Knowledge Vault

Google's Knowledge Vault is a large-scale probabilistic knowledge base automatically constructed from billions of facts extracted from web pages, used as the ground truth reference when computing KBT scores for web sources.

What the Knowledge Vault Is

The Knowledge Vault was introduced by Dong et al. in a separate 2014 paper and serves as the factual backbone of the Knowledge-Based Trust framework. Rather than being manually curated (like the Knowledge Graph), the Vault is assembled automatically by running information-extraction pipelines over the entire web and aggregating the resulting subject–predicate–object triples.

Each triple in the Vault comes with a probability score reflecting confidence in its accuracy. Triples with very high confidence — typically because they appear consistently across many independent, credible sources — are treated as near-ground-truth facts for the purpose of KBT scoring.

Scale and Coverage

At the time of the KBT paper (2015), the Knowledge Vault contained approximately 1.6 billion facts with associated confidence scores, covering entities, relationships, and attributes across virtually every topic domain. This made it by far the largest automatically constructed knowledge base in existence at the time.

Coverage is uneven across domains: facts about widely documented entities (prominent people, countries, major companies, scientific concepts) are densely represented with high-confidence triples, while niche or obscure entities may have sparse or low-confidence coverage.

How It Is Used in KBT

When a KBT pipeline encounters an extracted fact from a web page, it queries the Knowledge Vault for matching or conflicting triples:

• If the Vault contains a high-confidence matching triple, the page's claim is labelled supported.
• If the Vault contains a high-confidence contradicting triple, the claim is labelled contradicted.
• If coverage is sparse or confidence is low, the claim is treated as uncertain and handled probabilistically.

The quality of the Knowledge Vault therefore directly determines the quality of KBT scores: domains well-covered by the Vault yield reliable KBT scores; domains with sparse Vault coverage yield noisier estimates.

Knowledge Vault vs. Knowledge Graph

Google's publicly visible Knowledge Graph (which powers info boxes in search results) is a curated subset of high-confidence facts. The Knowledge Vault is its much larger, probabilistic, automatically-generated counterpart — an internal research system rather than a production product. The two systems share structural similarities but differ radically in scale and curation level.

Public Alternatives

While the Knowledge Vault is internal to Google, open knowledge bases such as Wikidata and the now-retired Freebase operate on similar principles. LLM training pipelines frequently use Wikidata as a structured-fact reference, making it the closest publicly accessible analogue to the Knowledge Vault for practitioners who want to align their content with machine-readable knowledge representations.