A real incident
A fabricated-citation scandal became the most cited example of an AI knowledge failure, and the chapters next door build the layer that prevents it.
The chapters
60 min total
- Context failures7 minSort a hallucination into fabrication, staleness, or loss, and place it on the supply chain.
- Retrieval: look it up8 minChoose one-shot or agentic retrieval by question type and stakes, and catch its failure modes.
- The search stack8 minRoute each question to keyword, vector, hybrid, or structured lookup, and size your chunks.
- Memory: what it keeps7 minSort memory on two axes, run it as a consented pipeline, and let users see, edit, and delete.
- Context budgets7 minSplit each request across rules, memory, facts, and history, and cut filler that rots answers.
- Freshness and conflicts7 minGive every source an owner, a shelf life, and a recall path, and rule which wins a conflict.
- Grounding evals8 minMeasure retrieval hit rate and grounded-answer rate, then gate every release on both.
- The Knowledge Charter8 minWrite one signed page recording every knowledge decision, and pressure-test it before you ship.
The sum
Put together, these chapters are the knowledge layer of your product: a lookup step, a memory policy, a context budget, governed sources, and a grounding eval, each one a decision you can defend rather than a default you inherited.
The Knowledge Charter is one signed page: fill it and every source has an owner, every memory has a rule, and every answer has a bar you can defend.
The Knowledge CharterFillable PDFDownload →