A retrieval service where every answer resolves to an exact document, page and paragraph, and tenant isolation is a CI test rather than a paragraph in a design doc.

01 Architecture

• Retrieval as a service: the engine is the sole owner of the indices. Producers upload, consumers call retrieve(), and nothing else touches the stores.
• Split by cost profile: ingestion is async and heavy, retrieval is sync and fast, so a large upload can never slow a query.
• Four stores behind one interface: a canonical DB, a vector index, a sparse BM25 index and blob storage, with hybrid search plus reranking across them.
• A citation and provenance resolver walks chunk to blocks to page, paragraph and bounding box, so every claim points at the exact place it came from.
• Tenant isolation is enforced at the index layer off a signed JWT claim, not in application code, with an adversarial CI test asserting one school's queries never return another's chunks.
• Object storage sits beside the engine: uploads land there first and emit an ingest event.

02 What is in it

• Ingest covers books, class notes, scanned handwriting, slides, images and lecture audio and video.
• Six AI consumers share one retrieval path: assignment generation, test generation, answer evaluation, flash cards, video generation and the learner model.
• Staged across four phases with a prior-art review and an implementation roadmap. Phase 0 is born-digital PDFs, a single tenant and dense-only pgvector.
• The security claim is a test, not a sentence: the isolation guarantee fails CI if it regresses.
• Knowledge tracing and an explicit latency budget are part of the spec rather than an afterthought.