Why not just use an LLM to extract every triple in your...

Using an LLM to extract triples from every document in a corpus is often impractical due to the nature of the underlying data. When dealing with a structured taxonomy, the schema is stable and hierarchical, allowing for consistent extraction. In contrast, unstructured data such as call transcripts contain relations that must be inferred on a per-call basis. Because these relations rarely recur, the extraction prompts and the required depth of traversal do not transfer effectively between documents, leading to high computational costs and inconsistent results.