Retrieval Grounding Handbook
How to ground LLM answers in retrieved evidence, cite it, and abstain when the pack does not say — including prompt-injection boundaries.
Portable policy knowledge for local retrieval.
Retrieval Grounding Handbook is original guidance for hosts that query Knowledge Images or other corpora and then talk to a model.
The pack is the authority. The model is a speaker. If retrieval is empty, abstain.
Agent metadata in a Knowledge Image is not executed by Hub.
Intended use
Grounding product and agent design when answers must point at a source, not a vibe.
Out of scope
Not a hosted model, not a vector database, and not a guarantee of factual truth outside the mounted pack.
Sources
Sample questions
This Knowledge Image contains prompts, tool policy, namespace grants, or other agent registry metadata.
Mounting the pack for retrieval does not execute tools. Applying these policies affects your application only if your host explicitly honors them.
Credentials must never be stored inside a pack.