Knowledge bases (KBs) are a cornerstone of the Semantic Web, yet they still struggle with scale and scope, and their construction and curation still involve a lot of manual efort. Large language models (LLMs) have recently emerged as powerful tools for a range of tasks, yet their potential for automated KB construction is still poorly understood. In this demonstrator, we showcase GPTKB, a methodology and KB entirely built from GPT-4.1. GPTKB is constructed by massive-recursive LLM knowledge materialization [1], using over 9M API calls for $14,000 to construct a 100M-triple knowledge base with over 6M entities. Our demonstration focuses on two use cases: (i) Link-based KG exploration and (ii) SPARQL-based analysis and comparison to Wikidata. The GPTKB demonstrator is accessible at https://gptkb.org.
1. Link-based interactive knowledge graph (KG) exploration; 2. SPARQL-based analysis and comparison to Wikidata.
The GPTKB methodology [
Knowledge elicitation Starting from a seed subject, the LLM is prompted to return knowledge about it in the form of triples. New named entities in these triple objects are identified via LLM-based named-entity recognition (NER) and are enqueued for further elicitation in a recursive BFS-based graph
Triples 100M (120M with meta-relations) Relations 936k (381k after canonicalization) Classes 220k (32k after canonicalization) Triple objects 59M entities, 41M literals
Subject-precision 85.3% Verifiable, 3.4% Plausible 11.3% Unverifiable
Triple-precision 75.5% True, 5.0% Plausible, 19.5% False
Wikidata Wikidata5m Yago 4.5 DBpedia
NELL ReVerb
GPTKB v1.5 #entities
#assertions 113M 5M 50M 3.8M ? ? 6.1M 1.62B 20M 140M 75M 12M 15M 100M exploration process. Constrained decoding is used to make sure that outputs stay within the triple format.
Knowledge consolidation To address the redundancy and variance introduced during knowledge elicitation, post-hoc knowledge consolidation is performed. In particular, we apply a greedy clustering algorithm to iteratively merge relations and classes into more frequent ones, given a suficiently high label embedding similarity.
Further methodological details can be found in Hu et al. [
Two versions of GPTKB are available, GPTKB v1.1, based on GPT-4o-mini [
Following the paradigm described in Section 2, We extracted knowledge from GPT-4.1 starting with the seed entity Vannevar Bush. The whole process cost $14,136 for OpenAI API calls and took 18 days. The final KB contains 100 million triples derived from 6.1 million entities in total, organized into 381k relations and 32k classes. We provide statistics of GPTKB v1.5 in Table 1.
Since crawl parallelization distorts BFS search order, we post-hoc recomputed the shortest paths of each node to the root, and stored this information in two meta-relations, bfsLayer and bfsParent, to enable structural insights. To facilitate data interchange, we also converted GPTKB into RDF format, and serialized it into Turtle syntax.
We performed two quality evaluations. An automated method based on web search, like in Hu
et al. [
We provide multiple modes of access to GPTKB. URI: https://gptkb.org/entity/Nara_City GPTKB entity Statements (31) Predicate Object gptkbp:instanceOf gptkb:city gptkbp:area 276.84 km² gptkbp:country gptkb:Japan gptkbp:famousFor gptkb:Todai-ji_Temple gptkb:Kasuga-taisha_Shrine gptkb:Nara_Park
Nara deer gptkbp:founded February 1, 1898 gptkbp:capitalOf gptkb:Nara_Prefecture gptkbp:formerName gptkb:Heijō-kyō gptkbp:region gptkb:Kansai
. . . truncated . . . gptkbp:bfsParent gptkb:Honshu
gptkb:Nara_Park gptkbp:bfsLayer 5
Todai-ji Temple URI: https://gptkb.org/entity/Todai-ji_Temple GPTKB entity Statements (51) Predicate Object gptkbp:instanceOf gptkb:Buddhist_temple gptkbp:afiliation gptkb:Kegon_school gptkbp:annualEvent gptkb:Omizutori
gptkb:Shuni-e gptkbp:annualVisitors millions gptkbp:architecturalStyle gptkb:religious_art gptkbp:contains gptkb:Hokkedō gptkb:Nandaimon_(Great_South_Gate) gptkb:Nigatsu-dō gptkb:Shōsōin gptkbp:coordinates 34.6889°N 135.8399°E gptkbp:country gptkb:Japan
. . . truncated . . . gptkbp:bfsParent gptkb:Nara_City gptkbp:bfsLayer 6
Firstly, GPTKB is hosted on the https://gptkb.org web server that provides a user interface to search entities via keyword queries and to perform link-based exploration to discover new connections and entities. Section 5.1 and Figure 1 provide a demonstration experience of this link-based exploration. The web interface is implemented by using the Python Django framework, and hosted on a Nginx web server. The KB is stored in an OpenLink Virtuoso server.
Secondly, we provide a SPARQL endpoint at https://gptkb.org/query/ that supports structured queries within a timeout window of 100 seconds.
Thirdly, we provide the RDF dump under the CC BY 4.0 license on the HuggingFace datasets library at https://huggingface.co/Knowledge-aware-AI.
We divide the demonstration experience in two parts: (1) link-based KB exploration, and (2) structured SPARQL analytics and Wikidata comparison. 5.1. Link-based KB Exploration Data about specific entities can be accessed via multiple routes: 1. The start page features a selection of direct links to entities such as Vannevar Bush and San
Francisco. 2. The web portal features a search field in the top-right corner, which can be used for string-based search. 3. If an unambiguous entity name is known, one can directly access the entity’s KB entry via the
URL https://gptkb.org/entity/<NAME>.
Figure 1 shows how to initiate an entity search and continue with a link-based exploration. We start by typing this year’s ISWC venue, Nara, which returns 314 results. Clicking on the Nara City result takes us to its entity page. The entity Nara City contains 31 statements, and a user can click on any of the entity objects, here for instance, Todai-ji Temple, to further explore connected entities. PREFIX gptkb: <https://gptkb.org/entity/> PREFIX gptkbp: <https://gptkb.org/prop/> SELECT ?o (COUNT(*) AS ?ofreq) WHERE {
?s gptkbp:instanceOf ?o. } GROUP BY ?o ORDER BY DESC(?ofreq) LIMIT 100 o gptkb:person gptkb:human gptkb:film gptkb:company gptkb:book gptkb:song gptkb:fictional_character ofreq 1,077,803 138,646 120,497 118,993 111,414 103,538 90,499
Since each entity in GPTKB was identified as object from a parent entity during the BFS algorithm, we provide this information via the bfsParent relation. Additionally, the bfsLayer relation tells us at which layer knowledge elicitation was performed for the entity. In Figure 1, we learn that Nara City is a child entity of Honshu, and that its triples were elicited in layer 5. Clicking on the parent entity lets a user move up the layers of GPTKB. 5.2. SPARQL Querying for Analytics and Wikidata Comparison A core feature of structured query languages is that they allow statistical analysis at scale. For this purpose, the GPTKB content is stored in a Virtuoso Triple store, whose content is exposed via a SPARQL query interface available at https://gptkb.org/query/. In the following, we show enabled analyses. Most frequent classes Just what kind of entities does GPT know about? An overview is provided by the following query: The results are fundamentally diferent from, e.g., Wikidata, with a much stronger focus on digital artifacts (films, songs), and fiction.
Nationality bias Existing KBs as well as LLM training corpora are known to be Western- and Englishlanguage dominated [9], can this bias also be observed at the factual level of GPTKB? A quick glimpse can be obtained by counting the number of citizens per country known to GPTKB: PREFIX gptkb: <https://gptkb.org/entity/> PREFIX gptkbp: <https://gptkb.org/prop/> SELECT ?o (COUNT(*) AS ?ofreq) WHERE {
?s gptkbp:nationality ?o. gptkb:American British gptkb:French gptkb:German gptkb:Indian gptkb:Canadian gptkb:Australian Japanese ofreq Notably, English language nationalities occupy the top places, at a much stronger bias than existing resources like Wikidata.
Factual knowledge of LLMs is intensively researched, mostly via sample-based benchmarks or probes, such as the seminal LAMA probe by Petroni et al. [10]. However, these works typically draw sample from existing web resources, thereby introducing a confirmation bias that prevents the discovery of unexpected knowledge (or errors). For example, LAMA drew 50k triples from Wikidata, Wikipedia, and ConceptNet.
Few works have harvested LLM knowledge at scale. Nguyen and Razniewski [11] harvested one
million commonsense assertions from BART and GPT-2, based on a pre-defined subject list. Cohen
et al. [12] proposed to crawl factual LLM knowledge by recursively prompting them. Parović et al. [13]
proposed domain-specific KB construction from LLMs, but did this only at the scale of a few hundred
thousand entities. In this demo we build upon the GPTKB methodology by Hu et al. [
Several large knowledge bases are deployed online [14], most notably Wikidata [2], Yago [3] and DBpedia [4]. Our web browsing and query interfaces are inspired by those.
In terms of LLM-generated datasets, the closest to ours might be Cosmopedia [15], an LLM-generated 25 billion token text corpus. However, Cosmopedia is intentionally designed to synthesize realisticlooking but invented texts, and has no goal of collecting factual LLM knowledge.
We have presented the https://gptkb.org web demonstrator, a knowledge base browser and query
interface to GPTKB, a massive 100-million-triple KB built from GPT-4.1 using the GPTKB methodology
[
The author(s) have not employed any Generative AI tools. [2] D. Vrandecic, M. Krötzsch, Wikidata: a free collaborative knowledge base, Commun. ACM 57 (2014). doi:10.1145/2629489. [3] F. M. Suchanek, G. Kasneci, G. Weikum, Yago: a core of semantic knowledge, in: WWW, 2007.
doi:10.1145/1242572.1242667. [4] S. Auer, C. Bizer, G. Kobilarov, J. Lehmann, R. Cyganiak, Z. G. Ives, DBpedia: A nucleus for a web of open data, in: ISWC, 2007. doi:10.1007/978-3-540-76298-0_52. [5] A. Fader, S. Soderland, O. Etzioni, Identifying relations for open information extraction, in: EMNLP, 2011. URL: https://aclanthology.org/D11-1142. [6] T. Mitchell, W. Cohen, E. Hruschka, P. Talukdar, B. Yang, J. Betteridge, A. Carlson, B. Dalvi, M. Gardner, B. Kisiel, et al., Never-ending learning, Communications of the ACM 61 (2018). doi:10.1145/3191513. [7] J. Z. Pan, S. Razniewski, J.-C. Kalo, S. Singhania, J. Chen, S. Dietze, H. Jabeen, J. Omeliyanenko, W. Zhang, M. Lissandrini, et al., Large language models and knowledge graphs: Opportunities and challenges, Transactions on Graph Data and Knowledge (2023). doi:10.4230/TGDK.1.1.2. [8] Y. Hu, T.-P. Nguyen, S. Ghosh, M. Müller, S. Razniewski, GPTKB v1.5: A massive knowledge base for exploring factual LLM knowledge, arXiv (2025). doi:10.48550/arXiv.2507.05740. [9] G. Wenzek, M.-A. Lachaux, A. Conneau, V. Chaudhary, F. Guzmán, A. Joulin, É. Grave, CCNet: Extracting high quality monolingual datasets from web crawl data, in: LREC, 2020. doi:10.48550/ arXiv.1911.00359. [10] F. Petroni, T. Rocktäschel, S. Riedel, P. Lewis, A. Bakhtin, Y. Wu, A. Miller, Language models as knowledge bases?, in: EMNLP, 2019. doi:10.18653/v1/D19-1250. [11] T.-P. Nguyen, S. Razniewski, Materialized knowledge bases from commonsense transformers, in: Workshop on Commonsense Representation and Reasoning (CSRR), 2022. doi:10.18653/v1/ 2022.csrr-1.5. [12] R. Cohen, M. Geva, J. Berant, A. Globerson, Crawling the internal knowledge-base of language models, in: EACL Findings, 2023. doi:10.18653/v1/2023.findings-eacl.139. [13] M. Parović, Z. Li, J. Du, Generating domain-specific knowledge graphs from large language models, in: Findings of the Association for Computational Linguistics: ACL 2025, 2025. [14] G. Weikum, X. L. Dong, S. Razniewski, F. Suchanek, Machine knowledge: Creation and curation of comprehensive knowledge bases, Foundations and Trends in Databases 10 (2021). doi:10.1561/ 1900000064. [15] L. B. Allal, A. Lozhkov, D. van Strien, Cosmopedia: A new frontier for open-source language models, 2024. URL: https://huggingface.co/blog/cosmopedia, Hugging Face blog.