The target capture flow, cloud-boundary enforcement, and librarian are not all complete. Independent probes have already found path traversal, concurrency loss, and chunker corruption in earlier waves. Those probes now belong in the build, and they must be calibrated against deliberately broken code before a clean run counts.
THE HONEST VERSION
What I am actually counting.
Corpus is deliberately described as building. The Markdown vault and write contract exist; the capture-time classifier, enforcement test, and librarian remain the work that makes the privacy promise real.
TRY IT
Try the interaction.
TRY IT · VAULT ROUTING
A hard-coded simulation. It never reads a real vault.
TOOLS AND CONSTRAINTS
- Markdown + Git
- MCP / REST target
- Derived vector indexes
- Local classifier
- Probe-driven verification