What is Knowledge Base?
Also called Corpus, Document Store.
A knowledge base is a curated collection of documents and structured facts assembled so that a system or a person can look up answers. In AI applications it usually means the indexed corpus a retrieval system searches, covering policies, product documentation, past tickets, and reference material. Its scope, freshness, and accuracy set the ceiling on answer quality.
The term predates modern language models and once referred mainly to structured rule and fact repositories in expert systems. In current usage it most often means a body of text prepared for retrieval, including source documents, extracted text, chunks, embeddings, and metadata. Both senses remain in circulation, so context decides which one is meant.
A useful knowledge base is more than a folder of files. It needs clear ownership so entries are corrected when reality changes, deduplication so contradicting versions do not both surface, and metadata recording origin, date, and audience. Without those, a retrieval system faithfully returns outdated or conflicting material and appears to be at fault.
Access control belongs to the knowledge base rather than to the model. Because a retrieved passage will be repeated in an answer, permission must be enforced at retrieval time by filtering the index to what the requester may see. Filtering after generation is not equivalent, since the content has already influenced the output.
Freshness is the recurring operational problem. Source systems change continuously, so ingestion runs on a schedule or reacts to change events, and deleted sources must be removed from the index rather than merely marked. A stale entry is more damaging than a missing one, because it produces a confident and wrong answer.
Key points
- Curated corpus a retrieval system searches for answers.
- Sets the quality ceiling for any grounded AI answer.
- Needs owners, deduplication, and origin metadata.
- Permissions must be enforced during retrieval, not after.
- Deletions must propagate through to the index.
In practice
A team assembles a knowledge base from a product handbook, two hundred resolved support tickets, and a pricing policy page. Each source is tagged with an owner and a review date. When the refund policy changes, the owner updates one page, the ingestion job reindexes it that night, and assistants answering refund questions the next morning quote the new terms.