How is a knowledge base different from a database?

The distinction is not absolute. A knowledge base emphasizes represented meaning and usable knowledge, while databases can provide its storage and may also support rules and inference.

Is a knowledge graph a knowledge base?

It can be. A graph is one way to organize entities and relationships within a knowledge base.

Are a language model's parameters the same as an explicit knowledge base?

They are different representations. Learned parameters do not generally expose a maintained, individually sourced set of factual records and rules.

AI Glossary / Graphs and symbolic AI / Beginner

Knowledge base

Definition

A knowledge base is an organized body of information that a system can use to answer questions or support reasoning. In symbolic AI, it commonly contains explicit facts, rules and relationships expressed in a representation with defined meaning. The term also has a broader everyday use for collections of support articles and reference material.

What it means

The knowledge base provides content; an inference engine or query mechanism uses that content. For example, a stored fact may identify an item as a camera, while a rule states that every camera is an imaging device. The system can then derive a classification that was not stored as a separate fact.

Different knowledge bases support different operations. A collection of searchable documents, a graph of entities and a formal logical theory are not interchangeable merely because all are called knowledge bases. Their treatment of meaning, uncertainty, inference and missing information can differ substantially.

A knowledge base can be incomplete or wrong. Its reliability depends on evidence, maintenance and how it is used.

How it works

Define the domain and the questions the system must answer. Choose identifiers and a representation for relevant objects, categories, properties and rules. Add information with sources and applicable context, such as the time or product version for which a statement holds.

A query may retrieve an explicit statement or ask whether a conclusion follows from several statements. In a logical system, entailment means the conclusion holds in every interpretation consistent with the knowledge base. In other systems, answers may instead be retrieved, ranked or generated from documents.

Updates need care because one revised fact can affect dependent conclusions. Missing information also needs an explicit policy. Under an open-world approach, failure to establish a statement does not establish its negation. A closed-world application may make a different assumption, but that assumption must be justified for the data and task.

Three practical examples

A product assistant retrieves maintained reference articles and identifies which version each article describes.
A symbolic classification tool derives that a named camera belongs to a broader equipment category.
A robot model records where objects were observed, together with the observation time and relevant relations.

Think of it this way

Think of a reference library with a catalog and stated rules for combining its information. The collection supplies material, while the catalog and reasoning procedures help find or derive an answer.

A library can contain outdated books or omit a topic. Likewise, a knowledge base does not become complete or correct merely because information has been organized and made searchable.

Worked example

Stored facts, derived facts and an unknown answer

Create a small logical knowledge base for a fictional equipment catalog. Store Camera(C1) and LocatedIn(C1, StudioA). Add the rule that every camera is an imaging device, and the rule that every imaging device is equipment.

A query asking whether C1 is equipment can be answered through two steps. Camera(C1) gives ImagingDevice(C1), and that gives Equipment(C1). This classification is derived; it need not be entered separately.

Now ask whether C1 is available today. No availability statement or rule has been provided. Under the open-world interpretation used in this example, neither AvailableToday(C1) nor its negation follows. The answer is unknown, rather than unavailable.

Next, add AvailableToday(C1) with a stated date. The system can answer the date-specific question, but should not silently reuse that answer tomorrow. Location and availability are also separate properties: being in StudioA does not imply availability unless an explicit, justified rule connects them.

Finally, suppose an incoming source says C1 is not an imaging device. That conflicts with the camera fact and classification rule under the stated logic. The update requires review; simply adding both statements can make a classical logical knowledge base inconsistent. Contradiction handling differs across knowledge-base technologies.

Illustrative case study

The fictional Birch Catalog team builds an assistant around the equipment example. Its first interface answers no whenever a query returns no supporting statement. Reviewers notice that this turns missing availability information into a false claim that equipment cannot be used.

Birch changes the response states to distinguish supported, contradicted and unresolved answers under its chosen logic. It also adds date scope to availability records and keeps the justification chain for derived category membership.

The team reviews conflicting source statements before accepting them into the maintained catalog. This fictional case illustrates the importance of semantics and upkeep.

Common misconceptions and mistakes

A knowledge base is always a set of FAQ pages. Formal AI knowledge bases can contain machine-interpretable rules and relations.
Everything it says is true. Source errors, stale facts and modeling mistakes remain possible.
No answer means the statement is false. That depends on the system's assumptions about completeness.
A retrieved passage is a logical proof. Retrieval relevance and formal entailment are different kinds of support.

Business use

Knowledge bases can make recurring information easier to find, reuse and review. They can also reduce inconsistencies when multiple applications draw from the same maintained definitions and facts.

For a project, specify the intended answer behavior, content ownership and update process. Keep evidence attached to important assertions and test ordinary, missing-information and conflicting-information cases. If a language model uses the base, distinguish what was retrieved from what the model inferred or generated. Reliable retrieval does not guarantee that the final wording faithfully reflects the source.

Check your understanding

Why does the catalog know that C1 is equipment but not whether it is available today?

Answer: Category membership follows from the camera fact and two rules. Availability requires different evidence that the initial knowledge base does not contain.

Related terms

Learn next

Build a five-fact catalog and two classification rules. Write one query answered directly, one answered by inference and one left unresolved. Then change a fact and identify which derived answers need to be reconsidered.

Sources

AIMA's logical-agent material distinguishes the knowledge base from the inference engine and explains entailment: https://aima.cs.berkeley.edu/4th-ed/slides-pdf/chapter07.pdf

Its first-order-logic material illustrates querying and representing changing facts: https://aima.cs.berkeley.edu/4th-ed/slides-pdf/chapter08.pdf

The W3C OWL primer explains formal ontology reasoning and open-world semantics: https://www.w3.org/TR/owl2-primer/

The equipment catalog and Birch Catalog case are original fictional examples.

Previous: Knowledge as a service Next: Knowledge-based agents