The community members are discussing the behavior of the LlamaIndex Knowledge Graph Extractors, specifically the SimpleLLMPathExtractor. The main points are:
1. Each chunk of text is processed independently by the LLM model to extract triplets.
2. There is currently no good strategy in LlamaIndex to handle duplicate nodes or relationships across different chunks, unless the names are exactly the same (in which case they get merged).
3. This is a well-known problem with knowledge graphs, and there is no good automatic solution for it.
For Llamaindex Knowledge Graph Extractors (SimpleLLMPathExtractor), as I understand, each chunk text will be fetched into the LLM model to extract triplets. Then, will this LLM extractions of each chunk be independent from each other, or how do they know information regarding previous LLM calls (eg: how do they know about the other chunks and triplets from them)? If the calls are independent, then how does LlamaIndex ensure that there are no duplicate nodes / relationships in the knowledge graph, and how to deal with some information that is across chunks?
