What this is

Zhira is Shona for “path(s)”. Extraction and organisation were solved years ago by Nota and the CLI; what was left was retrieval, finding the passage half-remembered, and the connection between two books read four years apart.

The corpus is real: currently over 14,000 records from 299 books, roughly 620,000 words, going back 6 years.

The product is a versioned context pipeline: source data changes, and the system produces a new, checksummed artifact you can diff, verify, and serve, to a person or to a model with citations back to the book and location.

Where it currently is

Early build. What exists today:

• a fresh-ingest pipeline — parse, canonicalise, merge, index, atomic promote;
• sync classification, detecting added, changed, and deleted records against a prior artifact by source checksum;
• a doctor command that validates the manifest and re-verifies recorded checksums;
• BM25 search over an SQLite FTS5 index.

Dense and hybrid retrieval, cited answers, and the MCP server that exposes the corpus to agents are planned, not built.

Github