Beyond "Chunk and Pray"

Top-k is not relevance. Retrieved text is not evidence. An LLM judging another LLM is not verification.

Conventional RAG papers over all three. Chunk the corpus, embed it, retrieve the nearest neighbors, and hope the generator does something reasonable with what comes back. It demos well. Then someone asks which number in the answer came from which filing, and the architecture has no answer to give.

This book builds GEODE-RAG instead — a retrieval architecture grounded in the Geometric Memory System, where the thing you retrieve is a verified structure rather than a hopeful paragraph.

The path runs from ingestion through knowledge-graph construction, numerical precision, and verification, to what it takes to run in production. Financial reports are the worked example throughout, with executable Python and reproducible notebooks, because that’s the domain where a number that drifts by a decimal place is a reportable event.

For teams whose retrieval systems have to hold up under audit, not just under a demo.