In the RAGraph Ontological Engine (ROE) project, we took the opposite path. Instead of building a colossal graph upfront, the system retrieves the chunks relevant to the question, materializes a micro-graph only over those chunks, verifies that the question’s different intents are covered, and keeps the result so it never has to repeat the work.
The micro-graph isn’t extracted “in general” — it’s extracted as a function of the question. Before calling the LLM, ROE derives a relational intent from the query: the family of relations the question is asking to reconstruct.
https://andreabelvedere.medium.com/dont-build-a-giant-knowledge-graph-build-the-tiny-one-your-question-needs-8dcc2120ffb0
No comments:
Post a Comment