padosoft-rag-knowledge-baselisted

# RAG knowledge base A corpus is not a folder of files with embeddings next to it. It is a store with an identity contract, a lifecycle and a trust ordering — and the three defects that follow from not having them are always the same: the same document indexed twice, a document that cannot be deleted, and an answer grounded on something nobody approved. **The rule: a model may propose knowledge; only a human promotes it. And everything the corpus holds carries its identity, its status and its origin explicitly.** --- ## 1. One idempotency anchor, and everything hangs off it ```text (tenant or project) + (source path) + (content hash) ``` Re-ingesting the same content is a no-op; changed content is a new version of the same document, not a second document. Get this wrong and the corpus grows duplicates that all retrieve, so the model sees the same passage three times and treats the repetition as agreement. **Normalise the path through one function, and use it everywhere** — ingest, read, delete, the command line and the HTTP surface alike. This is the rule people break by reimplementing the normalisation inline: - separators unified, repeated separators collapsed, leading and trailing ones trimmed; - relative segments **rejected**, not stripped — that is the traversal guard, and the inline version is always the one that forgets it; - an empty result raises, and the caller surfaces it as a validation error rather than swallowing it. **When ingest and delete normalise