Humanized-RAG tests compressed vectors for cheaper RAG search
Humanized-RAG is an RAG search method for finding useful information inside a with less wasted searching. The main idea is to group document embeddings into a smaller compressed form, described as vectors made from other vectors. This lets a search move from broad areas to narrower ones instead of checking many small pieces from the start.
The structure works like a tree: , topics, subtopics, and then detailed leaf items. The tree is built by clustering similar embeddings, and HDBSCAN was used in the tests. The goal is to build a more efficient that could be compared with well-known RAG methods.
The HRAG is available, and a paper draft with tests and results is being prepared for improvement and possible publication.
Key points
- The method compresses embeddings into a higher-level search structure.
- Search moves through a hierarchy from broad topics to detailed items.
- Similar content is grouped with clustering, and HDBSCAN was used in testing.
- The aim is a more efficient for RAG.
- The HRAG and a draft paper are available for feedback.