Knowledge Graph Process

Hey Queso, llama index actually uses a single prompt to extract triplets across text chunks. It should make more sense when you check out the default prompt for this here: https://github.com/jerryjliu/llama_index/blob/main/gpt_index/prompts/default_prompts.py#L241

Also, check out the demo notebook as well: https://github.com/jerryjliu/llama_index/blob/main/examples/knowledge_graph/KnowledgeGraphDemo.ipynb

At query time, llama index extracts keywords from the query and look for triplets that have the same subject as each keyword.

If include_text is true (the deafult) llama index include the triplet and the text chunk where the triplet was found when we ask the LLM to answer the query

Thanks @Logan M!. Got it now! Do you know if there's a study that it's based on? Trying to get a gauge of how good it is and how good it works on a large document.

Hmm not sure if it was based on anything. Looking at Google scholar though, "LLM knowledge graph" brings up quite a few papers

In my experience, the performance is not too bad. You'll notice the demo limits to 2 triplets per 512 tokens, which means it doesn't really matter how big the document is since it only looks at chunks with 512 tokens (with some overlap of course)