This demo reads plaintext data from data/ and generates a Knowledge Graph Index (using llamaindex). It does this by creating subject, predicate, object triplets.
It then opens an interactive CLI prompt for you to ask questions in natural language. So far, I have gotten the best results with the dolphin-mistral model.
At inference time, Colbert reranking is done to ensure the most topical graph nodes are fetched.
- Ollama (must pull the model you want to use)
- All
requirements.txtdependencies - bge embedding model (will be downloaded automatically)
Note: this software is in beta. You may get unexpected responses, or your knowledge graph may be incomplete. Finally, a reminder: garbage in, garbage out. Make sure your data is structured and easy to parse into triplets, or else you may have missing segments in your final knowledge graph.