What is a knowledge base in AI?
A knowledge base, in the context of generative AI, is the curated set of documents and records that an assistant or agent draws on to answer questions. It typically includes policies, procedures, product documentation, contracts, resolved tickets, and FAQs, stored as searchable chunks with metadata in a vector database. The knowledge base grounds the model in one company’s facts, which its training data lacks.
The word curated distinguishes a knowledge base from a document collection. A knowledge base has an owner, a refresh process, and a rule for what counts as authoritative. A raw copy of a file share lacks all three, and an assistant built on one answers from drafts, duplicates, and superseded versions.
Knowledge base in practice
A managed service provider builds its knowledge base from runbooks, vendor documentation for supported products, and resolution notes from closed tickets that a senior technician has marked as verified. Each item carries the client it applies to, the product version, and a review date. The technician’s assistant answers from this set and cites the runbook step. When a vendor ships a new version, the affected runbooks are flagged and excluded from retrieval until updated.
What to watch for
- Assign an owner to each content area, because unowned content goes stale and the assistant repeats it confidently.
- Decide how superseded documents leave the index, since the model cannot tell an old policy from a current one.
- Track which questions get no good source and use that list to decide what to add next.