Google's AI research engine: upload PDFs, documents, slides, or web pages into a notebook; Gemini embeds and retrieves the relevant passages and answers with clickable citations back to the exact source, plus generated audio overviews. Near-zero setup, and scoped strictly to the sources you uploaded.
Grounded, cited answers over a bounded set of your own sources — the research describes it as answering "with clickable citations" and refusing "if evidence isn't present." This is arguably the best citation experience surveyed anywhere on this matrix, and it's a shipped, working feature rather than a design intent.
Both are explicitly grounded-retrieval systems built to refuse answering from thin air — Mimir's provenance chain (GET /provenance/:chunk_id) and NotebookLM's clickable citations are solving the same trust problem from different architectures.
NotebookLM is scoped to whatever you deliberately uploaded into one notebook; Mimir is designed as a standing, continuously-capturing memory across everything — Slack, Drive, Asana, and fast-lane thoughts — rather than a per-project research tool. Mimir also adds the compulsory governance layer and an agent-facing audited API; NotebookLM is a human-facing product with no public agent/API surface.
Actual semantic retrieval, right now. NotebookLM's embeddings are live and its citations are a shipped, polished feature. Mimir's embedding column exists in the schema but sits at 0% coverage today, so retrieval is lexical-only until the embedding decision closes — on the one row both systems compete on directly, grounded retrieval quality, NotebookLM is ahead.
research/07-zero-structure-rag.md