Integrate a Large Language Model (LLM) AI with a Knowledge Graph

Connect a Large Language Model AI service with a knowledge graph so that the information source is scoped and accurate. This provides a reliable information repository for using the LLM in a specific business area.
What can you do with an LLM and a knowledge graph?

LLMs can be used to get information into your knowledge graph, and to get intelligence out of your knowledge graph. Some examples are:-

  1. Ask questions about the information in your knowledge graph using natural language queries, both vocal and written.
  2. Generate natural language reports using information in your knowledge graph.
  3. Generate visual reports using information in your knowledge graph.
  4. Create structured data to feed into your knowledge graph from unstructured sources; websites, videos, books, archives etc.
  5. Fill in missing information in your knowledge graph; e.g. summaries, data.
What are the main problems when integrating LLMs with knowledge graphs?
  1. Hallucinations: The LLM creates fictitious data/information that looks convincing, but is in fact nonsense. This includes e.g. wikidata IDs and IAB IDs for datasets that it might claim to know.
  2. Control: Getting the LLMs to "stay on the rails" and do what you ask - e.g. sticking to constraints that you feed into the process.
  3. Repeatability: Getting reliable and consistent results back is currently an issue.

These FAQ answers are based on tests in summer 2023 on chatGPT and Bard. They should all improve during 2023-2024. We are keeping track of feature areas.

What are the benefits of using LLMs with Knowledge Graphs?
  1. An LLM can be used as a natural language interface onto your knowledge graph.
  2. A knowledge graph provides an information source with context and meaning to an LLM service.
  3. A knowledge graph can scope the information used by the LLM to your business area, so that you are not plagued with hallucinations from elsewhere in an LLMs training corpus.
  4. Some LLMs understand RDF models, SPARQL and Cypher graph querying languages, so they can make full use of your graph structure.