Grounding a knowledge base agent in your documents, not the open web

The pitch for a "knowledge base agent" is easy to say and easy to get wrong: an agent that answers questions using your own documents, policies, and records instead of guessing from whatever a general model picked up from the public internet.

Grounding means the agent can say "I don't know"

A knowledge base agent that always produces a confident answer, even when nothing in the client's documents supports it, is worse than one that says "I couldn't find anything about that in your policies." We tune retrieval and prompting specifically to prefer the second behavior.

The index lives inside the client's environment

Documents are indexed and queried inside the client's own infrastructure. There's no step where a client's internal policy documents get uploaded to a shared vector store outside their control.

Answers cite where they came from

Every answer the agent gives points back to the specific document or policy section it drew from. That's not a nice-to-have — it's what turns "the agent said so" into something a person can actually verify in a few seconds.

Where this has landed so far

This is the same underlying approach we use across HR policy questions, internal documentation, and — in early testing — CRM playbooks, adapted to whatever document set a given client actually has.