Get answers with evidence
Retrieve relevant facts from the graph and use them to ground an AI answer. Check the sources cited in the response.
Content knowledge that grounds AI answers
LETR K-Graph
Connect content facts, relationships, and sources to AI. Follow works, people, and original stories to produce answers with evidence you can check.
Content data
snapshot
- Entities
- 190K+
- Relationships
- 450K+
- Evidence records
- 572K+
Includes unverified candidates · Confirmed data tracked separately
Korean content facts · Internal comparison ·
More accurate answers than the LLM-only baselines
K-Graph retrieved connected facts to answer questions about roles, soundtracks, and adaptations. It answered 109 questions, with 108 scored correct.
K-Graph
Knowledge graph + LLM
99.1%Accuracy among answered questions
109 answered · 108 correct
Correct across all 119 questions 90.8%
General-purpose LLM A
Model knowledge · No web search
67.1%Accuracy among answered questions
82 answered · 55 correct
Correct across all 119 questions 46.2%
General-purpose LLM B
Model knowledge · No web search
43.2%Accuracy among answered questions
37 answered · 16 correct
Correct across all 119 questions 13.4%
Questions were built from K-Graph data, and the comparison LLMs had web search disabled. This is not a comparison across all domains or against AI services with search enabled.
Answering questions across connected relationships
For questions spanning an original work, its adaptation, and soundtrack performers, the system follows the connected facts.
| Question type | K-Graph | General-purpose LLM A | General-purpose LLM B |
|---|---|---|---|
| Actor for a role | 92.0% | 60.0% | 18.0% |
| OST performer | 100.0% | 42.0% | 8.0% |
| Original → adaptation → OST | 63.2% | 21.1% | 15.8% |
Evaluation conditions and interpretation
- 169 questions: 119 answerable and 50 deliberately unanswerable. All systems received the same questions, answer format, scoring, and instruction to abstain when uncertain.
- K-Graph combines a knowledge graph with an LLM. This was not an experiment adding a graph to the same baseline model. The comparison models are anonymized as A and B.
- Each system gave one incorrect answer among the 50 unanswerable questions. These 50 questions are excluded from the accuracy figures above.
- The report attributes K-Graph’s one incorrect score to a name-format mismatch; its original score is retained. There were also four questions a comparison model answered correctly while K-Graph abstained.
- A source citation does not mean the underlying fact has been verified. Source quality, review status, and graph coverage still need to be checked.
Source: TWIGFARM, K-Graph Accuracy Report, internal experiment dated 2026-09-01. This page summarizes the reported results.
How it helps your team
For content planning and business teams checking detailed facts, and service teams that need a dependable knowledge foundation for AI answers.
Follow several relationships at once
Explore questions that connect an original work to a series adaptation and its soundtrack performers.
See what can be supported
Review source and verification status. The system is designed to abstain when evidence is missing and distinguish information needing further review.
From works to connections
Explore content through its people, companies, roles, and related works.
Content works34,375
Films, series, music, webtoons, novels & more
Work ↔ original & adaptationsCast & creative credits36,033People
Production & distribution8,199Companies & organizations
Appearances & performers111,380Roles & characters
Performances & membership503Music groups
- Films8,270
- Series4,409
- Webtoons2,008
- Songs10,581
- Albums6,948
Corroborated or human verified5,133 factsRelationships and attributes
Track the evidence behind the scale
547,090 relationship and attribute records are managed with their sources and review status. The 5,133 confirmed facts comprise 5,124 corroborated across independent sources and 9 reviewed by a person.
The catalog connects content metadata collected from Wikidata, KMDb, MusicBrainz, KOMACON, official materials, and other sources.
Counting rules and review status
Entity totals exclude merged and rejected records. They include 551 minimal identity records. Work counts refer to metadata records, not original media files.
450,315 relationships connect one entity to another. Including attributes such as release dates and categories brings the total to 547,090 facts. Multiple evidence records may support one fact; evidence counts are not counts of websites or works.
- Awaiting verification
- 540,245
- Corroborated across independent sources
- 5,124
- Human verified
- 9
- Conflicting sources awaiting review
- 1,712
4,769 of the confirmed facts are entity-to-entity relationships. All figures use the same snapshot and may change with collection and review.
AI answers need a knowledge foundation you can check
When using AI in planning, research, or content recommendations, teams need to know which facts support an answer. Detailed questions about supporting roles, soundtrack credits, and adaptations can turn a plausible response into a consequential mistake.
K-Graph structures content facts, relationships, and sources for AI to use. It identifies works and people in a question, retrieves connected facts, and supplies those facts as evidence to an LLM. The graph retrieves the evidence; the LLM turns it into a readable answer.
From retrieved facts to answers with evidence
A question about soundtrack performers in a series adapted from a webtoon requires following original work → adaptation → soundtrack → performer. Graph relationships help connect facts that would otherwise need to be assembled from separate search results.
The answer’s evidence exposes sources and review status. A sourced fact may still be awaiting review, and the system is designed to abstain when the graph lacks the necessary evidence. Both data coverage and review quality therefore matter.
Find the content and its connections
Information about films, series, music, webtoons, and novels is spread across many sources. LETR K-Graph organizes works, people, companies, roles, and groups as distinct entities, then connects who participated, who produced a work, and which original work it adapts.
Search by name or alias and follow related works and people. External identifiers and contextual information help distinguish entities that share a name.
Put accumulated knowledge to work
Explore works, people, and companies
Move from a work to its cast, crew, production companies, and related works. Examine the connections needed for content research, planning, and partner discovery.
Trace originals and adaptations
Follow relationships between novels, webtoons, series, and films to understand how an IP develops across media. Review connections that would otherwise remain scattered across separate records.
Review the evidence behind the data
Relationships and attributes retain their sources, observation times, and review status. Conflicting information and items awaiting review are distinguished, with an audit history of reviewed changes.
Current scope
The system supports content, person, organization, and character search; detail and neighboring-relationship views; original/adaptation links; and source review. The accuracy report evaluates natural-language question answering with a knowledge graph and an LLM. Adoption includes agreeing on question types, data coverage, and how users check the evidence behind an answer.
Scheduled collection and source checks enrich the catalog and expand the questions it can support.
The API integration for finding K-Graph entities and adding them as observation targets in AI Signal within LETR WORKS is planned. This is the next step toward using connected content knowledge in AI visibility diagnostics.
Collection scale and intended use
Metadata for films, series, music, webtoons, and novels is connected from Wikidata, KMDb, MusicBrainz, KOMACON, official materials, and other sources. Relationships and attributes carry source evidence, observation dates, and review states, with conflicting sources identified.
Entity, relationship, and evidence figures on this page share a snapshot date. They are not counts of owned media, IP rights, or customer deliveries. Confirmed facts and unverified candidates are kept distinct when defining data for research or analysis.
An adoption discussion starts with the content domain, use case, and relationships needed, then reviews available coverage, sources, and verification criteria.
How it works
-
01 Ask your question
Ask in natural language about the works, people, or original stories you need to understand.
-
02 Connect the evidence
Identify the relevant entities and follow cast, production, and adaptation relationships to find supporting facts.
-
03 Review the answer and sources
Check the AI response, cited evidence, and review status before using it in research or planning.
See how it fits your team
Tell us about your work and materials. We will help you define the right scope.