Contents
FAQ
Can we use a SQL database instead of a graph database? Technically yes, but graph queries become expensive joins. Graph databases optimized for relationship queries 100-1000x faster.
How frequently should knowledge graphs update? Depends on domain. Financial data: hourly. Personnel data: daily. Taxonomies: weekly.
Do we need to fine-tune our LLM with knowledge graph data? Not required. Querying the graph at inference time (RAG) provides grounding without retraining.
Can we version knowledge graphs like code? Yes, using snapshots or event logs. Changes propagate like database migrations.
How large can knowledge graphs get? 1B+ entities possible (Google's Knowledge Graph). Requires distributed systems and optimization expertise.
Should we use public knowledge graphs or build custom? Hybrid: Start with DBpedia/Wikidata for common entities. Add domain-specific custom knowledge.
Related Resources
- Best Model Serving Platforms
- Fine-Tune LLM for Chatbot
- Best Feature Store Platforms
- Self-Hosting LLMs
Sources
- Neo4j Graph Database Documentation (neo4j.com/docs)
- Amazon Neptune Documentation (docs.aws.amazon.com/neptune)
- DBpedia Knowledge Base (dbpedia.org)
- Wikidata (wikidata.org)
- TransE Paper (arxiv.org/abs/1301.3485)