Glossary

AI knowledge base

An AI knowledge base is an organization's documents and records, collected from the systems where they live and indexed so AI assistants can search them and give cited answers. It is the content side of retrieval-augmented generation, and its answers are only as current and as correctly permissioned as its index.

Four parts make it up

  • Connectors pull content from file shares, object stores, CRM and ticketing systems, and watch them for changes.
  • Processing splits text through document chunking and converts each passage into embeddings.
  • Indexes, one by meaning and one by keyword, make the content searchable; hybrid search queries both.
  • Metadata records the source, version, date and permitted readers of each document.

When someone asks a question, the assistant retrieves the best-matching passages and a language model writes the answer and cites them. That flow is retrieval-augmented generation.

Drift between source and index

Drift is the gap between what the source systems hold and what the index returns, and it produces more bad answers than the model does. It takes three forms. Deleted content stays answerable, so an assistant quotes a withdrawn policy, complete with a citation. New content takes days to appear. Permission changes do not carry over, so people receive answers drawn from files they could never open. Each form is an event the index failed to process. A deletion or a permission change at the source is as much an event as a new file, and systems that only add content and re-index on a schedule leave all three gaps open between runs.

Measuring freshness with index lag

Index lag is the time from a change at the source to that change being reflected in search results. It is worth measuring separately for new documents, edits, deletions and permission changes, because each travels a different path. A policy withdrawn on Monday and still cited on Thursday is a deletion lag of three days, even if new documents appear within the hour. Citations support audits only when each one carries a document version, since a link to a file that has changed since proves little.

How does it differ from a wiki or a data platform?

A wiki is content people write and browse; a knowledge base indexes content wherever it already sits, including tickets and file shares. An AI data platform is broader, serving training and analytics as well as retrieval. The index itself lives in a vector database or search engine and can be rebuilt. The source documents cannot, so they are kept durable and versioned. Wrong answers drawn from good sources usually trace to retrieval rather than content, a failure mode known as RAG hallucination.

Frequently asked questions

Can an AI knowledge base run without sending data outside the organization?

Yes, when the connectors, the index and the model all run in an environment the organization controls. Where data is processed is a design choice, not a property of the term.

Why do AI answers cite the wrong version of a document?

Usually because version information was not stored in the index, or an old copy was never removed. Versions and dates held as metadata let retrieval prefer the current document.